The Experimental Record

"The increase in fitness is in qualitative disagreement with the quantitative theory applied." — Springman, Keller, Molineux and Bull, reporting that a phage got fitter under a hundredfold mutation rate¹

Theory is cheap.  This article is what happened when people stopped selection and watched, and it does not all point one way.  Where the results cut against genetic entropy they are reported first and in full, because a reader who finds them somewhere else will stop trusting everything around them.

Mutation accumulation lines

The design

Breed an organism through a bottleneck so tight that selection can barely operate — one individual, or one brother-sister pair, per generation — for as long as patience allows.  Then measure the survivors against a frozen or cryopreserved ancestor.  Whatever has changed is mutation, because nothing else was allowed to act.

This is the closest thing to a direct test that exists.

What was measured

Four things this table says

Animals lose fitness under relaxed selection, and by orders of magnitude more than microbes.  Flies lose a few tenths of a percent per generation; E. coli loses 0.00027%.  That is a difference of roughly a thousandfold, and it tracks exactly what the theory predicts: microbes have small genomes, so a smaller share of their mutations lands anywhere that matters.

How you measure fitness changes the answer by a factor of ten.  Shabalina's group measured the same kind of Drosophila lines two ways: in comfortable vials, and in competition against rivals.  Benign measurement gave 0.2% per generation.  Competitive measurement gave 2.0%.²  Nature measures competitively.  This gap recurs across the literature and it always runs the same direction.

The estimates disagree by a factor of 700.  Katju and Bergthorsson's review of every eukaryote mutation accumulation study puts the measured genomic deleterious rate between 0.00065 and 0.47 per generation.³  Anybody quoting a single number from this literature — in either direction — has chosen it.

And the results that cut against genetic entropy are in the table too.  Wild-type yeast lost nothing measurable in 600 generations.  Arabidopsis lost nothing measurable in 17.  Most importantly, Katju's 2015 experiment found no significant decline at all in C. elegans at population sizes of 10 or 100, over 409 generations.⁴

Recovery is fast

The sharpest counter-result belongs to Estes and Lynch.  They took C. elegans lines that had been degrading for 240 generations under a single-individual bottleneck, put them back into a large population under normal selection, and measured full restoration of fitness in under 80 generations — a recovery rate "at least three times that of mutational degradation."⁵

If degradation reverses three times faster than it accumulates, genetic entropy in its strong form is in trouble.  That has to be conceded.

Two things limit how far it generalises, and both are quantitative rather than rhetorical.  A single-individual bottleneck accumulates strongly harmful mutations fastest, because those are the ones selection would normally remove and nothing was removing them — so a degraded MA line is loaded with exactly the class selection is best at clearing.  And a 409-generation experiment at N = 100 cannot detect a process whose characteristic timescale is 4N generations and whose effect sizes are of order 1/(2N).  The mutations at issue in genetic entropy are the ones such an experiment is designed not to see.

Error catastrophe in viruses

It works

Raise a virus's mutation rate far enough and it dies.  This is not controversial; it is the basis of real antiviral drugs.

Four reasons not to overclaim it

It is not Eigen's error catastrophe.  Bull, Sanjuán and Wilke showed that killing a virus with mutations is ordinary lethal mutation accumulation, not the exotic information-melting phenomenon the name suggests.¹¹  Creationist material that uses the two interchangeably is wrong on the mechanism.

The required rate increases are enormous.  A 2025 analysis across six viruses found that lethal mutagenesis needs mutation-rate increases of 11-fold to 2,700-fold, while real drugs deliver at most about tenfold.  Their conclusion: "lethal mutagenesis seems difficult to achieve."¹²

Molnupiravir did not kill the virus, and did not obviously drive it either.  Sanderson and colleagues found SARS-CoV-2 branches carrying twenty or more mutations with the drug's characteristic signature, including a 20-tip transmission cluster in Australia.¹³  Raising the mutation rate produced viable, transmissible descendants.

And the cleanest counter-experiment is in this literature, not outside it.  Springman, Keller, Molineux and Bull mutagenised phage T7 at a hundred to a thousand 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.  After 200 generations it had risen to 21.9.¹

That result deserves its own paragraph rather than a footnote.  The most likely explanation is arithmetic: at 2.6 deleterious mutations per generation, an organism producing more than e^2.6 ≈ 13 offspring will, on average, produce one carrying no new mutations at all — and selection can then pick it.  Phage burst sizes are in the hundreds.  Mammals have litters.

Note also what the same experiment found alongside the growth-rate gain: average burst size fell 80%, and 245 mutations per genome accumulated.  The authors say the two results "appear to be contradictory."  The virus got faster at the one thing being measured while getting worse at another.

Mutator strains

Bacteria that lose their mismatch repair machinery mutate 10 to 100 times faster than normal.  In Lenski's long-term evolution experiment, several populations did exactly that.  Their genomes decayed — and their fitness kept rising for tens of thousands of generations.¹⁴

That is a real and awkward result for a simple decay model, and it has the same explanation as the T7 phage: E. coli in a flask has an enormous population, a huge supply of beneficial mutations relative to a mammal, and a genome in which a much smaller fraction of sites are constrained.  Which is to say the experiment is a poor model of the case at issue — and that is a limitation of the argument's scope, not a refutation of it.

The human data

The most misused result in the subject

Fu and colleagues sequenced 15,336 genes in 6,515 people and found that about 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 in European Americans.¹⁵

Creationist material quotes this constantly and usually stops there.  The authors' own explanation is demographic: human populations exploded recently, and a rapidly growing population accumulates rare variants faster than selection can purge them.  Their model fits the acceleration beginning 5,115 years ago at about 1.95% growth per generation.¹⁵

The finding is consistent with genetic entropy.  It is not evidence for it over the standard explanation, which fits the data quantitatively and was proposed by the people who collected it.

What the rest of the human data shows

• About 100 genuine loss-of-function variants per person, with roughly 20 genes completely inactivated.¹⁶
• 7.7% of Icelanders are complete knockouts for at least one gene, across 1,171 genes.¹⁷
• Roughly 400 highly damaging missense variants per person, and about 2 bona fide disease mutations.¹⁸
• Neanderthals carried a load implying median fitness around 0.63 of modern humans; the Altai Neanderthal's genome scores worse than 97% of people alive today.^{19 20}

The ancient-DNA test, which nobody has run properly

This is the sharpest available empirical test of the claim, and its current answer is mixed.

The same study that scores the Altai Neanderthal poorly also finds that genomic health improved from the deep past toward the recent past — the opposite of a straightforward decay model.  And it finds that individuals from the last few thousand years score healthier than people alive today, a pattern the authors themselves call perplexing.²⁰

Neither side gets a clean result.  Genetic entropy predicts a monotonic decline that the deep-time data do not show.  The critics' position predicts no systematic trend, and there is one in the recent data.  A properly designed ancient-DNA study of mutational load over time would settle a great deal, and it has not been done.

Is anything measurably declining now?

Yes, slightly, and the size of the effect is the point.

Kong and colleagues measured polygenic scores for educational attainment across 129,808 Icelanders and found a decline of about 0.010 standard units per decade of birth year — implying roughly 0.30 IQ points per decade in the genetic component.²¹  Beauchamp measured selection against educational attainment in the United States at about 1.5 months of education per generation.²²

These are real, statistically solid measurements of ongoing genetic change in a fitness-relevant trait.  They are also, notably, the kind of change Crow expected — he called mutation accumulation "something like the population bomb, but it has a much longer fuse."²³  And they are running against a Flynn effect of +3.0 IQ points per decade — the environment improving ten times faster than the genetics decline.  Which is exactly what Crow said in 1997: mutation accumulation "is more than compensated for by much more rapid environmental improvements."²³

That is the honest shape of the human evidence.  Something is happening.  It is small, it is slow, and it is currently invisible under an environmental improvement running an order of magnitude faster.

What the experiments add up to

For genetic entropy: fitness declines measurably in animals under relaxed selection; the decline is worse when fitness is measured competitively; raising mutation rates kills organisms; the effect scales with genome size exactly as the theory predicts; and humans carry a large, measurable burden of loss-of-function variation.

Against it: degraded lines recover quickly under selection; modest population sizes stop the decline entirely in the one experiment that tested it; mutator bacteria and mutagenised phage gained fitness rather than losing it; and ancient DNA does not show the monotonic decline a simple version of the claim predicts.

What neither side has: a direct experimental test at the parameters that matter — a large-genome, low-fecundity, sexually reproducing species with an effective population size near 10,000, run for tens of thousands of generations.  That experiment is impossible, which is why this argument is conducted in equations and simulations.  It is also why the simulations have to be checkable, and why Drift Lab exists.

Sources