Skip to content

Anthropic, Please Don't Ruin Bio

Source: Anthropic, Please Don't Ruin Bio for Everyone \ Date Published: June 14, 2026 \ Author/Org: Anonymous Biologist (Guest Post on Cremieux Recueil)

TL;DR

A practicing biologist delivers a scathing critique of Anthropic's approach to AI bio-risks, arguing the company engages in fear-mongering while producing poor public research that threatens to restrict legitimate biological inquiry. The critique centers on four pillars: (1) the Fable 5 precedent — Anthropic CEO Dario Amodei called for government regulation of AI, and the US responded with export controls; applying the same logic to biology will block legitimate queries. (2) The burden of proof: all technology is dual-use, and Anthropic must demonstrate a unique harm-benefit imbalance for AI in biology. (3) The real bottleneck in biological research is slow real-world experimentation, not cognitive labor. (4) Anthropic's public benchmarks (BioMysteryBench) contain "comically sloppy" errors — wrong species identification, wrong contamination IDs, ambiguous conventions, and incorrect taxonomy — a textbook case of Gell-Mann Amnesia.

The Fable 5 Precedent

Anthropic's track record on AI regulation is instructive. When CEO Dario Amodei called for government intervention on AI risks, the US government responded with export controls that restricted access to advanced compute and model weights. The biologist argues the same pattern will play out in biology: Anthropic sounding alarms about AI-enabled bio-risks will lead to restrictions that block legitimate researchers — not hypothetical bioterrorists.

The Burden of Proof Problem

The author makes a principled argument about dual-use technology:

All technology is dual-use. A kitchen knife can be used for cooking or murder. The question is not whether a technology could be misused, but whether the harm-benefit balance is uniquely unfavorable.

The burden is on Anthropic to demonstrate that AI in biology presents a harm-benefit imbalance significantly worse than existing tools — gene synthesis, CRISPR, chemical synthesis, or even standard search engines. The author argues this case has not been made.

The Real Bottleneck in Biology

A crucial but overlooked point: the rate-limiting step in biological research is not cognitive labor (which AI could accelerate) but slow real-world sampling.

  • Experiments take weeks or months
  • Cell cultures grow on biological timescales
  • Clinical trials take years
  • Regulatory approval takes years more

AI can generate hypotheses faster, but the testing bottleneck remains unchanged. This fundamentally limits the marginal risk AI adds to biological research — the bottleneck is physical, not cognitive.

Gell-Mann Amnesia and BioMysteryBench

The most damning section of the essay applies Gell-Mann Amnesia to Anthropic's public benchmarks. The author, an expert in the field, can spot obvious errors in Anthropic's BioMysteryBench:

  • Wrong species: A question about human biology was answered using data from mice
  • Wrong contamination IDs: Samples were assigned incorrect contamination statuses
  • Ambiguous conventions: Questions had multiple valid interpretations depending on lab conventions
  • Wrong taxonomy: Organisms were misclassified at the genus or species level

The author's point: when an expert reads Anthropic's work in a field they know intimately, they find "comically sloppy" errors. This should make us skeptical of Anthropic's claims in domains where we lack expertise — the very domains (AI risk, alignment) where they claim authority.

Key Takeaways

  1. Anthropic's pattern of calling for government regulation has led to export controls — applying the same approach to biology will restrict legitimate researchers.
  2. All technology is dual-use; Anthropic must demonstrate a unique harm-benefit imbalance for AI in biology, which it has not done.
  3. The real bottleneck in biology is slow real-world experimentation, not cognitive labor — limiting the marginal risk from AI.
  4. Anthropic's BioMysteryBench contains multiple "comically sloppy" scientific errors: wrong species, wrong contamination IDs, ambiguous conventions, wrong taxonomy.
  5. Gell-Mann Amnesia applies: experts finding obvious errors in Anthropic's bio work should make us skeptical of their claims in domains where we lack expertise.
  6. The essay is a plea for evidence-based risk assessment rather than fear-driven restriction of biological research.