What AI Protein and Materials Discovery Can and Cannot Claim

The fight is not about whether the software did something. It is about what the word discovery is allowed to cover. Anthropic said its system of 950 agents flagged a repeating pattern around a known enzyme after 21 hours of work, according to MIT Technology Review, which first reported the dispute on 28 September.

Biologists pushed back within days.

Lucas Harrington, the biologist who wrote the widely shared critique, argued that finding a strange cluster of genes and repeats is often the easy part, and that the hard part is working out what the system actually does.

That critique was endorsed by the chair and CEO of Eli Lilly, MIT Technology Review reported. A second biologist, Mario Rodríguez Mestre at the University of Copenhagen, said over the weekend that his team had already found the same pattern, and that he regularly chatted with Claude in his work. He wondered whether Anthropic's team had learned from those conversations. Anthropic denies it. Mestre says he is stopping all use of Claude anyway.

Why the bar matters for protein work

The same ambiguity runs through the rest of the AI-for-biology field, and it is worth separating three things that are routinely bundled together: predicting a structure, designing a molecule, and proving that a molecule does something useful in a living system.

Structure prediction is the most mature.

Recent work in the dossier shows how far the design side has moved: an AI-directed protein-engineering cloud lab received $20 million from the National Science Foundation, Hansa Biopharma said on 1 October that it would work with Cradle on AI-driven protein design, and researchers reported using AI structure prediction to speed the search for molecular glues, small molecules that force two proteins together. Each of those is a real capability. None of them is a clinical result.

Verification is where the field is now putting its effort. On 1 October, DeepMind watermarked proteins so that labs could check them, and a related report on 30 September described SynthID Bio as a way to secure AI-generated protein sequences. The logic is straightforward: if a model can output thousands of plausible designs, the bottleneck shifts from generation to checking whether any of them behave as claimed.

That shift is visible outside biology too.

The same pattern shows up in materials, where AI models propose candidate compounds faster than any lab can synthesise and test them. The dossier does not give a single headline materials result this week, so the honest framing is that the biology dispute is the leading indicator for a problem the whole field faces.

"Finding a weird cluster of genes and repeats is often the easy part. The hard part, and where the real discoveries come from, is figuring out what the system actually does."

That quote, from the viral post by Harrington, is the crux. It was reported by MIT Technology Review, not by Anthropic, and it has not been answered on the merits by the company in the material available here.

A separate discovery, and a cleaner one

Not every recent claim is contested in the same way. Astronomers reported on 15 September, in a paper posted to arXiv, the first direct detection of auroral radio emission from an exoplanet, the giant planet beta Pictoris b, using the MeerKAT array. The team, Kevin N. Ortiz Ceballos, Edo Berger and Yvette Cendes, detected rapid, recurring and highly circularly polarized bursts, plus persistent emission, at frequencies of 0.85 to 3.5 GHz.

They identify the signal as electron cyclotron maser radiation, which implies a magnetic field of at least 1.25 kG at the planet. That is the first direct field-strength measurement for an exoplanet. Note what makes this claim durable: a specific instrument, a specific signal, a specific physical mechanism, and a number other groups can try to reproduce with their own telescopes. No AI agent is asked to take credit.

The contrast is not an argument that AI cannot contribute to discovery. It is an argument about how the claim is framed. Whittling 200,000 candidates down to a few worth testing is legitimate scientific work, as MIT Technology Review noted, and a general-purpose chatbot doing that work is notable even if humans steered it and ran the experiments.

The mathematics precedent

The pattern is not unique to biology. Earlier in September, OpenAI said its own agents had cracked a million-dollar problem in mathematics. A couple of weeks later, as MIT Technology Review reported, AI sceptics were sharing an article asking whether it was the math problem that really mattered. The piece did not argue that OpenAI's solution was wrong. It argued that the result may not be the one mathematicians care most about. A mathematician also accused the models of using some of his work without credit.

Once a company frames its output as a discovery rather than a tool result, the debate collapses into two answers, breakthrough or bust, and everything in between gets lost. That is the mechanism the biologists are objecting to, more than the specific enzyme pattern.

Harrington's proposed fix was explicit: AI companies should set the bar high now, so that when an AI actually discovers a fundamentally new biological mechanism, everyone appreciates how big a deal it is. MIT Technology Review noted that as Sam Altman and Dario Amodei race to one-up each other, raising that bar might be the last thing on their minds.

What to watch

Two dates matter for the next phase. MIT Technology Review will publish its 2026 list of Climate Tech Companies to Watch on 6 October, which will not settle the biology argument but will show which companies the publication judges to be delivering measurable results rather than announcements.

The second is the replication question around beta Pictoris b, which is exactly the kind of claim that survives or dies on independent observation. If another array confirms the bursts, the field gets a magnetic field measurement it can build on. If not, the paper joins the long list of single-detection results awaiting confirmation.

For AI protein and materials work, the test is narrower. A model output becomes a discovery when someone else can reproduce the behaviour in a wet lab or a furnace, and when the result is the thing the field was actually looking for. Until then, the accurate description is a promising candidate, not a finding.

Anthropic's own framing invites the harsher reading. Its announcement called the pattern reminiscent of what led to CRISPR, which has already transformed science and medicine, according to MIT Technology Review's account of the text. That is a comparison to a Nobel-winning mechanism, made about a repeating pattern around a known enzyme, and it is why the response was so sharp.

The dispute will not be resolved by argument. It will be resolved by whether independent labs can take the flagged pattern and show what it does. That work is slow, unglamorous, and exactly the part the current claims tend to skip.