Genetic Entropy

"Since most mutations, if they have any effect at all, are harmful, the overall impact of the mutation process must be deleterious." — James Crow, University of Wisconsin¹

"Some unconditionally deleterious mutations of smaller effect size will inevitably fix, producing cumulative degradation termed 'Ohta's ratchet'." — Matheson, Hernández, Bertram and Masel, University of Arizona, 2025²

"If substitutions at 10% of all nucleotide sites have selection coefficients within this range with the mean 10^-6, an average individual carries approximately 100 lethal equivalents." — Alexey Kondrashov, asking why we have not died a hundred times over³

Overview

Every human is born with roughly 70 new mutations that neither parent had.^{4 5}  Some fraction of them land in DNA that is doing a job, and when a mutation lands in working DNA it is far more likely to break something than improve it.  Natural selection removes the worst of them.  But selection is not a microscope.  It cannot see a mutation whose effect on survival is smaller than the ordinary noise of luck, and the size of that blind spot is set by population size.

The claim called genetic entropy is that a large fraction of harmful mutations fall inside that blind spot, so they accumulate permanently, generation after generation, and the genome degrades.

Three things about that claim need separating, because they are usually run together and only one of them survives scrutiny.

1. The load argument — that the sheer number of harmful mutations should already have killed us. This one does not work. It depends on comparing every human to a hypothetical mutation-free human who does not exist, and mainstream population geneticists have a good answer to it. This article explains their answer and concedes the point.
2. The ratchet argument — that mutations too weak for selection to see fix permanently in the whole species, one after another, forever. The fixations are real and counted, by mainstream labs, and that part is not in dispute. Whether "degrading" is the right word for them is disputed, and this article takes on the best version of the objection.
3. The accounting argument — that this destroys specific genetic information faster than evolution builds it. This one is live and unsettled, and it is where the real disagreement now is.

The number that decides everything is U: how many harmful mutations each person gets per generation.  Mainstream estimates run from about 2 to about 10.^{6 2}  If the human genome is more functional than the low end assumes, U is higher.

What both sides now agree on

This debate has moved, and the movement is worth stating plainly, because the popular versions on both sides are ten years out of date.

In 2025 a team at the University of Arizona led by Joanna Masel — no creationists, funded in part by the NIH and the John Templeton Foundation — simulated a human-scale population for 200,000 generations to test exactly this question.  (Their paper is a preprint, not yet peer reviewed, and that should be weighed.)  They counted 56,255 harmful mutations that fixed permanently against 6,387 beneficial ones.²  They call the process Ohta's ratchet, they treat it as real, and their paper is titled to reassure: the population survives.

It survives, in their model, because the beneficial fixations were individually about a hundred times larger in effect than the harmful ones, so the books balance.  Their own words: "the accumulation of mildly deleterious fixations (Ohta's ratchet) is countered by a smaller number of larger-effect beneficial fixations under quite conservative assumptions."²

So the question is no longer whether mutations classified as deleterious fix permanently.  On the counts, both sides agree.  The questions now are:

• Is the beneficial mutation rate their answer requires one that anyone has measured? Their own paper says "there is great uncertainty in the true values of U_b and s̄_b."²
• Does compensation restore lost information, or only mask it? A better chaperone protein that stops a hundred slightly-misfolded proteins from causing trouble is a real fix for fitness. It does not restore the hundred specific sequences that were degraded.

One caution before going further, because it is the strongest thing that can be said against this whole article.  A major school of population genetics denies that "degrading" describes those fixations at all.  Nick Barton, in PNAS in 2022, names both the drift barrier and Kondrashov's version of this argument and rejects the framing outright:

"Lynch (21) has applied this concept to argue that molecular adaptations that are under weak selection cannot be established or maintained in (relatively) smaller populations, imposing a 'drift barrier' to adaptation.  Along the same lines, Kondrashov (55) has argued that deleterious mutations with N_e s ≈ 1 will accumulate, steadily degrading the population.  Both ideas seem problematic if we view adaptation as due to optimization of polygenic traits: Organisms can be well adapted even if drift dominates selection on individual alleles, and, under a model of stabilizing selection on very many traits, any change that degrades fitness can be compensated."⁷

That is not a fringe position and it is not answered by counting fixations.  It gets a full section below, because it — not truncation selection — is the objection this argument actually has to survive.

What this article does

It works the numbers out in the open, and it gives you a simulator to check them with.  Every headline figure below comes out of standard population genetics — Kimura's diffusion equations, Haldane's load formula, Watterson's estimator — with parameters taken from mainstream measurements, and the code that produces them runs in your browser at Drift Lab.

It also says plainly where the creationist case is weak.  John Sanford's book Genetic Entropy, which put this argument on the map, contains a quotation of a Nature paper that deletes the conditional clause the authors' whole conclusion depends on.  That is documented below, with both versions side by side.  A reader who suspects creationists of quote-mining should be able to find that out here rather than somewhere else.

Three arguments that get called genetic entropy

Most of the confusion in this subject comes from three separate claims wearing the same name.  They have different evidence, different opponents, and different verdicts.

Taking them in order.

The four numbers everything depends on

How many mutations

Direct sequencing of parents and their children settled this.  A 2012 study of 78 Icelandic trios found an average of 63.2 new mutations per child at a paternal age of 29.7 years, rising by about 2 mutations for every year of the father's age.⁴  A larger deCODE study in 2017 confirmed the pattern.⁵  Long-read sequencing in 2025 pushed the count of all new genetic changes — including repeat-length changes and structural variants — to an average of about 152 per generation, but two-thirds of that increase is tandem-repeat length changes whose harmfulness is unmeasured.⁸

This article uses 70 point mutations per person per generation — the low end of the direct-sequencing measurements, and lower than the 100 to 175 that older creationist material quotes.

It is also lower than the number this argument's leading critic uses, which is worth saying because it runs against the article.  Larry Moran calls the direct-sequencing method unreliable, puts the biochemical estimate at 130 and the phylogenetic estimate at 112 to 160, and has used 100 as his working figure since 2015.⁹  Every number below would be larger on his figure.

How harmful they are

Not very, individually — and that is the problem, not the reassurance.

The distribution of fitness effects has been measured in humans, flies, yeast and bacteria, and it is always the same shape: a long tail of barely-harmful mutations, with a thin tail of severe ones.  Fitted to a gamma distribution, the shape parameter comes out near 0.2.^{10 11 12}  In plain terms: most mutations that do anything at all do very little, and a few are disasters.

One caveat belongs here rather than in a footnote, because a critic will find it.  Those gamma fits describe the deleterious side only, by construction — a beneficial fraction of zero is assumed, not measured.  When Böndel and colleagues fitted a two-sided distribution to Chlamydomonas mutation-accumulation lines, they got a shape of 0.30 and a fitted positive-effect frequency of 0.84, while stating plainly that for most of their mutations "we do not know whether their effects are positive or negative."¹³  The one-sided fit is standard practice and it is also an assumption.

That is exactly the wrong shape.  Selection is good at removing disasters.  It is blind to the long tail.

Kimura made the point himself in 1979.  He proposed the gamma distribution for this purpose and shaded in the region he called "effectively neutral" — mutations with selection coefficients between 0 and 1/(2N_e).  His printed caption: "Frequency distribution of selection coefficients among mutants at different sites within a cistron (gene).  The shaded area represents the fraction of effectively neutral mutations."¹⁴  His own summary: "even if the frequency of strictly neutral mutations (for which s' = 0) is zero in the present model, a large fraction of mutations can be effectively neutral if β is small."¹⁴

He also drew the consequence, on the paper's last page:  "effectively neutral, but, in fact, very slightly deleterious mutants accumulate continuously in every species."  He put the cost at about 10^-7 of fitness per generation, called its threat to the species "a moot point," and said it "will easily be taken care of by adaptive gene substitutions that must occur from time to time (say once every few hundred generations)."¹⁴  So the founder of the neutral theory granted the mechanism in 1979.  What he did not do was measure the beneficial supply that his answer needs, which is the question Argument 3 takes up.

Modern data put numbers on Kimura's shaded region for human proteins.  Kim, Huber and Lohmueller fitted the distribution to large human samples in 2017 and found "34–37% of new mutations have |s| < 10^-4," with about a quarter of new amino-acid changes below 10^-5.¹²  The human drift barrier used in this article is about 4 × 10^-5, so roughly a quarter to a third of new protein-changing mutations sit in the zone where selection cannot see them.  That is the measured part.  How much of the rest of the genome behaves the same way is the unmeasured part.

How much of the genome can be damaged

This is the biggest uncertainty in the whole subject, and it is worth being blunt about it: U scales linearly with the functional fraction of the genome.  Double the functional fraction and you double the problem.  This is why the junk-DNA fight matters so much to both sides.

Two qualifications on that scaling, both conceded to critics who raised them.  It holds only for sequence-specific function, and the cleanest statement of that limit is in Joe Felsenstein's own population genetics textbook:

"If much of the DNA is simply 'spacer' DNA whose sequence is irrelevant, then there will be a far smaller mutational load.  But notice that the sequence must be truly irrelevant, not just of unknown function.  If the 'extra' DNA has regulatory or chromosome-pairing function requiring it to have a specific base sequence, then mutations in that sequence will still cause a mutational load, even if these loci are not producing polypeptide chains."¹⁵

That cuts both ways and both should be stated.  DNA of unknown function still counts toward U if its sequence matters — which is the article's position.  And DNA that has function without needing a specific sequence does not count, which Moran presses: "there are bulk DNA hypotheses that attribute non-sequence specific function to most of the genome and if they are correct the genetic load argument carries no weight."¹⁶  He thinks those hypotheses are unsupported, and so does this article — but it means the companion argument has to establish sequence-specific function, not function of any kind.  And the scaling is an approximation, because functional classes differ in how much mutation they tolerate: an RNA secondary-structure element absorbs point mutations far better than a coding exon.  That cuts both ways — such elements are more tolerant per site, and they also occupy more sites.

Comparative genomics — looking for sequence conserved across mammals — gives estimates from 4% to 15% depending on method and on how much correction is made for turnover:

That last row is the honest floor and it deserves to be quoted rather than buried.  A direct measurement inside 71,702 human genomes found only 0.4% to 0.7% of the genome under strong purifying selection, implying just 0.26 to 0.51 strongly deleterious mutations per generation.²²  If genetic entropy were a claim about strongly harmful mutations, that number would sink it.  It is not — it is a claim about the mutations below that threshold, which is precisely the region Dukler's method is designed not to detect.

Multiplying through, with 70 mutations per generation:

• Low. Peter Keightley's 2012 review puts the whole-genome deleterious rate at U ≈ 2.2.⁶ Lesecque and colleagues get 2.1.²³
• Middle. Using published constraint directly: Rands's turnover-corrected 8.2% gives U ≈ 5.6. Zoonomia needs more care, because it publishes two numbers that measure different things. Its 10.7% is element-level — the share of the genome lying inside constrained elements, which counts every base in an element whether or not that base is itself constrained — and multiplying by it as though it were per base gives U ≈ 7.3, which overstates the case. Its single-base figure is 3.26%, which gives U ≈ 2.25, Keightley's number again.²⁰ That one runs low for the opposite reason: a base counts only if it passes a conservative significance test (FDR below 5%), so weakly constrained bases are missed. This article therefore treats Zoonomia as a bracket, 2.25 as a floor and 7.3 as a ceiling, and does not use 7.3 as an estimate. The Masel group states its own working range as "U_d = 2-10."²
• High. Kellis's union of constraint estimates (16–26%) gives U ≈ 11 to 18. If the higher functional-DNA estimates are right — the case is made in the companion article Functional DNA is Evidence of Design — U is above 20.

Dan Graur, an evolutionary biologist and one of the fiercest critics of the ENCODE project, worked the same arithmetic in the opposite direction and reached the same place.  His 2017 paper is titled "An upper limit on the functional fraction of the human genome," and its method is to ask how much function the mutation load can tolerate.²⁴  He concluded that the functional fraction "cannot exceed 15%," and that if ENCODE's 80% figure were right, the required fertility would be absurd:  102 children per couple.  His table carries a factor-of-two slip, and correcting it helps ENCODE.  Table 1 computes the required fertility as e^2U where e^U belongs, because the diploid genome size he multiplies by already counts both copies of every site.  A 2019 corrigendum added the missing "2" to his printed equation (3) so that it matched the table, announced that "The correct Table 1 is shown below," and reprinted Table 1 with every cell unchanged.²⁵  Corrected, his ceiling at his own most permissive mutation rate is about 28% rather than 14–15%, and ENCODE's 80% needs about 14 children per couple rather than 102.  That is still far beyond any human population, so the conclusion against ENCODE survives;  the ceiling itself is roughly twice what he printed.  Both sides of this argument agree on the arithmetic.  They disagree about which premise gives.

Which of these numbers are measurements, and which are model outputs

Before going further, one distinction has to be made, because it changes how much weight each number can carry — and it is a distinction this article's sources do not always make for you.

Some of these quantities are measured.  The mutation rate comes from sequencing parents and their children and counting what is new.  That is a direct observation.  It needs no assumption about how long humans have existed, how big the population was, or when anything diverged.  Laboratory measurements of mutational effects are the same: a fitness assay in a flask is a measurement, whatever else is wrong with it.

Others are outputs of a fitted model, and the model contains a population history.  Distributions of fitness effects inferred from human polymorphism data are obtained by fitting demography and selection jointly — you cannot get one without assuming the other.  So is the effective population size of 10,000, which is a harmonic mean recovered from linkage disequilibrium under an assumed demographic history.  So, in a different way, are constrained fractions from comparative genomics, which depend on a phylogeny and its divergence dates.

Johri, Pfeifer and Jensen address that dependence directly in "Developing an Evolutionary Baseline Model for Humans: Jointly Inferring Purifying Selection with Population History."  They fit population history and purifying selection together, including the effect of selection on linked sites.²⁶

And the coupling runs both ways, which is worse than a simple assumption.  Soni, Pfeifer and Jensen showed in 2024 that background selection skews the site frequency spectrum, that the skew "may then be mis-inferred as population size changes using neutral demographic estimators, which may in turn result in a mis-inference of the DFE."²⁷  Errors in the history and errors in the selection estimate feed each other.  With realistic variation in mutation and recombination rates they found "variable mutation and recombination rates resulting in the inference of an excess of effectively neutral mutations," and in one case "DFE-alpha inferred a completely neutral DFE for population contraction scenarios."²⁷

What follows for this article.  There is a difference between adopting an opponent's timescale for the sake of argument — which is legitimate, and is what this article does throughout — and importing a number whose value was produced by assuming that timescale.  The first is arguing on someone else's premises.  The second is quietly inheriting them.

Of the four numbers this section is about, one is a measurement and three are model outputs.  That does not make the three wrong.  It does mean that anyone who rejects the demographic history also has to recompute them rather than borrow them, and that a reader who accepts the history should treat their error bars as wider than the published intervals suggest, because the published intervals condition on the model being right.

This article uses the mainstream values throughout and says so.  Where a conclusion depends on one of the three rather than on the measurement, that dependency is now named at the point it is used.

This article uses U = 5 as its working figure — the middle of the mainstream range, derived from a mainstream constraint estimate.  Every result below is also reported for U = 1.5 and U = 30 so you can see how much it matters.

And U = 5 sits above the ceiling the critics themselves name, which is a problem for them rather than for this article.  The clearest statement of it is not in a blog post but in PLoS Genetics, in a paper defending junk DNA.  Palazzo and Gregory set out the mechanism — "If the rate at which these mutations are generated is higher than the rate at which natural selection can weed them out, then the collective genomes of the organisms in the species will suffer a meltdown as the total number of deleterious alleles increases with each generation" — then give the ceiling, "approximately one deleterious mutation per generation," and then, in the same passage, the observed figure: "it has been estimated that humans sustain 2.1–10 deleterious mutations per generation."²⁸

They resolve it the only way available to them, and they state the conclusion plainly: "By this line of reasoning, we would estimate that, at most, only 1% of the nucleotides in the genome are essential for viability in a strict sequence-specific way."²⁸

Larry Moran makes the same argument less formally, putting the tolerable rate at "no more than 1 or 2 deleterious mutations per generation," and saying that at four per generation "that's an intolerable genetic load; it suggests that most of our genome is not a target (i.e. not functional)."^{29 29}

Notice where that ceiling comes from.  It is derived from the load formula — the same formula this article concedes below does not work, on the authority of Lesecque, Charlesworth and Galeota-Sprung.  The two positions cannot both be held.  If the load formula does not predict extinction, then "no more than 1 or 2" is not a real ceiling and the junk-DNA inference drawn from it does not follow.  If it does predict extinction, then U = 5 is a prediction of extinction and the argument in this article is conservative.  Anyone who takes the relative-fitness escape has to give up the ceiling with it.

This article takes the first horn: the load formula does not predict extinction, U = 5 is tolerable, and nothing in Argument 1 is a proof of anything.  The argument that survives is the ratchet, which does not use the load formula at all.

The size of selection's blind spot

A mutation is invisible to natural selection when its effect is smaller than the random noise of who happens to survive.  The threshold is sharp and it is one of the most solid results in population genetics: a mutation is effectively neutral when

2 × N_e × s < 1

where N_e is the effective population size and s is the fitness cost.  For humans, N_e is estimated at about 10,000.²  So the blind spot covers every mutation costing less than about 0.005% of fitness.

Three qualifications, all of which cut against the argument this article is making, and all of which belong here rather than buried later.

It is not a cliff.  "Effectively neutral" is a convenience, not a wall, and Tomoko Ohta — whose name the ratchet carries — says so herself: "the fixation probability is a continuous monotone function of Ns.  Thus, when discussing molecular evolution by nearly neutral mutants, one has to consider all mutants around Ns ≈ 0."³⁰  Right at the boundary, where N_e·s = 0.5, a beneficial mutation is still 2.31 times more likely to fix than a neutral one and 7.39 times more likely than its harmful mirror image.  Selection is weak there; it is not absent.  The figure below shows the transition running about a decade and a half wide, and the prose should be read against the figure, not against the inequality.

N_e = 10,000 is the critic's number, not ours.  It is a harmonic mean over hundreds of thousands of years, inferred from linkage disequilibrium under a demographic model that assumes deep time.  This article uses it anyway, on purpose: every quantity here is taken from mainstream measurement so the arithmetic can be checked by people who reject the conclusion.  An argument that needed its own population sizes would not be worth making.

And going forward it is collapsing.  For a mutation arising now, the relevant effective size is the present and future one, not the ancient harmonic mean.  Joe Felsenstein — who worked out the overlapping-generations correction in 1971 — applies it here:

"The overlap of generations reduced the Ne to 1/3 of the census size.  So maybe 2.7 billion.  The result would be that the dynamics of any new mutations, from now on, would be nearly-neutral if the selection coefficients were less than +/- 0.0000000003."³¹

That is a blind spot more than five orders of magnitude narrower than the one this article works with, and on that number the ratchet very nearly stops.  The objection is real and the reply is narrow: the load already fixed was fixed at the small ancient N_e, and fixation takes on the order of 4N_e generations, so a population explosion roughly ten generations old has not yet had time to change what fixes.  The ratchet's rate as computed here is a statement about the past, not a forecast about the future.  Anyone quoting these figures as a prediction about modern humanity is misusing them.

That said, the blind spot in the past sounds like nothing.  It is not nothing, for a reason Kimura himself pointed out: mutations of small effect are more damaging in the long run than mutations of large effect, because the large ones get removed and the small ones do not.  A mutation that kills you is a mutation that leaves no descendants.  A mutation that costs you 0.002% is a mutation your species keeps forever.

Kimura's 1962 diffusion result, drawn for a population of effective size 10,000.  Below about s = 0.00001 a harmful mutation is as likely to spread through the entire species as a neutral one.  Above about s = 0.0005 it essentially never does.  The transition is only about one and a half orders of magnitude wide, and the whole argument lives inside it.

Argument 1: the load paradox, and why it is not a proof

The arithmetic

Haldane worked this out in 1937 and it has been rederived many times since.  At mutation-selection equilibrium, each harmful mutation reduces the population's mean fitness by twice its mutation rate — regardless of how harmful it is.  Multiply across all sites and mean fitness settles at

w̄ = e^(−U)

So the fraction of the population that must fail to reproduce, relative to a mutation-free individual, is 1 − e^(−U).[crow-1997] ³²  The numbers get ugly fast:

Humans do not have 40 children, let alone 807.  Hunter-gatherer women reach menarche around 16 and menopause around 47, with an inter-birth interval near 3 years, giving a ceiling of about 11 children.²³

The clearest sign that this is not a creationist construction is that it is worked out, with these numbers, in the standard textbook written by one of the argument's sharpest critics.  Felsenstein takes 20 million loci mutating at 10^-7, gets a mean fitness of e^-2 ≈ 0.15, and concludes: "Clearly an organism with as much DNA as we have would be in severe trouble.  Yet in humans well over 98% of all newborns survive to adulthood in most industrial countries."¹⁵

He then lists three ways out, and it is worth noticing that the article's whole structure is a response to his list.  The first is junk DNA, qualified as above.  The second is that added function raises the fitness ceiling at the same time as it raises the load — which he objects to himself, because "it seems to predict that as more loci enter the genome the discrepancy between maximum possible fitness and average fitness increases.  This is hard to reconcile with the high viability of humans in industrial countries."¹⁵  The third is interaction between mutants, so that one selective death removes many mutations — the truncation family of models, which he traces to Sved, Reed and Bodmer, King, and Milkman, all in 1967.  That third answer gets its own section below.

This is why the load problem has its own literature, most of it written by people with no theological stake in it.  Hermann Muller raised it in 1950.³³  Kondrashov's 1995 paper is titled "Contamination of the genome by very slightly deleterious mutations: why have we not died 100 times over?"³  Crow's 1997 paper asks whether the mutation rate is a health risk and answers yes.¹  Eyre-Walker and Keightley's 1999 Nature paper found the coding-sequence rate alone was "close to the upper limit tolerable."³⁴  Nachman and Crowell in 2000: "each female would need to produce 40 offspring for 2 to survive."³²

The strongest mainstream answers

There are two, and creationist sources usually engage only the first — which is the weaker of them, and the one the field has largely moved past.

(a) Truncation selection — the historical answer

Truncation selection works, and it is cheap.  If everyone below a fitness threshold fails to reproduce, mutations are removed far more efficiently than if each is selected against independently.  Crow and Kimura put the gain at 5 to 20 fold.³⁵  Their point is not that truncation demands enormous fecundity — it is the reverse.  Truncation is what lets a low-fecundity animal pay a load that multiplicative arithmetic says it cannot, and the example they work culls only 10% of the population, which is Haldane's own budget and which any vertebrate can afford.

That has to be granted plainly.  It is the single best reason to think the load arithmetic overstates the problem.

The mainstream limit on it comes from the same population geneticist this article quotes on Mendel's Accountant.  Asked directly why truncation is absent from standard models, and pointed at the Crow and Kimura paper above, Joe Felsenstein answered: "There are papers, such as that one, that try to work out the consequences of truncation selection.  But generally we don't see such a mechanism at work in populations.  That and it's harder to model mathematically."³⁶  That is a stronger objection than Crow's, because Crow was calling strict truncation unrealistic while Felsenstein is talking about the mechanism as a class.

And Sanford's own group demonstrated it.  In Biological Information: New Perspectives — a symposium volume Sanford co-edited — Brewer, Baumgardner and Sanford set Mendel's Accountant to ten new deleterious mutations per person per generation, six offspring per female, and strict truncation selection.  Mutation count stopped rising at about 125 per individual by generation 100, and fitness settled at 0.94, flat for the rest of the run.³⁷  Halve the truncation so that half the culling is luck, and it still holds.  Sanford wrote the admission into the volume's own section introduction: the mechanism "actually can work, but only when mutation effects are relatively uniform, when there is truncation selection, and where there is sexual recombination."³⁸

Two things follow, and neither is comfortable for the argument this article is making.  The fecundity is charged and it is within reach — six births per woman with most not reproducing is inside the ceiling of about eleven that hunter-gatherer biology allows.  And the equilibrium load is small: a 6% fitness deficit, not the 89% the multiplicative formula demands.

Their escape, and the problem with it.  Their answer is that real mutations do not all have the same effect.  Once effects span orders of magnitude, an individual's mutation count stops predicting its fitness, so selection on fitness stops removing counts.  Re-run with a realistic spread of effects and every selection scheme declines to extinction.³⁷

That figure has a problem its authors never mention.  Rank the four selection schemes by how long the population survived and they come out backwards: full truncation dies first, at about generation 4,700, and ordinary fitness-proportional selection lasts longest, to about generation 8,100.  More truncation made the population die sooner.  That is the reverse of Crow and Kimura, the reverse of the same paper's own earlier figure, and the reverse of what the paper concludes from it.  The likely cause is Mendel's additive fitness model: when mean fitness is allowed to fall toward zero while variance stays finite, fitness-proportional selection outruns truncation, because the response to selection scales with variance divided by the mean.  Whatever the cause, the volume's central argument against truncation rests on a run in which truncation behaves anti-selectively, and nobody flagged it.

Three things bound it.

Price.  The gain scales with how hard you cull.  Culling 10% is nearly free; culling half means every breeding female must raise four offspring to adulthood, and culling 90% means twenty.  Crow, who thought quasi-truncation realistic, still called strict truncation "totally unrealistic."¹  And when Milkman introduced the threshold model in 1967, he justified its affordability by naming the organism it was affordable in: culling to the top 13.5% is "a practical procedure in Drosophila, whose females produce scores of eggs."³⁹  He was not conceding a difficulty — he was arguing the mechanism works.  But the reason he gives for its being practical is fecundity humans do not have.

Heritability.  Truncation only removes mutations to the extent that the trait being truncated is heritable.  How steeply the response falls with heritability depends on the yardstick:  the share of the imposed selection that is passed on scales with h² itself (R/S = h²), while the move measured in additive-genetic standard deviations scales with its square root (R/σ_A = i·h).  Measured heritability of total fitness in wild vertebrates is about 3%.⁴⁰

And whether real selection looks like this at all.  The Masel group tested Haldane's assumption directly in Genetics in 2025 and found the proportion of deaths that are genuinely selective can far exceed his 10% ceiling — between 8.5% and 95% depending on conditions.⁴¹  That is a real result and it cuts against the load argument.  Their data are from Arabidopsis, an annual selfer with enormous fecundity, and they flag less-fecund species as future work.  Humans are the paradigm less-fecund species.

Drift Lab reports the bill rather than assuming it away: run truncation selection and it computes the fecundity that regime demands and whether the species you specified can pay it.

But this is not where the argument is anymore, and creationist sources that stop here are fighting a dead war.  The Masel lab's own summary calls synergistic epistasis "one historical solution to the mutation load paradox, often modeled in extreme forms such as truncation selection," and then closes it: "Since then, data has come out showing that the mean effects of two deleterious mutations are extraordinarily close to multiplicative."⁴²  Truncation was answered by the data, not by argument.  The live answer is different.

(b) Polygenic stabilizing selection — the live answer

Zach Hancock, a population geneticist who co-authored the formal rebuttal to Basener and Sanford's 2018 model, was asked directly what his strongest argument against genetic entropy is.  His answer was not truncation:

"The argument I presented during the debate is what I consider the strongest argument against GE — that is, that most complex traits are highly polygenic, and thus selection is exceptionally effective despite being weak on each individual allele (e.g., Barton 2022)."⁴³

The mechanism is the one Barton states in the passage quoted at the top of this article.  Under the infinitesimal model, a trait is the sum of thousands of alleles of tiny individual effect.  Selection acts on the sum.  It can hold the sum at its optimum with enormous precision even though no single allele is under selection strong enough to beat drift.  And because many different combinations of alleles produce the same sum, a mutation that pushes the trait one way can be offset by anything that pushes it back — which is exactly what Barton means by "any change that degrades fitness can be compensated."

Charlesworth put numbers on it in 2013, in the paper whose title asks the load question outright.  With human parameters — Sa² = 5 × 10⁻⁷, N_e = 20,000, π = 0.001 — his stabilizing-selection load comes to L*stab = 0.05.⁴⁴  Five percent, not eighty-nine.  His verdict on that model is unambiguous: "the load can be small to moderate, even with tens of millions of sites subject to selection of the strength estimated from population genetic data.  There is thus no threat to the survival of the population."⁴⁴

Read the next sentence of that paragraph, though, because it is the whole ballgame:

"However, the relative importance of stabilizing versus purifying selection for synonymous and noncoding sites is currently unknown, so the question of the load under weak purifying selection also needs consideration."⁴⁴

Charlesworth's escape is conditional on which kind of selection those sites are actually under, and he says nobody knows.  So he works the other branch — and this is the part of the paper that never gets quoted:

"in contrast to the situation with stabilizing selection, if there is genome-wide purifying selection of the type described by equation (4) on the scale suggested by the recent findings mentioned earlier, an extremely high level of genetic load is inevitable, in agreement with Kondrashov (1995a)."⁴⁴

His numbers on that branch are not small.  For Drosophila, with 14 million synonymous codons under purifying selection, the load is 0.83 with completely synergistic selection and 0.97 with multiplicative fitnesses; use the larger noncoding count and they become 0.998 and 1.00.  And then: "The situation is even worse in species with effective population sizes that are 10-fold or more smaller, and that have many more noncoding sites under selection, such as humans."⁴⁴

So the paper routinely cited as the answer to the load paradox says the load is near-total if the sites are under purifying selection, that this is worse in humans than in flies, that it agrees with Kondrashov, and that which model applies is unknown.  Its resolution is to reinterpret the sites as being under stabilizing selection instead — which is a hypothesis about what those sites are doing, not a measurement of it.

That is the same open question this article keeps arriving at from other directions, and it is now stated by the strongest source on the other side.

Three things have to be granted.  Trait-level selection is real and powerful; the entire history of plant and animal breeding is the demonstration.  Complex traits genuinely are polygenic, and the omnigenic evidence for it has only gotten stronger.  And selection on a sum genuinely is far more effective than selection on any of its terms — this is not a trick, it is the reason quantitative genetics works.

A note on a quotation you will meet.  Charlesworth is often quoted as writing that "the mean fitness of the population relative to the fitness of a hypothetical optimal genotype that has a very low chance of being present in the population is essentially irrelevant."  He does write that, and this article accepts it — it is the lag-load point conceded below.  But the sentence immediately after is usually dropped, and it matters: "However, we need to be assured that the level of variation in fitness is not so large as to be implausible, but is large enough to be compatible with a selection coefficient of the order of 1/(4N_e)."⁴⁴  He is replacing one test with another, not abolishing the problem.

And here is the reply.  Read Barton's sentence again.  Not repaired — compensated.

The infinitesimal model is a model of trait optimisation, and it delivers exactly what it promises: the trait mean stays at the optimum.  It says nothing whatever about whether the alleles holding it there are the ones that were there before.  In fact it says the opposite, and says it as a feature: the whole reason the model tolerates drift at individual sites is that the specific alleles do not matter, only the sum.  Sequences turn over underneath a stable phenotype.  Hartl and Taubes showed the same thing formally in 1996: compensatory evolution "results in no change or permanent improvement in enzymatic function."⁴⁵

So the polygenic answer is not a refutation of the accounting argument.  It is the accounting argument, stated by the other side, and counted as good news.  Fitness is maintained; sequence identity is not.  Whether that difference matters is the whole of Argument 3 below, and it is not settled by pointing at the load number.

Where the polygenic answer is strong and where it is weak.  It is strongest for traits that really are sums — height, yield, risk scores, anything where a thousand small contributions add up and many combinations reach the same total.  It is weakest for the parts of the genome that are not doing arithmetic at all.  A splice acceptor site is not a polygenic trait.  A transcription factor binding motif either matches or it does not.  A protein's active site is a specific arrangement of specific residues.  For those, there is no compensating allele elsewhere that restores the sum, because there is no sum.  The infinitesimal model is a good description of the additive, quantitative fraction of the genome and a poor description of the switch-like fraction, and the argument here has always been about the second.

That is a real division of the genome, and neither side has measured how it splits.  It is the sharpest open question in this dispute and it is stated as such in what this does and does not show.

And the objection has a mainstream answer that does not come from this side of the argument at all.  Michael Lynch, whose drift-barrier work Barton was criticising in the passage quoted above, has taken the argument onto Barton's own ground — stabilizing selection on quantitative traits — and reached the opposite conclusion.  From his 2025 paper with Menor in Genetics:

"Although multigenic traits are often assumed to be under some form of stabilizing selection, numerous aspects of the population-genetic environment can cause mean phenotypes to deviate from presumed optima, often in ways that effectively transform the fitness landscape to one of directional selection."⁴⁶

And on whether the barrier stops at the molecular level:

"Drift barriers, which are universal properties of all finite populations (i.e. all of biology), denote the limits to which mean phenotypes are expected to wander over evolutionary time."⁴⁶

So the question of whether a polygenic trait actually sits at its optimum in a finite population is live inside mainstream population genetics, argued by the two people best placed to argue it, and it has not been settled in Barton's favour.  This article does not need it settled in the other direction either.  It needs the reader to know that "selection on polygenic traits handles it" is a position with an opponent, not a closed case.

One more thing should be said, and it comes from Charlesworth rather than from anyone arguing this article's side.  In a separate paper the same year, he worked out whether the polygenic answer can be told apart from the ordinary one using sequence data, and found that the frequency spectra produced by these models can be difficult to distinguish:

"These results mean that studies utilizing data such as the frequency spectra of DNA sequence variants may find it virtually impossible to distinguish stabilizing from purifying selection."⁴⁷

So the strongest objection to this article is a model that fits the data, not a measurement that decides between models.  That is not a criticism of it — the same is true of the position argued here.  It is a reason to hold both loosely, and a second place where the dispute reduces to a measurement nobody has made.

Why it does not prove the genome is decaying

Because the comparison is to a genotype that does not exist.

Every one of those calculations measures the population against a hypothetical human with zero harmful mutations.  In a population where everybody carries thousands, that individual has never been born and never will be.  What actually determines who reproduces is how you compare to the people around you.

Lesecque, Keightley and Eyre-Walker made this formal in 2012 in a paper called "A resolution of the mutation load paradox in humans."  Under a relative-fitness model, they showed, "a species could tolerate 10's or even 100's of new deleterious mutations per genome each generation."²³  Brian Charlesworth reached the same destination by a different road the following year, in the paper quoted above: his second resolution is soft selection, in which "individuals compete for a limiting resource whose abundance determines the absolute fitness of the population," and weak purifying selection across many sites "produces only a small variance in fitness among individuals within the population, due to the fact that most sites are fixed rather than polymorphic."⁴⁴  Galeota-Sprung, Sniegowski and Ewens took the same objection to Graur's functional-fraction bound in 2020 and concluded that "an argument from mutational load does not appear to be particularly limiting on f" -- f being the functional fraction of the genome.⁴⁸

This objection is correct, and any honest presentation of genetic entropy has to concede it.  The load formula assumes hard selection against an absolute standard.  Real selection is largely soft: a fixed number of slots, filled by whoever competes best on the day.  Under soft selection nobody has to have 807 children, because nobody is being measured against a perfect genome.

A second concession: the load's demographic cost depends on when you die

There is a further step, and it is worth taking on its own because it does not depend on relative fitness at all.  Even granting a load in absolute terms, the demographic price depends on how much of a lifetime's resources the individuals who die have already consumed.

Agrawal and Whitlock worked this out in 2012.  If a fraction L of the population dies of deleterious alleles before adulthood, and those individuals consume only a fraction β of what a healthy adult consumes, the equilibrium population size is

N̂ = bI(1 − L) / [d(1 − L(1 − β))]

Their conclusion: "if the selective deaths occur early in life, such that β is small, the demographic consequences of load can be greatly reduced, at least at the adult stage."⁴⁹  A miscarried embryo costs the population almost nothing.  So the "807 children" column is not what the load implies unless selection is acting on grown offspring.

This is a good argument, and it has a limit that is easy to miss: it is strongly dependent on U.  Run their formula across the same range as the table above.

Equilibrium adult population as a percentage of the unloaded population, from Agrawal and Whitlock's expression, at each mutation rate.  β is the fraction of an adult's lifetime resource consumption used up by an individual who dies a selective death.

At the low end of the U bracket the escape is complete: a load of 89% costs 7% of carrying capacity if the dying are embryos.  At the middle of the bracket it is not.  At U = 6 with β = 0.01, the population sits at a fifth of its unloaded size; to keep it above 70%, β has to fall to about 0.001 — every selective death has to cost one-thousandth of what raising an adult costs.

And humans are the worst case for small β.  Human reproduction is exceptionally expensive per offspring: nine months of gestation, single births, and fifteen or more years to maturity.  A human who dies selectively at fifteen has consumed a large fraction of an adult's lifetime resources.  Only very early embryo loss is cheap.  β is unmeasured in humans, on both sides of this argument, and neither side should pretend otherwise — but the direction of the human life history pushes it up, not down.

And their supplement narrows the escape further, in a way the headline formula hides.  Which vital rate the load falls on decides everything: "Load on birth rate or survival causes a nearly proportional decline in equilibrium population density, while load on acquisition ability tends to have little effect on density."⁴⁹  The cheap case is load on how well an organism gathers resources.  Deleterious mutations in humans do not mostly reduce foraging efficiency — they reduce survival and fertility, which is the expensive case.  Their own summary is that "the effects of load depend on the details of the interaction within and among species," and those details are not favourable here.

Agrawal and Whitlock say as much themselves, and it is the reason this concession is bounded rather than total: "In fact, ecological processes are unlikely to completely mask the effects of genetic load."⁴⁹

Felsenstein made the same point earlier and added the part that matters most for the species that are actually at risk.  Granting that culling 1000 offspring to 500 may barely move the adult population under density dependence, he writes:

"It would be a mistake to conclude from this that the mutational load has no effect whatsoever on population dynamics.  The reproductive excess acts as a buffer allowing the population to survive various misfortunes, and a reduction of the size of this buffer, even if it has little effect on average population size, may make the population more vulnerable to extinction in time of crisis."¹⁵

That is the honest shape of the load concession.  The load does not kill a large stable population.  It spends the margin that population has for surviving a bad century, which is exactly the quantity that decides the small-population cases below.

There is a real reply, and it is not a dodge.  Michael Lynch put it in 2016: relative fitness explains why nobody is dying at 89% per generation, but it does not make the deterioration unreal.  "Physical defects involving cancer, metabolic disease, and psychiatric disorders have very real costs regardless of the average population state."⁵⁰  Sickness measured against your neighbours is still sickness.

There is a second reply, and it comes from the lab whose simulation this article leans on most.  The Masel group's own research page states the problem without softening it:

"Most work on mutation load periodically renormalizes relative fitness to deal with the fact that fitness keeps declining no matter what, or else treats only a subset of the genome.  Either way, this is dodging the problem, not tackling its fundamentals."⁴²

The same page gives their answer, and it should be quoted alongside rather than left out: "We think the answer is a 'ratchet' in which many small-effect deleterious fixations are counterbalanced by a much smaller number of large-effect beneficial fixations."⁴²  They add a sentence that matters for what follows: "This has some similarities with 'drift barrier theory', but emphasizes that the cause of the ratchet is not small population size, but a high genome-wide deleterious mutation rate."⁴²  That is Argument 2's claim exactly — the ratchet runs on U, not on N_e being small.

So they are not saying the load paradox is unsolved and walking away.  They are saying the usual way of making it go away is bookkeeping, and offering the ratchet as the real answer.  That is this article's framing — load fails, ratchet holds, the accounting is what is left to argue about — arrived at independently by people who reach the opposite conclusion from it.

And the same page corrects this article's own arithmetic, which is worth stating rather than hiding.  The load in the table above is what population genetics calls the lag load: the gap between the population's mean fitness and a theoretical optimum nobody ever reaches.  Masel's page names that as a mistake:

"The 'lag load' is the difference between population mean fitness and some theoretical optimal fitness that is never obtained; inappropriate use of the lag load is one contributor to confusion (panel B).  The more appropriate load is the difference between the population mean fitness and the highest fitness individual actually found in a population — this quantity was rediscovered and named the 'lead' (Desai & Fisher 2007)."⁴²

The lead is the right quantity, and it is small where the lag load is enormous, because the best individual actually alive is not mutation-free either.  This is the same objection as the relative-fitness one, stated more precisely, and it is a second reason the "807 children" row should not be read as a prediction about anybody.  It is also what Galeota-Sprung and colleagues compute, so two of the sources this article leans on are making one coherent point rather than two scattered ones.

None of this touches Argument 2, because fixations are counted, not inferred from a load formula.

But that is an argument about consequences, not a proof of decline.  The proof has to come from somewhere else.

Argument 2: the ratchet, and why it holds

The mechanism

Selection acts on differences.  Once every member of a species carries the same mutation, there is no longer any difference to act on.  The mutation is fixed, and no amount of future selection can remove it, because there is nothing left to compare it against.  That is not a controversial claim; it is the definition of fixation.

The question is how often a harmful mutation gets all the way to fixation.  Kimura answered it in 1962.  For a new mutation starting in a single copy out of 2N, with a fitness cost s:

P(fixation) = (e^(2·N_e·s/2N) − 1) / (e^(2·N_e·s) − 1)

Below the blind spot this collapses to 1/(2N) — the same odds a completely neutral mutation has.  And Kimura's other famous result is that neutral mutations fix at exactly the rate they arise, no matter how big the population is.

Put those together and the conclusion is unavoidable.  If a fraction of new harmful mutations fall below the blind spot, that fraction fixes permanently at close to the full mutation rate, forever.

The numbers

Working the integral over a measured distribution of fitness effects, for a human-sized population:

Effective population size 10,000; gamma-distributed effects with shape 0.2 and mean 0.001; free recombination.  Computed from Kimura's fixation probability, and confirmed by simulation — see below.

The relationship is linear in U, which is why the functional fraction of the genome matters so much.  A larger population helps, but not nearly enough: raising the effective size a hundredfold, from 1,000 to 100,000, only cuts the rate by about 60%.

At the working figure of U = 5, roughly one harmful mutation becomes permanent in every human being, every generation.  Over the six million years since the human and chimpanzee lineages are supposed to have split, that is about a third of a million permanent losses.

Checking the arithmetic against the people who disagree

The strongest available check is not a creationist source.  It is the Masel lab's 2025 paper, which ran a far more elaborate simulation — real chromosomes, linkage blocks, background selection, beneficial mutations, 200,000 generations — and reported its fixation counts.

Feeding their published parameters into the calculation above gives:

Their parameters:  effective population size ~7,500, U_d = 2, and a gamma distribution of fitness effects after Kim et al. 2017 with shape 0.169 and mean heterozygous effect 0.009.²  The two figures are the two ways of feeding that heterozygous effect into Kimura's equation:  doubled to a homozygous effect (63,212) or taken as it stands (71,056).  An earlier version of this table printed 60,780 and called the agreement eight percent.  That figure used a gamma shape of 0.2 and treated 0.009 as a homozygous effect;  the paper states neither, and the two slips happened to cancel.

The two routes agree to within 12% to 26%, depending on how the paper's own effect sizes are read — one a C program written by a mainstream evolutionary genetics lab, the other a diffusion equation from 1962.  The gap runs in the expected direction:  the diffusion equation treats every site as evolving on its own, while real linkage and background selection in the simulation lower the fixation rate, so the simple calculation should come out higher, and it does.  Whatever else is in dispute, the rate of permanent genetic damage is agreed to within about a quarter.

Run it yourself

Drift Lab simulates this in your browser.  The default run is a diploid population with the parameters above.  Three things about it are worth knowing before you trust anything it says.

It is deliberately generous to natural selection.  Every site is inherited independently of every other, so no good gene is dragged down by a broken neighbour — no Hill-Robertson interference, no hitchhiking, no Muller's ratchet in the classical sense.  An individual's chance of reproducing is exactly its genome, with no luck involved.  The population can never shrink, so there is no death spiral.  All three of those assumptions make selection more effective than it can possibly be in a real organism.

It grades itself against results that predate computers.  The Verify tab runs the simulation against Kimura's 1968 neutral substitution rate, Watterson's 1975 formula for the number of segregating sites, Kimura's 1962 fixation probability, and Haldane's 1937 equilibrium.  All four pass:

PASS  Kimura 1968: neutral substitution rate = mutation rate
      measured 0.4976, expected 0.5000
PASS  Watterson 1975: number of segregating sites
      measured 3417, expected 3395
PASS  Haldane 1937: fitness at mutation-selection balance = e^-U
      measured 0.3693, expected 0.3679
PASS  Kimura 1962: probability a harmful mutation fixes anyway
      measured 0.0005820, expected 0.0005823

It reproduces the stable case too.  Set every mutation to cost 1% and the decline stops dead: mean fitness parks at e^(−U) and stays there for as long as you care to run it, with zero fixations.  Mutation-selection balance is real, and the simulator finds it.  That matters, because it means the decline in the other runs is not an artefact of the code.  It is the biology of the parameters.

And a caveat that applies to every simulation in this dispute, including this one.  A simulation's output is a consequence of its assumptions, not independent evidence for them.  Drift Lab cannot tell you whether U is 2 or 7, what the real distribution of fitness effects looks like below the drift barrier, or how much of the genome is functional.  It can only tell you what follows if you assume particular answers.  That is worth stating because simulation output is routinely presented as empirical in this argument — by both sides — and it is not.  What the self-tests above establish is narrower and more useful: that the code correctly implements the population genetics, so any disagreement is about the parameters rather than the arithmetic.  The parameters are where the argument actually is.

Where the ratchet ends

The most substantial objection to all of this comes from Dan Stern Cardinale and others, and it is correct as far as it goes: there is no infinite runaway.  Sites are finite.  Once a site has lost its letter, further mutations there do nothing more, and mutation runs both ways — a broken site can mutate back.  So an equilibrium exists, and the genome cannot decay forever.

True.  The two questions this leaves open are where the equilibrium sits and how long it takes to arrive, and both have answers.

Where.  In the regime real genomes are in, a site spends most of its time fixed one way or the other, and the ratio between the two is set by how much more likely a beneficial fixation is than a harmful one.  With three wrong letters for every right one:

fraction of sites sitting wrong at equilibrium = 3 / (3 + e^(2·N_e·s))

At an effective population size of 10,000:

Verified three ways: by numerical integration of Wright's stationary distribution, by a direct Wright-Fisher simulation in a rescaled regime, and by the Genome kernel in Drift Lab.  All three agree to within a percent.

A third force this formula leaves out, and why it does not change the answer.  Mutation and selection are not the only things that decide which letter a site ends up with.  GC-biased gene conversion — gBGC — is a repair bias during recombination that favours G and C over A and T, and it acts regardless of fitness.  It is not a mutation-rate effect; it behaves like a selection coefficient, which means it operates in exactly the near-neutral zone this whole argument is about.  A critic is right to ask whether it breaks the equilibrium above.

Worked through, it does not, and the reason is a coincidence worth stating.  gBGC pushes sites whose required letter is G or C toward the right answer and sites whose required letter is A or T toward the wrong one, so it splits functional sites into two groups shifted in opposite directions.  Whether that raises or lowers the average fraction sitting wrong depends on the curvature of the equilibrium curve — and that curve is a logistic centred at 2·N_e·s = ln 3 ≈ 1.10, which is precisely the point at which it is 50% wrong.  At the inflection the two shifts cancel exactly.  The article's headline claim about weakly-constrained sites sits on the one point in the whole range where gBGC has no net effect at all.

Away from that point it has some effect, but not much, because gBGC is weak.  Glémin and colleagues measured the population-scaled coefficient B = 4N_e b across three human populations and got 0.27 to 0.43 — "in the nearly neutral area" — with B above 5 in only 1–2% of the genome, near recombination hotspots.⁵¹  At the genome-average value, the largest shift anywhere in the table above is under one percentage point, whichever way the sites are split between G/C-requiring and A/T-requiring.⁵¹  In the 1–2% of the genome that sits in hotspots the effect is real and large, and this article's equilibrium numbers should not be applied there.

The bigger simplification is elsewhere, and it should be admitted.  The "3" in the formula assumes the three wrong letters are reached at equal rates.  Real mutation is nothing like uniform — transitions outnumber transversions roughly two to one, and CpG sites deaminate about an order of magnitude faster than anything else.  That approximation distorts the equilibrium more than gBGC does, and correcting it properly would require a full 4×4 substitution model rather than a two-state one.  Drift Lab does not model nucleotides at all: it stores one bit per site, working or broken, and the four-letter alphabet enters only as a 3:1 ratio of forward to back mutation.  That is a real limit on both the formula and the simulator, and it is the honest place to point a critic.

The flat left-hand end is the neutral answer: three of the four possible letters are wrong, so a site selection cannot see is wrong 75% of the time.  A site has to be worth more than about 0.01% of an organism's fitness before selection holds it in place.

So the equilibrium the objection points to is one in which every weakly-constrained functional site is wrong about half the time.  Genes do not work at that level of corruption.  The equilibrium is not a floor the genome rests on; it is the state of a genome that has stopped being a genome.

And it has now been measured in real genomes rather than argued about.  Latrille and colleagues analysed 28 mammalian populations across six genera, including humans, and reached a conclusion that supports this article's Argument 2 and complicates its Argument 3 in the same sentence:

"Our work confirms that deleterious substitutions are accumulating in mammals and are being reverted, generating a balance in which genomes are damaged and restored simultaneously at different loci."⁵²

The first half is the ratchet, observed in sequence data rather than simulated.  The second half is a bigger repair flux than this article's treatment of back mutation allows for: they estimate that beneficial non-adaptive mutations — reversions restoring a previously lost state — make up 15% to 45% of all beneficial mutations in 24 of their 28 populations.⁵²

That number deserves to be sat with rather than explained away.  If a third of beneficial mutations are repairs rather than innovations, the accounting question changes shape: the genome is not only being compensated, it is partly being fixed.  What it does not do is make the ledger balance by itself, because a reversion restores a site that was already broken — it does not build a new one, and the sites available to revert are exactly the ones the ratchet has already damaged.  Reversion is the mechanism that produces an equilibrium, which this section is about, not a mechanism that produces gain.

How long.  The waiting time for a site to be destroyed is about 1/μ generations — around 80 million generations at the human per-site mutation rate.  Nothing living is anywhere near equilibrium.  Every real genome is in the transient, and the transient is a straight line downward.

That is the honest answer to the equilibrium objection: technically right, practically irrelevant, and its own implications are worse than the problem it was raised against.

But there is a second brake, and it is a real one

The mutation-selection equilibrium above is not the only thing that can stop the decline.  There is a feedback loop, it is measured, and it is the strongest objection in this section.

As a population's fitness falls, the fraction of mutations that can improve it rises.  This is a consequence of Fisher's geometric model — a population far from its optimum has many ways to get closer — and it has now been measured about as sharply as anyone could ask.

Ardell, Martsul, Johnson and Kryazhimskiy measured the fitness effects of about a hundred insertion mutations across 42 yeast strains in six environments, and published the result in Science in 2024.  Their finding is that the scaling of mutational effect with background fitness is nearly the same in every environment, and it has a consequence they state plainly: "the proportion of beneficial mutations should increase with the decreasing fitness of the background strain, reaching up to 100% in the lowest-fitness genotypes."⁵³

Their raw counts make the same point from the other end.  Of the mutations they tested, "94% (88/94) of our mutations are beneficial in at least one strain and condition, and, of those, 96% (85/88) are also deleterious in at least one strain and condition."⁵³  Almost every mutation is both, depending on where it lands.

That is not an isolated result.  Couce and colleagues found the beneficial fraction in E. coli contracting from 6.8% in the ancestor to 3.2% after 15,000 generations of adaptation.⁵⁴  Silander and colleagues found the reverse in bacteriophage: high-fitness lines under drift declined, low-fitness lines improved, and both converged on the same middle.⁵⁵

Run that forward and the ratchet is self-limiting.  Decline raises the beneficial supply; the raised supply slows the decline; the system settles at a fitness where the two rates match.  That is not a terminal process, and this article should not pretend otherwise.

The reply is the one this whole argument turns on, and it is not a dodge: the feedback bounds fitness, not sequence.

Look closely at what Ardell's result says a "beneficial" mutation is.  The environment, in their model, "tunes one global parameter, the background fitness at which most mutations switch sign."⁵³  A mutation does not become beneficial because it built something.  It becomes beneficial because the background got worse — the same change that was harmful in a healthy genome is an improvement in a degraded one.  That is regression toward a mean, and the mean it regresses toward is falling.

The compensating mutations are somewhere else.  Nothing in the mechanism sends a repair to the site that broke — Hartl and Taubes showed formally that compensatory evolution "results in no change or permanent improvement in enzymatic function."⁴⁵  So the equilibrium the feedback produces is one where fitness has stopped falling and specific sequences are still turning over underneath it.  That is the same structure as the polygenic answer above, reached from a different direction, and the same reply applies.

And one feedback runs the other way

Both brakes above slow the decline.  There is one measured mechanism that speeds it up, and it is the only one in this subject.

Sharp and Agrawal found in PNAS in 2012 that the mutation rate in Drosophila is not a constant of the species but depends on the genetic quality of the individual.  Flies carrying deleterious alleles mutated faster than unimpaired controls, on chromosomes unlinked to the damage.  Their conclusion is the one sentence in this literature that describes an accelerator:

"individuals with low-quality genotypes will produce offspring of even lower genetic quality, in a mutational positive feedback loop."⁵⁶

The size is not trivial.  For genotypes showing a 10% reduction in body mass, they report roughly a 77% increase in the deleterious mutation rate and about a 104% increase in the rate of mutational decline.⁵⁶

What it does and does not establish.  It is one species, and the mechanism is plausible but not pinned down — a damaged genome may invest less in DNA repair, or suffer more oxidative damage.  It has not been measured in a mammal.  But it means the standard treatment, in which U is a fixed parameter of the species, is an approximation in the direction that flatters the population: a genome that has already degraded mutates faster than the one used to compute its future.  Every projection in this article, including its own, holds U constant and is therefore conservative if this result generalises.

Which means the honest version of this article's claim is narrower than the popular version.  Genetic entropy as "fitness declines to extinction" has a real mechanism working against it and probably does not happen in large populations.  Genetic entropy as "specific functional sequence is destroyed faster than it is built" survives the feedback, because the feedback does not act on that quantity at all.  Everything below is about the second claim.

Three results that point the other way

Before going further, three mainstream findings belong on the table, because they are the best evidence against everything above and a reader will find them eventually.

Mutation accumulation lines sometimes get better.  Bao, Melde and Sharp catalogue experiments in Arabidopsis, C. elegans, Daphnia, yeast and Chlamydomonas that show no fitness decline at all, and report Shaw and colleagues' conclusion that about half the mutations in their study were beneficial.⁵⁷

A two-sided fit puts most mutations on the positive side.  Böndel's Chlamydomonas analysis fitted a positive-effect frequency of 0.84.¹³  The same group went further in 2022, tracking 254 individual spontaneous mutations by how their frequencies actually moved, and reported that "as many mutations were observed to increase as decrease in frequency" — a positive-effect proportion of 0.50, with a confidence interval of 0.43 to 0.58.⁵⁸  Peter Keightley is senior author on both.  This is the most direct measurement in the list and it says the distribution is symmetric.

Beneficial mutations are commoner than the argument wants.  Joseph and Hall titled their yeast paper "more beneficial than expected" and measured 5.75%.⁵⁹  Firnberg's million-variant scan of β-lactamase found 7.0%.⁶⁰  Goyal and colleagues calculate that roughly 2% beneficial is enough to hold the line indefinitely against Muller's ratchet.⁶¹  Every measured fraction in that list clears the bar.

There are three legitimate replies and all come from mainstream sources.

The first is that the high beneficial fractions are probably an artefact, and this is now the mainstream position.  Crombie and colleagues broke the linkage between mutations in C. elegans mutation-accumulation lines by advanced intercrossing — the confound that makes every line's mutations inseparable — and concluded: "unless it can be shown otherwise, estimates of the distribution of fitness effect that infer a substantial fraction of mutations with positive effects are likely confounded by linkage disequilibrium."⁶²  Their verdict on the older picture is blunt: "the conventional wisdom — most mutations have effects near 0, a handful of mutations have effects that are substantially negative, and mutations with positive effects are very rare — is likely correct."⁶²

Their abstract reports roughly 40% positive effects in the full dataset before the correction, so a reader who stops partway gets the paper backwards — but its closing sentence states the conclusion plainly, and quoting the abstract in full is safe.

Second, mutation accumulation experiments are biased upward regardless, because selection cannot be switched off entirely during colony expansion.⁶³  And every one of those systems is assayed in a laboratory environment it is not adapted to — which is precisely the condition under which Fisher's model predicts an inflated beneficial fraction.  A population far from its optimum has plenty of ways to improve.  A well-adapted one does not, which is what Couce and Silander measured.

That second reply has a price, and it should be paid out loud rather than left for a critic to collect: it is the same mechanism as the feedback brake described above, pointed the other way.  If being far from the optimum inflates the beneficial fraction in a laboratory flask, then it also inflates it in a genome that has been decaying for ten thousand generations.  This article cannot use the mechanism to discount inconvenient measurements and then ignore it when it bounds the decline.  It does not get to have it only once, and the section above concedes it.

Two further concessions belong here, and both cost this article something.

Böndel's 2022 symmetry result has an answer, and it is the one their own authors give: the mean absolute effect they measure is "just over 2%," which is five hundred times the human drift barrier, and they note that loss-of-function mutations in genes not needed in laboratory culture may be advantageous there.⁵⁸  So the result is real and it does not reach the zone in dispute.  But "it does not reach the zone" is a reply this article has now used against several measurements, and a reader is entitled to notice that the zone in dispute is one where nobody has data at all.

And Crombie's linkage correction, quoted above in this article's favour, carries a second finding that runs against it: "almost all mutations have effects that are very close to 0, and … the decline in fitness with MA is the result of a small number of mutations with large negative effects — perhaps only one."⁶²  If the observed decline in mutation-accumulation lines is driven by a handful of large-effect hits, then those lines are a poor model for the genome-wide drizzle of near-neutral damage this article is actually about.  They cannot be used as evidence for it.

That is the honest state of it: the objection is strong, the replies are real, and the question is not closed.

Argument 3: the accounting

Here is where the disagreement actually lives, and it will not be settled by arithmetic alone.

Over six million years, at U = 5, the human lineage fixes about 333,000 harmful mutations.  What does it fix in the other direction?

Estimates of adaptive substitution in the human lineage are low.  Larry Moran, a biochemist at the University of Toronto and a prominent critic of intelligent design, was asked in 2014 how many of the roughly 46 million mutations separating humans and chimpanzees were adaptive.  His answer, quoted in full because the qualification matters: "We don't know how many of those were beneficial (adaptive) leading to ways in which modern chimps are better adapted than the common ancestor.  (Same for humans.)  My guess would be only a few thousand in each lineage."⁶⁴  It is a guess, he says so, and it should not be reported as a measurement.  Ian Musgrave, a pharmacologist at the University of Adelaide arguing against creationist claims about Haldane's dilemma, worked with a ceiling of 1,667 beneficial fixations and argued the true number is well below it.⁶⁵  Estimates of the proportion of amino-acid substitutions driven by positive selection in humans are among the lowest measured for any organism.

Set those side by side:

The ratio runs somewhere between 17:1 and 300:1 against.

One caveat on that table, raised by Felsenstein and fatal to a naive reading of it.  Adaptation does not run only on new mutations: "much of the variation that is being acted on by selection would not be new mutations, but would be existing alleles floating around in the population."³⁶  Standing variation appears in neither column.  The comparison above is between two flows of new fixations, and it ignores a third pool that selection draws on without waiting for mutation.  That does not dissolve the ratio — standing variation is itself the product of past mutation, and the question is what replenishes it — but anyone quoting "300 to 1" as the whole accounting is overstating what the table shows.

The best counter-argument, and what is actually wrong with it

The Masel group's answer is that this comparison uses the wrong currency.  Count effects, not events.  In their simulation the beneficial fixations averaged s = +0.0042 while the harmful ones averaged s = −0.000039 — a factor of about a hundred.²  A few thousand large gains can outweigh sixty thousand small losses.

Work their own numbers and something more interesting appears.  The beneficial flux is 6,387 × 0.00423 = +27.0 in log fitness.  The deleterious flux is 56,255 × 0.0000391 = −2.20.  Those do not balance.  The beneficial side is twelve times larger, and over the 200,000 generations simulated the population ends up something like 10¹¹ times fitter than it started.  Their own text confirms the direction: their reference case gains 10% fitness every 151 generations, indefinitely.²

So their model does not show a population breaking even against Ohta's ratchet.  It shows a population adapting without limit, with the ratchet as a modest tax on the way up.  That is only possible because the model has an inexhaustible supply of beneficial mutations of constant mean effect — nothing in it makes good mutations harder to find as the genome gets better.

Real genomes do not behave that way, and this is measured rather than argued.  In the long-term E. coli experiment the beneficial fraction contracted from 6.8% in the ancestor to 4.3% after 2,000 generations and 3.2% after 15,000, as the population adapted.⁵⁴  In bacteriophage, Silander and colleagues found the fraction of beneficial mutations was zero in every high-fitness population and greater than zero in every low-fitness one — high-fitness lines under drift declined, low-fitness lines improved, and both converged on the same middle.⁵⁵

That is genetic entropy, measured, in a mainstream journal — together with the mechanism that stops it.  A model without that feedback cannot locate the equilibrium.  It can only show that a parameter region with positive flux exists.

None of which is an answer to the information question, and the difference matters.

Their own example of how compensation works: "many proteins accumulate small deleterious mutations that slightly inhibit folding, which can be compensated for by a novel or improved or overexpressed chaperone protein."²

This is also where the claim that adaptation usually proceeds by breaking things has to be sourced rather than asserted.  Michael Behe surveyed four decades of laboratory evolution experiments in the Quarterly Review of Biology in 2010 and concluded that adaptive changes are dominated by loss or modification of existing function rather than gain.⁶⁶  Behe is an intelligent-design advocate, which is a reason to check his survey rather than to ignore it — and he states the limit himself: it "is called a 'rule' in the sense of being a rule of thumb.  It is a heuristic, useful generalization, rather than a strict law."⁶⁶  Jerry Coyne, reviewing it, granted that the observed laboratory changes are indeed mostly loss-of-function while disputing that this extrapolates to the long term.  For this article's purposes the granted half is the half that matters.  Take that literally.  A hundred proteins have been degraded, each at a specific position that used to hold a specific letter.  A chaperone is upregulated, and the organism's fitness recovers.

The hundred positions are still wrong.  The specific sequences are still gone.  What has happened is not repair but compensation — the genomic equivalent of a building with a hundred cracked beams and one very good jack.  The building stands.  The beams are still cracked, and next generation there will be a hundred more.

And it has been watched happen.  Denver and colleagues took a nematode line that had been bottlenecked for 323 generations until it lost nearly half its fitness, let five populations recover for 60 generations, and sequenced the whole genome before and after.  Fitness came back.  Nothing was repaired:

"We identified 68 base-substitution changes between N2 and MA12; in all 68 cases, the mutation was also detected in each of the three MA12-derived recovery lines.  Thus, no reversions of MA12 mutations were detected in the recovery lines."⁶⁷

"None of the recovery response is attributable to detectable reversion mutations, and no cases of putative intragenic compensatory mutations were identified.  Thus, intergenic compensatory mutations likely drive all of the change that we observe."⁶⁷

Every one of the 68 mutations that degraded the line was still there when it had recovered.  The repairs were at different loci from the breaks.  That is compensation rather than repair, measured in an animal, by people with no stake in this argument — and it is discussed in full, with its four substantial caveats, on the fitness-effects page.

This is not a creationist observation.  Hartl and Taubes made the same point in 1996, in a paper whose subtitle is the whole argument — "Selection without Adaptation."  Their abstract: "On a suitably long timescale... every gene incorporates favorable mutations that compensate for detrimental mutations previously fixed.  This form of evolution is driven primarily by natural selection, but it results in no change or permanent improvement in enzymatic function."⁴⁵

The body puts it harder still:

"in a metabolic system near saturation, as well as in a system with an intermediate optimum, most compensatory mutations merely restore, partially or completely, the level of enzyme function to that of the ancestral molecule.  There will usually be no overall increase in absolute fitness.  On this treadmill, each molecule evolves adaptively but does not improve."⁴⁵

Selection is running. Fitness is being maintained. Nothing is being built. That is the accounting argument, in Journal of Theoretical Biology, by a Harvard population geneticist, in 1996.

H. Allen Orr put the general version the same year, reviewing Daniel Dennett, and put it better than this article can.  Natural selection, he wrote, "'sees' only brute birth, death, and reproduction, and knows nothing of Design," and is "just as happy to lay waste to the kind of Design we associate with engineering as to build it."  His example is the cave amphipod Gammarus minus, which retains an optic nerve with little or no eye: "If eyes are expensive to make, selection can wreck their exquisite engineering just as surely as it built it."⁶⁸  That is the accounting argument, stated by a mainstream evolutionary geneticist who was attacking intelligent design at the time.

This is why the argument cannot be settled by fitness alone.  Fitness is a measure of how well you compete today.  The question genetic entropy asks is whether the process that is running can have built the specific sequences in the first place, and for that you have to count sequences, not competitive success.

Kondrashov already answered this, in 1995, and the answer is the article's own

The Masel resolution is that a smaller number of larger-effect beneficial fixations counterbalance the deleterious ones.  That is not a new proposal.  It is Kimura's, and Kondrashov named it and rejected it in the paper that defined the paradox:

"According to Kimura (1983: 248) VSDMs do not cause any problem because (i) they accumulate very slowly and (ii) their impact can be easily counterbalanced by rare fixations of beneficial alleles.  I do not think that this is correct."³

His reason is a constraint on where beneficial mutations can even occur, and it is the sharpest statement of this article's argument that exists anywhere:

"A beneficial mutation is possible only in a site occupied by a suboptimal nucleotide.  This implies that either a VSDM was previously fixed at this site (in this case the beneficial mutation can have only a very small advantage) or that the environment has changed (then the mutation can be significantly beneficial)."³

Read what that does to the balance sheet.  A beneficial mutation has somewhere to happen only if a site is already wrong.  If it is fixing a site that a very slightly deleterious mutation broke, its advantage is by definition very slight — you cannot recover more than was lost.  The only way to get a large beneficial effect is if the environment moved, and that is a lag load being paid down, not a repair of the genome.

So the Masel model's parameters are in tension with the mechanism.  Their resolution needs beneficial fixations averaging about a hundred times the effect of the deleterious ones.²  Kondrashov's constraint says the repairs available to a degraded site are the same size as the degradation, and anything bigger is coming from environmental change rather than from fixing what broke.  Neither paper cites the other on this point, and it is the crux.

Ohta had described the same pairing sixteen years earlier and got the size right: "it is likely that a slightly deleterious base substitution is followed by a slightly advantageous (compensatory) substitution."³⁰  Slightly deleterious, then slightly advantageous.  Nobody in this literature who works the mechanism through arrives at large compensating effects.

The honest counter, which their side would make.  Compensation need not occur at the broken site — a better chaperone can offset a hundred misfolded proteins at once, which is their own worked example.  That is a large-effect beneficial mutation that is not repairing any of the broken sites.  Kondrashov's constraint bounds repair; it does not bound compensation.

Which is exactly the distinction this whole article turns on, arrived at from the opposite direction.  Fitness can be rescued by compensation of unbounded size.  Sequence can only be restored by repair, and repair is bounded by how small the original damage was.  Kondrashov saw the fitness half of this in 1995; the sequence half is what remains unmeasured.

Two more sentences from the same paper belong here, because both anticipate objections raised against this article thirty years later.  On whether an equilibrium rescues the situation: "the assumption that there is an equilibrium contamination can only underestimate the impact of VSDMs."³  And on the lag-load complaint that Argument 1 concedes: "If all individuals indeed carry many lethal equivalents of VSDMs, the perfect mutation-free genotype is rather different from the best available one, even if this does not directly translate into high mutation load."³  He is conceding the load point and keeping the contamination point — which is this article's structure exactly.

The information objection, stated by the people best placed to state it

There is a formal mainstream treatment of exactly this question, and it does not go the way this article would prefer.

Hledík, Barton and Tkačik asked in 2022 how much information a population can accumulate and hold, measured properly — as the divergence between the actual distribution of genotypes and what drift alone would produce — and how much fitness cost that maintenance charges.  Their result is the opposite of the intuitive one:

"Selection is most efficient when it is relatively weak (Ns≪1), maintaining up to N/μ(1+4Nμ) bits per unit cost at the genotype level, and inefficient when strong (Ns≫1), maintaining only about ln(2)/μ bits per unit cost."⁶⁹

Weak selection — the regime just at and below the drift barrier, the regime this article has spent ten thousand words calling a blind spot — is where information is held most cheaply.  And they add that capacity is not the binding constraint: "Typical populations have Ne/μ much greater than the genome size, suggesting that the genome size or other factors are more limiting."⁶⁹

Three things about this.

It is right, and it is a genuine problem for the crude version of the argument.  Anyone claiming that sub-drift-barrier sites carry no information, or that the genome's information capacity is exhausted, is contradicted by a careful PNAS paper.  This article does not make that claim and should not be read as making it.

But their information is frequency information, not sequence identity.  Their measure asks how far a site's allele frequency has been pushed away from the neutral expectation.  A site sitting at 60/40 instead of 50/50 contributes a fraction of a bit.  That is a real and well-defined quantity, and it is not the quantity in dispute.  The question here is whether a splice site still reads GT, not whether its frequency distribution is skewed.  A population can be rich in the first kind of information while losing the second, and their framework does not distinguish the cases.

And they say the limit themselves.  Their own closing caution is that "more work is needed to make claims about the information content of any real organism's genome."⁶⁹  That is the correct position, and it is this article's position too.  Nobody has done this measurement on a real genome.  Until somebody does, the accounting argument is an open question with a strong prior on each side and no decisive data — which is what the last section of this article says.

The measurement everybody skips: beneficial how?

Here is a gap that runs through this entire literature, including the parts of it this article relies on.

Every number quoted for "the beneficial fraction" is a fitness measurement, and fitness measurements are blind to mechanism.  When Firnberg's scan reports that 7.0% of β-lactamase missense mutations raise fitness, the measured quantity is growth under ampicillin.⁶⁰  Nothing in that assay distinguishes a mutation that improved the enzyme's catalytic machinery from one that broke a regulatory brake, disabled a competing pathway, or dumped a protein the cell was paying for and did not need in a flask.  Couce's E. coli scan, Böndel's Chlamydomonas fits, and the mutation-accumulation lines are the same: fitness in, fitness out.^{54 13}

That matters because Argument 3 is not a claim about fitness.  It is a claim about whether specific functional sequence is being built as fast as it is destroyed.  A beneficial-mutation rate cannot be converted into a construction rate without knowing what fraction of those beneficial mutations built anything — and that fraction is not in the papers the rate comes from.

What has been measured points one way, and it is not a creationist result.  Hottes and colleagues screened transposon libraries and found that "even under conditions of extreme nutrient limitation, substantial adaptation can be achieved solely through loss of function mutations, which rewire the metabolism of the cell without gain of enzymatic or sensory function."  Across their survey, "at least one beneficial null mutation was identified in all but five of the 144 conditions considered."⁷⁰  Adaptive destruction is available almost everywhere anyone has looked.  A 2024 meta-analysis of 58 studies and 65 evolution experiments concludes that "loss-of-function mutations play a basic role in rapid adaptation to ecological conditions."⁷¹  Michael Behe's survey of laboratory evolution reached the same conclusion earlier and was granted on the observation by Jerry Coyne, as discussed above.

Now the honest limits, because this point is routinely overstated.

The meta-analysis found nonsense substitution significantly exceeding synonymous substitution in 17 of 65 experiments, not in most of them — and its own summary of the aggregate is a null: "the rate of accumulation of nonsense mutations and synonymous mutations was approximately equal during short-term evolution (dNONSENSE ≃ dS)."⁷¹  That sentence sits in the body of a paper whose title is "Loss-of-function mutations are main drivers of adaptations," and anyone citing the title owes the reader the sentence.

So "loss of function is a major route in short-term lab adaptation" is supported.  "99% of beneficial mutations are destructive" is not a number anybody has measured, and stating it would be inventing precision the literature does not have.

Not all adaptive change is destruction, and two of the counter-examples come from inside the argument's own camp.

Regulatory mutations that retune expression are not obviously breaking anything, and the one study that sorts adaptive mutations by locus finds the answer depends on which gene you look at.  Lang and colleagues sequenced forty evolving yeast populations.  Their adaptive drivers overall are "enriched for nonsense and frameshift mutations," which is the destruction thesis holding.  But the spectrum splits by function.  At mating-pathway and Ras-regulator loci they saw 14 missense, 8 nonsense and 10 frameshift, "suggesting that selection at these loci is for loss of function."  At cell-wall assembly loci all 13 mutations were missense, "suggesting that at these loci selection is for alteration or attenuation, not loss, of function."⁷²

That is the shape of the real answer.  Destruction is a large and probably dominant share of short-term adaptation, and it is not the whole of it.  Which one you get depends on the gene.

And the E. coli long-term experiment's most celebrated event — aerobic citrate use — arose when a duplication placed an existing transporter under a new promoter.  By Michael Behe's own published definition that is a gain of function, since his criterion for gain-of-FCT is "the construction by mutation of a new promoter."  It is the single clearest counter-example to the destruction thesis, and it should be conceded by anyone advancing that thesis rather than left for a critic to produce.

Genuine construction from nothing is rarer but documented base by base.  The antifreeze glycoprotein of northern cod was traced to its origin, and the authors state the result without hedging: "We experimentally proved that all genic components of the extant gadid AFGP originated from entirely nongenic DNA."⁷³  The same function evolved independently a second time in Antarctic fish, from a digestive enzyme gene.  Two independent origins of a real innovation.

And gene loss is not always a dead end.  Helsen and colleagues deleted genes in over 200 yeast lineages and found that loss "can facilitate adaptation by opening alternative evolutionary paths," with some deletion strains eventually beating adapted wild-type.⁷⁴  Destruction that increases evolvability is a real mainstream answer, and it has to be counted.

And when the "less is more" literature is asked for a fraction rather than a slogan, the fractions are small.  This is the part that never gets quoted, so it is quoted here.

Xu and colleagues identified 60,819 loss-of-function variants across 1,071 Arabidopsis genomes — the largest such survey there is.  Their adaptive yield: "in the Yangtze River basin population, 1% of genes with LoF mutations were under positive selection," and "0.9% of LoF variants with minor allele frequency larger than 0.5% are associated with climate change."⁷⁵  The review built on that work is titled Less Is More, and concedes in its own abstract that "although most LoF mutations are neutral or deleterious, at least some of them are under positive selection."⁷⁶

The cleanest adaptive-versus-neutral denominator in the field comes from whales.  Huelsmann and colleagues screened the cetacean stem lineage and "found 85 gene losses," of which about eleven are given an adaptive story.  Their own summary of the rest: "Many of these gene losses were likely neutral, and their loss happened because of relaxed selection to maintain their function."⁷⁷

On crop domestication the literature flatly contradicts itself.  Gross and Olsen, counting the same genes, write that "most of the domestication alleles characterized to date are not loss-of-function mutations."⁷⁸  Anyone citing domestication as a showcase of adaptive destruction has to deal with that sentence first.

A caution that cuts against the comparative-genomics gene-loss numbers generally.  Natsidis and colleagues fed simulated data containing zero real gene loss into the standard orthology pipelines and reproduced the published pattern anyway, concluding that such results "may well have been based on a misleading signal."  In fairness the same critique damages the gain counts — homology detection failure inflates those too — so the net effect on any loss-to-gain ratio is unknown rather than known to favour either side.⁷⁹

Albalat and Cañestro's review poses this article's exact question — what fraction of adaptive mutations are loss-of-function? — declines to answer it for want of data, and reports that in multicellular life "the fixation of most gene losses is driven by neutral evolution."⁸⁰

The strongest form of the argument, then, is not a percentage.  It is that the field measures the wrong variable for this question and then uses the answer as though it were the right one.  Until somebody reports a distribution of fitness effects with the beneficial side classified by mechanism — repair, retuning, deletion, duplication, genuinely novel function — the beneficial rate that every balance calculation in this dispute depends on cannot tell you whether the genome is being built or merely re-tuned while it erodes.  That measurement is achievable with existing technology.  Nobody has published it.

The construction side, measured by the other camp

The article's own arithmetic covers the destruction rate.  The best quantitative attempt at the construction rate comes from Sanford's group, and it is peer reviewed in a secular journal, so it belongs here on its merits rather than as a target.

Sanford, Brewer, Smith and Baumgardner simulated a pre-human hominin population of at least 10,000 with strong selection — half the population eliminated each generation — and asked how long it takes to establish a specific short string of nucleotides.  Two nucleotides took an average of 84 million years; five took 2 billion; eight took 18 billion, longer than the age of the universe.⁸¹

Two honest caveats.  The model asks for a pre-specified target string, and evolution does not work toward targets — a critic will say the waiting time for "some useful string" is far shorter than for "this string," and they are right in principle.  And their simulator is the one discussed above.  But the qualitative point survives both objections and is not seriously disputed: multi-nucleotide functional elements that require several specific changes before conferring any benefit are extremely slow to arise, and that is the side of the ledger evolution has to fill.

What would settle it

Not the beneficial mutation rate.  That is worth stating carefully, because it is the thing creationist sources usually attack and they are attacking the wrong number.  Masel's reference value of U_b = 0.002 works out to about one beneficial mutation per 35,000 — which is between 300 and 2,000 times lower than every direct measurement of the beneficial fraction, which run from 1% to 7%.^{60 54 82}  Their assumed rate is conservative, not optimistic, and their own analysis says the rate is not the binding constraint anyway: "The mean beneficial effect size s̄_b is more important than the beneficial mutation rate U_b."²

What would settle it is a harder question, and it is the one nobody can currently answer:

What is the mean effect of the beneficial mutations actually available to a genome that is already well adapted, and does that supply contract as fitness rises?

Couce and Silander both measured a contraction.  Neither measured it in a large-genome sexual species over evolutionary time.  Masel's own paper concedes the uncertainty: "there is great uncertainty in the true values of U_b and s̄_b."²

And the reason nobody has a better number is not neglect.  It is that the measurement cannot currently be made, and the people who build the instruments say so.  From Dmitri Petrov's laboratory, in 2023:

"Given the importance of measuring microbial fitness in diverse fields, it is frustrating that we cannot measure fitness more precisely.  Reported precision remains capped at detecting fitness differences on the order of 0.1–1%."⁸³

And four sentences later, they state this article's question in their own words:

"natural selection can distinguish fitness differences that are orders of magnitude smaller than we can measure in the laboratory, depending on effective population size.  This creates a problem for scientists: how do we understand the fitness effects of mutations if many of these effects are too small for us to measure precisely?"⁸³

A related limitation appears in one inference from human sequence data.  Johri, Pfeifer and Jensen tested how much positive selection human exome data can rule in or out, and concluded: "although strong and/or frequent recurrent positive selection is inconsistent with observed data, weak to moderate positive selection is consistent but unidentifiable if rare."²⁶  With 0.1% or 1% of new mutations beneficial at 2N_e s_b of 10 or 100, "there is almost no difference between the distribution of statistics across the 465 exons in the absence versus presence of positive selection," which they say "suggests a general inability to identify this class of mutations, if present."²⁶

That test leaves the specified rare beneficial classes difficult to distinguish from their absence.  It does not establish the beneficial rate, whether beneficial effects balance harmful ones, or what a different method could identify.

And there is a second missing measurement, which is easier than the first and which nobody has run:

Of the beneficial mutations that do occur, what fraction build functional sequence rather than remove it?

Every beneficial fraction in this dispute is a fitness number with no mechanism attached.  Classifying the beneficial side of a deep mutational scan into repair, retuning, deletion, duplication and novel function is achievable with current sequencing and structural tools.  Until it is done, the construction rate that Argument 3 turns on has never actually been measured — only a proxy for it that cannot distinguish building from pruning.

What the experiments show

Theory is cheap.  This section is what happens when people actually stop selection and watch.

Mutation accumulation lines

The direct test is to breed organisms through a bottleneck so tight that selection can barely operate — a single individual or a single brother-sister pair per generation — and measure fitness against a frozen ancestor.  Anything that accumulates has to be mutation.

The results, honestly reported, are a mess, and the mess is the point:

Three things follow.

First, fitness does decline in animals under relaxed selection, and by a lot more than in microbes.  Flies lose a fraction of a percent per generation, and Shabalina's competitive assay — which measures fitness the way nature does, against rivals, rather than in a comfortable vial — found 2% per generation, ten times the benign estimate.⁸⁴  That gap between "benign" and "competitive" measurement recurs throughout this literature and it always runs the same way.

Second, the estimates disagree by a factor of 700.  Katju and Bergthorsson's review of every eukaryote MA study puts the range of measured genomic deleterious rates at 0.00065 to 0.47 per generation.⁸⁵  Anyone quoting a single figure from this literature — in either direction — is choosing.

Third, and this cuts against genetic entropy: degraded lines recover fast.  Estes and Lynch took C. elegans lines that had been degrading for 240 generations, put them back in a large population under selection, and measured full restoration of fitness in under 80 generations — a recovery rate "at least three times that of mutational degradation."⁸⁶  Katju's 2015 experiment found that at population sizes of 10 or 100 rather than 1, there was no significant decline at all over 409 generations.⁸⁷

That last result is the strongest empirical evidence against the strong form of genetic entropy, and it should be stated without softening.  Selection at modest population sizes stopped the decline.  What it does not settle is the fate of the mutations below the blind spot at that population size, which is where the whole argument lives — a 409-generation experiment at N = 100 cannot see a process that operates on the scale of 4N generations and effect sizes of 1/(2N).

Error catastrophe in viruses

Raise a virus's mutation rate with a drug and it dies.  This is not disputed and it is the basis of real antiviral therapy.

Crotty, Cameron and Andino gave poliovirus ribavirin at 1,000 μM, raised the mutation rate 9.7-fold, and measured a 99.3% loss of genome infectivity in a single round of replication.⁸⁸  Loeb and colleagues drove HIV to extinction in 7 of 9 cultures with a mutagenic nucleoside, against 0 of 28 controls.⁸⁹  Favipiravir cleared norovirus from 7 of 9 treated mice.⁹⁰  Mutator mice — carrying a defective DNA polymerase, with a 17-fold elevated mutation rate — showed 4.1 times more abnormal phenotypes, dropping fertility to 36% of baseline, and many breeding lines died out.⁹¹

But the honest reading of this literature is more limited than creationist summaries usually admit, in four ways.

The mechanism is not what it is usually called.  Bull, Sanjuán and Wilke showed that killing a virus with mutations is not Eigen's "error catastrophe" — it is simply lethal mutation accumulation, a different and less exotic thing.⁹²

The rates required are enormous.  A 2025 analysis across six viruses concluded that lethal mutagenesis needs mutation-rate increases of 11-fold to 2,700-fold, while actual drugs deliver at most about 10-fold: "lethal mutagenesis seems difficult to achieve."⁹³

Molnupiravir did not sterilise SARS-CoV-2, and it did not obviously drive it either.  Sanderson and colleagues found branches carrying 20 or more mutations with the drug's characteristic signature, including a 20-tip transmission cluster in Australia.⁹⁴  Raising the mutation rate produced viable, transmissible descendants — which is the point that matters here.  It should not be stretched further than that: no molnupiravir-derived lineage has become a named variant, and the honest summary is that heavily mutated lineages survived and went nowhere in particular.

And there is a clean counter-experiment.  Springman, Keller, Molineux and Bull mutagenised phage T7 at 100 to 1,000 times its normal rate, giving 4.3 non-lethal mutations per genome per generation.  Theory predicted fitness would fall from 18.3 to 8.6 doublings per hour.  Instead, after 200 generations, fitness rose to 21.9.  The authors write that the "failure of the theory challenges the quantitative basis of lethal mutagenesis."⁹⁵

That last result is the single best experimental argument against genetic entropy and it belongs in any honest treatment.  The most likely explanation is arithmetic: at 2.6 deleterious mutations per generation, a virus producing more than e^2.6 ≈ 13 copies of itself will, on average, produce one with no new mutations at all.  Selection can then pick that one.  Viruses have burst sizes in the hundreds.  Mammals have litters.  The difference between a phage and a human here is not a detail — it is the whole argument, and it runs in genetic entropy's favour rather than against it.

The human data

The strongest single result is also the most misused.  Fu and colleagues sequenced 15,336 genes in 6,515 people and found that 86% of the variants they classed as deleterious arose within the last 5,000 to 10,000 years, with a mean age of 5,200 years for European Americans.⁹⁶

And there is a second thing creationist material never quotes: how the 5,000-year figure was obtained.  It is not a measurement of variant ages.  It is the output of a fitted demographic model — specifically "a modified Out-of-African model in which accelerated population growth began 5,115 years ago with a per generation growth rate of 1.95% and 1.66%" for the two ancestry groups.⁹⁶  The ages come out of the model; they do not go into it.

To their credit the authors tested how much that matters, and report that estimates "were generally robust across different demographic models, with the largest discrepancies resulting in a two-fold difference in average age across all SNVs."⁹⁶  That is a real robustness check and it should be reported alongside the criticism.

But notice what it does and does not cover.  Every model in the family they tested is a deep-time out-of-Africa history.  Robustness within that family is not robustness of that family.  Anyone using this result while rejecting the demographic history is quoting a number that was manufactured by the history they reject — which is the distinction drawn in which numbers are measurements above, and this is the cleanest example of it in the whole article.

Creationist material quotes this constantly.  It should also quote the authors' explanation, which is not genetic entropy: human populations exploded in size recently, and a rapidly growing population accumulates rare variants faster than selection can purge them.  Fu's own model fits the acceleration starting 5,115 years ago with per-generation growth of about 1.95%.⁹⁶  The recent origin of deleterious variants is a demographic signature.  It is consistent with genetic entropy but it is not evidence for it over the mainstream explanation.

What the rest of the human data shows:

• The average person carries about 100 genuine loss-of-function variants, with about 20 genes completely inactivated.⁹⁷
• 7.7% of Icelanders are complete knockouts for at least one gene, across 1,171 genes.⁹⁸
• Every person carries roughly 400 highly damaging missense variants and around 2 bona fide disease mutations.⁹⁹
• Neanderthals carried a mutational load implying median fitness around 0.63 of modern humans, and the Altai Neanderthal's genome scores worse than 97% of present-day people.^{100 101}

And the counter-evidence, which is real and cuts both ways.  The same ancient-DNA study reports two findings that do not sit comfortably together.  Genomic health improved from the deep past toward the recent past — the opposite of a simple "decaying since the beginning" model.  But individuals from the last few thousand years score healthier, on average, than people alive today.¹⁰¹  The authors themselves note the pattern is "perplexing."  Neither side gets a clean result out of this, and any article claiming otherwise is picking half the paper.

On measurable decline in modern populations: Kong and colleagues found that polygenic scores for educational attainment declined in Iceland by about 0.010 standard units per decade of birth year across 129,808 people — real, statistically solid, and small: an implied 0.30 IQ points per decade against a Flynn effect of +3.0 points per decade running the other way.¹⁰²  Beauchamp found selection against education in the United States at about 1.5 months per generation.¹⁰³  These are genuine measurements of ongoing genetic change in a fitness-relevant trait.  They are also two orders of magnitude too small to be visible in a human lifetime, and they measure selection under modern conditions, not the deep past.

The forecast from the drift-barrier school is larger than either measurement.  Michael Lynch, who holds that most of the human genome is junk and so cannot be accused of inflating U, worked the cost of relaxed selection in 2010:  "with a complete relaxation of natural selection, the expected decline in fitness associated with mutations in coding DNA alone appears to be on the order of 1% to 3% per generation."¹⁰⁴  That is the same size as the Crow figure Sanford uses, reached independently from mutation data.  Two limits come with it, both Lynch's own.  It is a forecast for industrialized societies under relaxed selection, not a measured decline.  And because the human population is now so large, the new harmful alleles "will remain at very low frequencies for many centuries, so the damage incurred under the scenario outlined earlier need not be permanent."¹⁰⁴  So a mainstream estimate does put the cost of fully relaxed selection at 1 to 3% of fitness per generation.  Selection is not fully relaxed, as the results below show, and the cost Lynch describes is rare harmful variants piling up, which is reversible.  Argument 2 is about fixations, which are not.

Three results published since 2022 cut hard against a simple decline picture, and they belong here rather than in a footnote.

Selection is still running on living people.  Chen and colleagues analysed 509,817 variants across 72,635 Han Taiwanese and looked for allele frequencies that shift with age — the signature of selection acting now.  They "detected 168 variants deviating from neutrality, with most showing declining frequencies in younger generations, consistent with purifying selection on deleterious alleles influencing disease risk."¹⁰⁵  That is purifying selection observed within living memory, in a population with universal health coverage.

And it has a mechanism medicine does not fix.  Gardner and colleagues found in Nature that variants which ablate protein-coding genes are "associated with markedly reduced reproductive success, primarily owing to increased childlessness," and that this "may account for 20% of purifying selection against heterozygous variants that ablate protein-coding genes."¹⁰⁶  The route is behavioural — carriers are less likely to find a partner — which no clinic addresses.

And over the Holocene, the trend ran the other way on the traits people worry about.  Akbari and colleagues measured selection across 15,836 ancient West Eurasians spanning 18,000 years and found "negative selection on combinations of alleles that today increase body fat percentage," against psychosis-associated alleles, against "alleles contributing to overall health decline," and for higher scores on cognitive and educational predictors.¹⁰⁷

How much do these cost the argument?  Less than they look, and the reason is the same one this article keeps arriving at.  Every one of them measures selection where selection can be seen.  Akbari's standard error on a selection coefficient is about 0.17%.  Agarwal and colleagues put the median cost of losing a gene copy at 1.8%.¹⁰⁸  Chen's method needs an effect large enough to move an allele's frequency across age groups within one lifetime.  The drift barrier for humans is 0.005%.

So these results establish something worth granting plainly — selection has not been switched off, and the popular claim that modern medicine has abolished it is false — without touching the band the argument is actually about.  None of them measures the sign of selection on a mutation costing less than 1/(2N_e).  That is not a criticism of the studies; it is a statement about what is currently measurable.

One honest complication in the other direction.  Two of the three depend on polygenic scores built in modern industrialised populations, and Akbari's authors say so themselves: those traits "are only relevant to modern societies and would have been unmeasurable in preliterate societies over the vast majority of the period during which selection acted."¹⁰⁷  They also note the behavioural signals largely stop about 2,000 years ago, and that the educational trend has reversed in Iceland over the past century — which is the Kong result quoted above.  The sign of selection on a measured trait is not a fixed thing.

Small populations: the cleanest natural experiment

Conservation genomics has been running the experiment this argument needs, on species that went through severe bottlenecks, and the result is a two-tier pattern both sides should find interesting.

The best single case is the saola, a critically endangered Vietnamese bovid whose genomes were published in Cell in 2025 by 34 authors with no stake in this debate.  Purging removed roughly two thirds of the genetic load the species would otherwise have carried -- and the load that actually gets expressed rose anyway over the last 10,000 generations.  Their own summary: "genetic drift leads to an increase in the mildly and slightly deleterious mutations, while purging reduces overall genetic load by removing the strongly deleterious recessive mutations."¹⁰⁹

Two more results sharpen it.  A 2024 study of 502 European barn owls found island populations enriched in every category of deleterious allele with no purging at all, and concluded that only populations "with an effective population size of a few hundred or less can undergo purging" -- while adding that such populations "will usually also incur an additional cost of increased lowly and mildly deleterious mutations."¹¹⁰  And the Florida panther's celebrated genetic rescue did not remove deleterious variants: loss-of-function alleles per individual actually increased, while homozygous ones fell.¹¹¹  The rescue worked by masking, not repair, and the authors' simulations have fitness declining again within twenty generations.

The counterexample, stated first.  A 2025 Heredity paper reports long-term purging of non-synonymous mutations as a class in North Atlantic killer whales, not only the strongly deleterious ones.¹¹²  If that holds it reaches into the tier this argument says selection cannot clean.  Three caveats: "non-synonymous" lumps strong and mild effects together, the authors attribute part of the signal to linked-selection artefacts rather than clean purging, and no fitness was measured.  The full text is embargoed until December 2026, so the published numbers cannot currently be checked.

And the strongest sentence against the whole thesis comes from the modellers whose distribution of fitness effects the saola study used: "empirical evidence for mutational meltdown occurring in wild populations remains scarce, thus its importance as a genetic driver of extinction remains in question."¹¹³  They are right about meltdown specifically, and an article claiming otherwise would be overstating it.

But there is one cross-taxon empirical test, and it is co-authored by the man who named the paradox.  Polishchuk, Popadin, Baranova and Kondrashov took 211 mammal species, used the ratio of non-synonymous to synonymous substitution in mitochondrial DNA as a genetic proxy, and asked how much of extinction risk it accounts for once the dominant ecological predictor is controlled for.  Their answer: "Ka/Ks and body mass are responsible for a 38% and a 62% increase in extinction risk, respectively," with genetic factors contributing roughly one quarter to one half of the combined effect.¹¹⁴

Three caveats, all theirs or obvious.  Ka/Ks is a proxy for accumulated mildly deleterious substitution, not a measurement of load.  It is mitochondrial, so it says nothing directly about the nuclear genome where most of the argument lives.  And a correlation across species is not a demonstration of mechanism in any one of them.

What it does establish is that the honest answer to "no field evidence" is "less than you would want, but not none" — and that the one study that looked found a genetic signal roughly half the size of the strongest ecological one.

A 2026 result splits the difference in a way that is bad for the popular creationist reading and interesting for this one.  Mawass, Hernández, Matheson, Berg and Masel — the same lab as the fixation counts above — asked what actually drives a small population into an extinction vortex, and separated two candidates: mutational meltdown, where deleterious mutations fix, and mutational drought, where a small population simply generates too few beneficial mutations to keep adapting.  Their result:

"In stable environments, we find that mutational drought is nearly as significant as mutational meltdown.  Drought is more important than meltdown when populations must also adapt to a changing environment unless the beneficial mutation rate is extremely high."¹¹⁵

Read that both ways, because it cuts both.  It undercuts pointing at bottlenecked species and calling their decline a demonstration of mutation accumulation — at best that is half the story, and in changing environments less than half.  And it makes the supply of beneficial mutations the quantity that decides whether a population survives, which is exactly the parameter Argument 3 says nobody has measured in a large sexual species.  Two independent lines of work now converge on the same missing number.

Where the creationist case is weak

An article that only listed the other side's problems would be worth nothing.  Here are ours.

Almost all of them are in one book, and repeating the audit here would only invite the two copies to drift apart — which is exactly how a misquotation got into this article once already.  The full page-by-page treatment is a separate article.  This is the summary.

Seven documented errors in Genetic Entropy, the book that made this argument known, and in the work that grew from it.  The first five are in every edition of the book;  the sixth is in a 2015 follow-up paper;  the seventh is in the 2014 edition's appendix.

1. A key quotation is cut where it becomes inconvenient. Sanford prints Eyre-Walker and Keightley's 1999 Nature sentence ending "...it is doubtful that such species could survive," with the condition removed. The original ends "could survive if mutational effects on fitness were to combine in a multiplicative way," and the authors' next sentence says their results indicate they do not. A rejected premise is converted into a bare prediction of extinction.
2. The headline decline rate is run backwards through history. Figure 4 projects Crow's 1% per generation over 300 generations. Crow's number is an extrapolation from Drosophila, located explicitly in "the past few centuries" and "the wealthy nations," and he says in the same paragraph that it "is more than compensated for by much more rapid environmental improvements."
3. The heritability of fitness is the wrong quantity. Three chapters rest on h² = 0.004, cited to Kimura. Kimura's figure is a broad-sense H², from one Drosophila study, deployed by him as an objection to a position Sanford is arguing for. And the book never says what its 1-in-250 measures: 250 is right for the share of selection passed on (R/S = h²), while the response in additive-genetic standard deviations scales with the square root (R/σ_A = i·h) and costs a factor of about 16. Neither the book nor its critics named the currency. A direct measurement now exists and lands between the two: about 3% across 19 wild vertebrate populations.⁴⁰
4. A number is attributed to the wrong person. Sanford writes that Kimura "states that only 138 sites can undergo selection simultaneously." The 138 is Maynard Smith's, for a model he was arguing against; the figure he advocates on the same page is 25,500.
5. The most-reproduced figure assumes its own conclusion. Figure 3d's mirrored beneficial hump is asserted — "a reverse image of Kimura's curve" — with no source. Kimura's axis is labelled "Selective disadvantage" and runs entirely on the harmful side, so there is nothing to mirror.^{14 14} Reflect the deleterious curve and every beneficial mutation is tiny by construction, therefore inside the box, therefore unselectable. The figure does not demonstrate that; it draws it.
6. The long-term E. coli experiment is read backwards. Christopher Rupe and Sanford's 2015 paper, published by Sanford's own foundation, calls Lenski's experiment "one of the most powerful demonstrations of genetic entropy and de-evolution."¹¹⁶ Its Figure 1 caption assigns the unbounded power-law curve to the claim that fitness "has reached a near maximum," and the bounded hyperbolic curve to Lenski.¹¹⁶ Both attributions are reversed: Wiser, Ribeck and Lenski's 2013 abstract says "Mean fitness appears to increase without bound, consistent with a power law."¹¹⁷ Fitness kept rising past 50,000 generations, in the mutator lines as well.
7. The cap on beneficial mutations is argued in one currency and used in another. Every Mendel's Accountant run caps how much a beneficial mutation can raise fitness, at 0.001 or 0.01, and the declines depend on that cap. The only defence of it, in the 2014 edition's Appendix 5, is about information: a beneficial mutation "should only increase information by about one part in 3 billion."¹¹⁸ The same appendix says an antibiotic-resistance mutation raised fitness "infinitely, going from zero to one" while losing information.¹¹⁸ By Sanford's own reasoning, then, the information cap does not limit fitness gains, and it is the fitness number that goes into the simulator.

Four editions across nine years fixed none of the first five.  The only textual changes are a widened parenthesis, one added clause, and "even as low as" softened to "perhaps as low as."

The sharpest criticism of the method comes from inside the intelligent-design camp.  Douglas Axe, publishing his own genetic-entropy result in 2023, writes that in models of this kind "the statistics that come out of these models are dictated by the statistics that went in," and notes that a Nelson and Sanford run reverses itself once large beneficial mutations are allowed to occur.  He accepts the conclusion and rejects the evidence for it.

And the movement has published almost no original work on its own flagship argument.  The Creation Research Society Quarterly has run for sixty-two years; a full-text search of its archive finds five papers that mention genetic entropy at all, every one citing Sanford's book in passing.  No creationist journal has published a measured or simulated distribution of fitness effects.  The quantitative work lives in conference proceedings, one symposium volume, and three papers in a journal that ceased publication in 2021.

The influenza argument is the weakest part of the case

Carter and Sanford's 2012 paper argues that human H1N1 influenza has been degrading since 1918 and went extinct twice.  In their words:  "We document multiple extinction events, including the previously known extinction of the human H1N1 lineage in the 1950s, and an apparent second extinction of the human H1N1 lineage in 2009.  These extinctions appear to be due to a continuous accumulation of mutations."  By 2009 the lineage had gathered "over 1400 point mutations (more than 10% of the genome)," at a steady 14.4 a year.¹¹⁹  They grant that "Herd immunity is undoubtedly an important factor in reduced H1N1 mortality since 1918," and argue it "may not be sufficient" to explain the decline.¹¹⁹  The mutation counts are real.  The inference from them is the weakest argument in the creationist case, for four reasons that do not require any theoretical commitment to see.

The paper measures no fitness.  Fitness in virology means replicative or transmission success, measured by competition assay.  None was run.  The two substitutes are case-fatality rate, which is not fitness — standard virulence theory expects selection to reduce lethality, since killing hosts truncates transmission — and codon bias, which assumes influenza was once adapted to human codon usage and has drifted away from it.  It never was.

The turnover dates do not fit a clock.  H2N2 replaced H1N1 in 1957 after 39 years.  H3N2 replaced H2N2 in 1968 after 11.  A lineage dying of accumulated load at 11 years of age, while another survived to 39, is not a clock.  Immune-driven replacement explains both, and the 2009 replacement as well.

The strain that should be furthest down the ratchet has not collapsed.  H3N2 has circulated in humans continuously since 1968, 58 years, longer than the H1N1 lineage did before 1957.  The mainstream modellers who study its load put the question carefully.  Katia Koelle and David Rasmussen grant that "one might expect the virus population's deleterious mutation load to increase in the long run, causing a long-term decline in influenza A/H3N2 fitness," name two brakes (a rising share of beneficial mutations as fitness falls, and reassortment that recombines low-load segments), and note a measured fall since 1968 in how well H3N2 binds human receptors.  Their verdict on that fall:  "Whether this finding can be interpreted as evidence for a 'weakening' of the virus over time is unclear, especially because 'weakening' by any epidemiological measure has not been empirically demonstrated."¹²⁰  A decline in H3N2 is therefore an open question in the mainstream literature, not a closed one, but nobody has shown one.

And there is a clean control.  Influenza B/Yamagata has not been confirmed in circulation since March 2020, after COVID-19 measures changed how influenza spread worldwide.¹²¹  A lineage that had circulated for decades disappeared within a season, and the timing points to transmission, not to accumulated damage.  Influenza lineages go extinct for ecological reasons, so extinction alone is not diagnostic of mutational decay.

What does survive is smaller and worth stating.  Slightly deleterious mutations genuinely do accumulate in influenza:  Koelle and Rasmussen cite three phylogenetic studies showing that human H3N2 "carries a deleterious mutation load."¹²⁰  And their own simulation of H3N2 shows how the stability of that load is set.  To stop Muller's ratchet in a simulated population, they had to make a share of non-antigenic mutations beneficial, and "we set εc to a value of 0.16, which resulted in a stable deleterious mutation load over time."¹²⁰  They put the need down to the small number of hosts a computer can simulate, and real influenza populations are vastly larger, so this is not evidence that real influenza is ratcheting down.  It does show that in this model the balance comes from a chosen beneficial fraction, not a measured one, which is the gap Argument 3 is about.  All of this illustrates the mechanism.  It is not proof of the conclusion, and this article does not use it as one.

Mendel's Accountant is not the neutral instrument it is presented as

Most creationist quantitative work on this subject runs through Mendel's Accountant, a simulator written by Sanford, John Baumgardner and colleagues.  Its central abstraction — dividing genomes into linkage blocks carrying summed mutation effects — is legitimate, and the Masel lab independently adopted the same design in their own code.¹²²  But its selection algorithm has genuine oddities that neither side has fully resolved.

The most interesting episode is worth recording because of who found it.  In 2015, arguing with a critic on Uncommon Descent, the author of this site's research notes ran a test on Mendel's "probability selection" scheme, comparing an individual of fitness 1.0 to one of fitness 1.1 a million times.  The fitter individual won 20% more often, not 10%.  Joe Felsenstein, a population geneticist and no friend of creationism, agreed with the finding: "If the fitnesses stand in the ratio of 1.1 : 1.0 then Genghis Khan should be 10% more likely to survive.  But owing to (perhaps) the truncation you find him 20% more likely to survive.  The truncation part is not in standard population-genetic models."³⁶

The direction of that bug matters, and it is itself disputed.  On the reading above it makes Mendel's selection stronger than standard theory, which means Mendel understates mutation accumulation rather than exaggerating it.  A 2009 analysis reached the opposite conclusion about the same line of Fortran — work_fitness(i)/(randomnum(1) + 1.d-15) — arguing that dividing by a uniform random number "largely erases any signal from adaptive fitness," making selection weaker and the meltdown "an artifact of bad programming."¹²³  Both claims are about one line, and no mainstream population geneticist has adjudicated between them; Felsenstein declined, saying he could never obtain the documentation and that the program "has too many arbitrary processes."  A year later the same author publicly withdrew the "twice as efficient" claim as possibly an error of his own: "I'm also no longer sure about the comment I made a year ago that selection in Mendel is 'TWICE as efficient as it should be.'  It's possible I could be off by two somewhere."¹²⁴

None of this is why the article does not rely on Mendel's Accountant.  It does not rely on it because it does not need to.  Every number here comes from equations published between 1937 and 1979, and the simulator that checks them is open, in your browser, and grades itself against those equations in public.

Objections and answers

The full inventory runs to more than forty objections and lives in a companion article.  These are the twelve that matter most, each stated in its strongest form.

1. "The load calculation compares everyone to a mutation-free human who does not exist." Correct.  Conceded in full above.  The load argument does not prove decline.  Everything in this article rests on fixation counts instead, which are unaffected by this objection because a fixed mutation is one that literally everyone carries — there is no relative comparison left to make.

2. "Selection is soft, not hard, so nobody needs 800 children." Also correct, and it dissolves the load paradox.  It does not touch the ratchet.  Soft selection decides who reproduces; it cannot remove a mutation that every candidate carries.

3. "Back mutation gives an equilibrium.  There is no runaway." Correct, and calculated above.  The equilibrium for weakly-selected sites is one where roughly half of them are wrong, and it takes on the order of 80 million generations to approach.  The objection is true and irrelevant on any real timescale.

4. "Within the effectively-neutral zone, mutations are as likely to be beneficial as harmful, so drift has no direction." This is the strongest live objection, made by Stephen Schaffner of the Broad Institute among others, and it is Fisher's own geometric model rather than an ad-hoc rescue.  Fisher's result guarantees an excess of harmful mutations at every effect size — but the excess shrinks toward an even split as the effect size falls, and it is exactly the small effects that matter here.

Measured distributions of fitness effects are skewed toward harm wherever they can be resolved.  The honest limit is that they cannot be resolved anywhere near the range in question.  The best precision ever achieved on a single selection coefficient is about 2 × 10^-4, for three mutations in E. coli.¹²⁵  A typical distribution study resolves effects above 4 × 10^-3.¹²⁶  The human drift barrier is 4 × 10^-5.  Nobody has measured the sign of a mutation in the zone the whole argument is about.

There is a second problem when an assay measures a molecular property, such as protein folding or fluorescence, rather than reproductive success.  If the same nucleotide also contributes to a splice signal or a regulatory binding site, a change that improves the measured property may harm a function the assay cannot see.  Montañez, Marks, Fernandez and Sanford model this concern: under their assumptions, the probability of a net beneficial change falls sharply as the number of overlapping codes increases.¹²⁷  Their example assumes that 99.9% of positions are already optimal for each code, that the codes' preferred letters are chosen independently, and that equal beneficial and harmful effects cancel.  Neither the number of codes affecting a typical nucleotide nor these fitness assumptions is measured by the model.

This is a reason to distinguish an improved molecular readout from improved overall fitness.  It does not establish that every reported beneficial fraction is an upper bound on the true fraction.  A competition or reproduction assay already combines the mutation's effects on success in the tested environment, including costs from overlapping functions that affect that success.  It can still miss small effects and tradeoffs in other environments.  Those limits must be assessed for each experiment, rather than imposed by the overlap calculation on all measurements.

5. "Compensatory and beneficial mutations offset the losses." This is the mainstream answer and it may well be right for fitness.  Its weakness is that the required beneficial rate is assumed rather than measured, by the authors' own admission, and that compensation does not restore the specific sequences lost.

6. "Synergistic epistasis makes harmful mutations easier to remove together." Proposed by Kondrashov, Crow, and Eyre-Walker and Keightley.  Sohail and colleagues reported evidence for it in humans and flies in 2017, though the effect they measured is small.¹²⁸  Direct tests mostly fail to find it.  Saber, Baer and colleagues ran C. elegans mutation-accumulation lines for two rounds of about 150 generations each and measured competitive fitness, the setting in which they expected synergy to be easiest to detect.  Their result:  "This study provides no support for the possibility that deleterious mutations interact synergistically," and the second round of mutations cost about half as much as the first, which points the other way.¹²⁹  They name what rides on it, "the long-term existence of species with modest effective population sizes (e.g., humans, Kondrashov 1995 ...)," and say the idea has been "as influential among evolutionary geneticists as the supporting evidence for it is weak."¹²⁹  A 2025 Drosophila inbreeding study, not yet peer reviewed, found synergy in one stage of the life cycle and the opposite in the next, adding up to a straight-line decline in offspring.¹³⁰

Sanford's 2014 edition makes the same empirical point, citing Elena and Lenski's 1997 test in E. coli:  "There is no reason to think that on a genomic level, the primary mode of interaction between mutations could involve synergistic epistasis."¹¹⁸  The same section overreaches when it says the mechanism "has all the appearance of representing deliberate obfuscation."  That is a claim about motive, and nothing supports it.  Synergy is an ordinary hypothesis, and it is being tested.

What this buys the argument is limited.  The skeptic's reply is fair:  one nematode experiment over 300 generations is a modest test, Saber's own group suspects fifty years of inconclusive results come from underpowered experiments, and Sohail's human signal is still on the other side.  And a weak case for synergy does not revive the load argument, which fails for a different reason, the relative-fitness point conceded under Objection 1.  What it removes is the escape Kondrashov himself thought most promising for his own paradox.  Even if synergy were common, it reduces the load but it does not remove nearly-neutral mutations from the blind spot, because a mutation with an effect below 1/(2N_e) is invisible whether or not its effect is amplified by others — unless the amplification lifts it above the threshold, which is the empirical question.

7. "Bacteria and viruses have far higher mutation rates and are fine." Per genome, yes; per functional site, no.  E. coli has a genomic deleterious rate on the order of 10^-4 per generation, and the source is the same paper this article cites against itself on the load argument — Galeota-Sprung, Sniegowski and Ewens, who add the concession that matters here: "For many microbes, then, Haldane load theory may be appropriate."⁴⁸  Humans are at 2 to 10.  A microbe also produces enough offspring that a mutation-free one appears in every generation, which a mammal cannot do.  This is exactly the T7 phage result above, and it argues that genetic entropy is a large-genome, low-fecundity problem specifically.

8. "Mutation accumulation lines recover when you put them back under selection." True — Estes and Lynch measured recovery three times faster than degradation.  But recovery in those experiments comes largely from compensatory mutations and from selection against the strongly harmful mutations that dominate a fast-degrading MA line.  It does not demonstrate the removal of the sub-threshold mutations at issue.

9. "Ancient DNA does not show the predicted decline." Partly true, and reported above.  Genomic health improved from the deep past to the recent past.  Any strong version of genetic entropy has to answer this.

10. "Fitness is relative and environment-dependent, so 'decline' is meaningless." This is a definitional move, not an argument.  A protein-coding exon converted to non-coding DNA is a measurable loss whether or not you call it a fitness decline.  The article's currency is fixed substitutions in constrained sequence, which is a count, not a value judgement.

11. "Your simulation ignores linkage, epistasis, changing environments, population structure..." It ignores linkage because including it makes the result worse, not better — the Full kernel in Drift Lab measures the difference and it runs in genetic entropy's favour.  The other three are genuinely absent, and the honest statement is that changing environments in particular could cut either way.

12. "This is a creationist argument dressed up." The mechanism is Ohta's ratchet, named for Tomoko Ohta, and the fixation counts in this article agree to within 12–26% with a 2025 paper by a mainstream lab that concludes the population is fine.  The disagreement is not about whether the ratchet turns.  It is about whether the compensating process is large enough, and that is an empirical question with an unmeasured answer.

What this does and does not show

It does not show that the human genome is collapsing.  Peter Keightley's group — the same Keightley who co-authored the 1999 paper the load literature rests on — ran the first replicated mutation-accumulation experiment in a mammal and extrapolated it: "We therefore predict a reduction in mean fitness from mutation accumulation of about 0.38% per generation in humans."  Their own verdict on what that means: "it is doubtful whether a reduction in mean fitness due to mutation accumulation will be observable in human populations in the near future."¹³¹  Note the number is a hundred times larger than this article's, and note also that their measured fitness proxies "decrease on average by about 0.2% per generation, but with confidence intervals overlapping zero."¹³¹  That is the honest state of the empirical estimate: not zero, not well constrained, and not fast enough to notice.

At U = 5 the fitness cost of accumulated fixations computed here is about 0.0035% per generation — real, permanent, and far too small to notice in a lifetime.  Kondrashov ran the same calculation with more conservative parameters in 1995 and got the same order of magnitude: with 100 mutations per genome, 10% of them very slightly deleterious at a mean selection coefficient of 10⁻⁶, "they cause the decline of fitness by ~10⁻⁵ per generation."³  Two independent routes, thirty years apart, land within a factor of four.

It does not show that evolution is impossible.  It shows that a particular accounting has to balance, and that whether it balances depends on a beneficial mutation rate nobody has measured well.

It does not support the 6,000-year timescale on its own.  Sanford's 2005 edition makes the case this way.  At the current rate, 300 generations (6,000 years) would cost "about .003% of our total information," which the genome's robustness makes "conceivable";  300,000 generations (6 million years) would cost 3%, and "This is inconceivable."¹³²  His rate assumes "more than 600" mutations per person per generation, six to twelve times the 50 to 100 that Michael Lynch estimated in 2010 for all kinds of mutation together,¹⁰⁴ and it counts every new mutation as lost information, a figure Sanford says "is not affected by what fraction of the genome is actually functional."  With this article's numbers, about 1.1 harmful mutations fix per generation among roughly 220 million constrained letters, so 6 million years would fix about 0.15% of them, a rate about twenty times slower than Sanford's.¹³³  Whether a genome can carry that is the unmeasured accounting question, not a settled one.  In fitness terms, the 0.0035% per generation below, compounded over 300,000 generations, is Kondrashov's paradox, and the answer to it is the compensation argument rather than a smaller number.  Either way the decline rate limits the timescale without fixing it.

What it does show is narrower than the popular version.  Slightly harmful mutations fix permanently in large-genome sexual species, and their count is calculable:  this article's figure and a mainstream lab's simulation agree to within 12–26%.  Whether those fixations add up to degradation is disputed by Barton's school, which holds that polygenic traits compensate them.  They also slow down:  the equilibrium and the fitness feedback above both brake the decline, and at today's human population size the forward ratchet has nearly stopped.  What remains is that the mainstream answer to the accounting question rests on a supply of beneficial mutations that its own advocates describe as uncertain.  That is a genuine open problem in evolutionary genetics, and it is stated as such by the people who work on it.

And the open question this article ends on is not the one Kondrashov left open.  His abstract closes off the exit that the current mainstream answer takes: "This paradox cannot be resolved by invoking beneficial mutations or environmental fluctuations."³

The abstract names two resolutions.  The body names five, and anyone quoting only the abstract — as an earlier version of this article did — is giving an incomplete list:³

1. The equilibrium is never reached, because species do not last long enough to arrive at it.
2. N_e is higher than assumed. He rejects this: hybridization and horizontal transfer occur, but "it is highly unlikely that they are so important."
3. The distribution of fitness effects is bimodal, with few sites in the dangerous range 10⁻⁹ < s < 10⁻⁵. "It is unclear why this should be the case."
4. Synergistic epistasis.
5. Soft selection.

Of the five, three have weakened since.  Synergistic epistasis has found little support:  the Masel lab's summary is that "the mean effects of two deleterious mutations are extraordinarily close to multiplicative,"⁴² and a 2026 test in C. elegans under competition found "no support" for synergy (Objection 6).¹²⁹  Sohail's small 2017 signal in humans and flies is the main result on the other side.  The bimodal-DFE option remains unmeasured and is still the best empirical hope for dissolving the whole problem — it is the same measurement gap named throughout this article, seen from the other side.  Soft selection survives, and this article concedes it.

Option 1 is the one worth stopping on, because it is genetic entropy, proposed by the author of the paradox, in a mainstream journal, and not dismissed:

"the suggestion that accumulation of VSDMs in a lineage with N_e ≪ G acts as a time-bomb, finally destroying it, is a modification of this idea.  If so, the existence of vertebrate lineages with N_e ≤ 10⁴–10⁵ should be limited to ~10⁶–10⁷ generations.  Perhaps, this does not contradict the existing data."³

That is not an endorsement, and it should not be read as one.  It is a working population geneticist saying that a finite lifespan for vertebrate lineages is consistent with what was known in 1995 — and 10⁶ to 10⁷ generations is tens of millions of years for a mammal, not thousands.  It supports the mechanism and not the timescale.  Nothing in this article gets closer to a mainstream statement of its own conclusion, and nothing in this article is further from a young-earth one.

Meanwhile the resolution now offered — a ratchet balanced by rarer beneficial fixations of larger effect — is the one Kondrashov ruled out by name, and it is offered without the measurement that would settle whether he was wrong to rule it out.

That is the honest end state.  Kondrashov may simply have been mistaken about beneficial mutations; the Masel group thinks so and has a model that says so.  But their own paper says "there is great uncertainty in the true values of U_b and s̄_b," and until somebody measures the mean effect of the beneficial mutations available to a well-adapted large genome, the answer rests on an assumed parameter rather than a measured one.  That is a real open problem in evolutionary genetics, and it is stated as such by the people who work on it.

One last limit on the scope of all this, because the version a reader is likely to have met elsewhere is narrower and worse.  Genetic entropy is sometimes defended as a claim about "large multicellular eukaryotes" only, with bacteria and viruses excluded when they fail to cooperate.  That retreat makes the thesis unfalsifiable in exactly the systems where it can be tested in a human lifetime.  This article does not take it.  Sanford's own 2014 statement of the scope is better than that retreat.  He argues that most microbes resist genetic entropy because "in microbial systems there is selection after every cell division.  So there are about 1,000 cycles of selection for every mutation that occurs.  In man it is reversed – there is only 1 cycle of selection for every 100 mutations."¹¹⁸  That is the same per-generation arithmetic as Objection 7, and it is sound.  The weak step comes next.  For RNA viruses he says mutational meltdown "is well documented," which overstates the lethal-mutagenesis record described above, and he explains their survival by strains that sit dormant and "re-emerge from 'natural reservoirs'," an explanation he offers without data on any lineage.¹¹⁸  The argument here is a function of two measurable quantities — the genome-wide deleterious mutation rate and the effective population size — and it predicts what it predicts in any organism, including the ones where the prediction comes out weak.  In E. coli it comes out weak, and this article says so.

Going deeper

• Objections and answers — thirty-four objections, each stated in its strongest form, with a response and a verdict of Fatal, Serious, Partial, Weak or Confused. Six are Fatal.
• The experimental record — mutation accumulation lines, lethal mutagenesis, mutator strains and the human data, including every result that cuts the other way.
• Simulating genetic entropy — Mendel's Accountant honestly assessed, the mainstream alternatives, and why this article does not use any of them.
• What has actually been measured about mutations — the distribution of fitness effects, catalogued. Which numbers are measurements and which are model outputs, what the instruments can and cannot resolve, and why "beneficial" in this literature usually means something was destroyed.
• The literature — every source this argument touches, tagged by topic, camp, and which way it cuts. Filter for undercuts to read the strongest material against this article.
• Sanford's Genetic Entropy, audited — the book that made this argument famous, checked page by page against the primary sources. Seven documented errors, and what four editions across nine years did and did not fix.

Test it yourself

• Drift Lab — the simulator used for every figure here. Change the parameters, run it, read the code, and run the self-tests against Kimura, Watterson and Haldane in front of you.
• Functional DNA is Evidence of Design — the companion argument for a high functional fraction, which is what sets U.

Sources