Anthropic's Biology Claim Sparks Scrutiny: The Real State of Autonomous AI Discovery

Days after Anthropic posted a pre-print claiming its Claude AI had autonomously discovered a novel molecular system buried in bacteriophage DNA, a computational biologist named Mario Rodríguez Mestre stepped forward with a complicating account: he had been studying the same enzyme system for four years, shared his findings with Anthropic while using the Claude API, and had not been credited for it. Anthropic said it was "not aware of any previously published work describing the ART system we recently found" and denied that Claude was trained on user transcripts. The dispute, which surfaced publicly on September 27, did not erase the scientific interest in what Claude found — but it exposed a deeper question that neither the pre-print nor Anthropic's initial announcement had addressed: what does it actually mean for an AI to autonomously discover something?

What Claude Found — and How

The object at the center of the dispute is a molecular system Anthropic's researchers called ART, for Array-associated Reverse Transcriptases. It was found in bacteriophages — viruses that infect bacteria — and it has an unusual three-part structure: a reverse transcriptase gene, a second gene of unknown function that consistently appears alongside it, and a set of short DNA sequences repeated at evenly spaced intervals.

Reverse transcriptases are enzymes that copy RNA into DNA, running the flow of genetic information in the direction opposite to the usual DNA-to-RNA transcription. HIV encodes one. Bacteria use structurally similar enzymes in immune-like defense systems called retrons, which store short DNA snippets and can trigger immune responses when they detect viral attack. The ART reverse transcriptase appears distinct from these known bacterial families based on sequence comparison, making it potentially a novel lineage.

The repeat array is what generated the most attention. CRISPR systems — the gene-editing machinery that has transformed molecular biology — also contain evenly spaced repeats, which bacteria use to store genetic "memories" of past phage infections, allowing recognition and targeted destruction of returning viruses. Anthropic described ART's repeat architecture as "CRISPR-like," a comparison that Feng Zhang of the Broad Institute, one of CRISPR's co-inventors, called "genuinely intriguing," adding that it was "an exciting example of how AI agents can contribute to biological discovery." Stanford professor Stanley Qi offered a more measured assessment, saying "What stands out is its ability to recognize an unusual biological pattern that was difficult to detect before." Both qualified their reactions in the same direction: the structural resemblance is real, but whether ART's repeats serve any analogous function is not yet known. The pre-print characterizes the ART system's architecture; it does not establish what the system does.

That distinction matters enormously for how the discovery should be evaluated. A molecular system whose structure is novel but whose function is unknown is an interesting finding for genome biologists and a starting point for experimental work — not a confirmed mechanism or a deployable tool. The pre-print was posted September 23, 2026 and has not yet undergone peer review.

Read more: Claude AI detects CRISPR-like enzyme system in phage DNA

How 950 Agents Scanned a DNA Database for 21 Hours

The methodological core of Anthropic's announcement was not that Claude generated a biological hypothesis from first principles. It was that approximately 950 Claude agents ran in parallel for 21 hours, consuming around 210 million tokens, scanning a massive DNA sequence database for uncharacterized protein families bearing structural patterns associated with reverse transcriptase domains and unusual repeat architectures.

Anthropic launched Claude Science in beta on June 30, 2026, positioning it as a research workbench that connects Claude to more than 60 scientific databases and lets researchers run agentic workflows that would be impossible to execute manually. The ART discovery was its first major public result — an attempt to demonstrate that multi-agent AI systems can identify novel biological findings at a scale and speed that exceeds what a single research group could achieve.

What the agents did is best understood as automated genome mining at unprecedented scale. Genome mining — systematically searching DNA sequence databases for uncharacterized or underappreciated protein families — is a recognized bioinformatics discipline. Traditional genome mining uses specialized tools like BLAST for sequence similarity searches or HMMer for profile-based pattern matching. Expert bioinformaticians design the queries, interpret the results, and determine what constitutes a novel finding. Claude Science's contribution is layering LLM-scale contextual reasoning on top of this process and parallelizing it to a degree that was previously computationally impractical.

The agents were not generating new scientific theories. They were searching for patterns in existing data using parameters set by Anthropic's researchers — parameters that already embodied substantial biological knowledge about what constitutes an interesting protein family. The ART system was identified because it matched a profile that human scientists had reason to search for. Whether that constitutes "discovery" in the full scientific sense, or a very powerful form of automated literature and database mining, is the crux of the dispute Mestre raised.

The Autonomy Claim in Context

The word "autonomous" in Anthropic's framing deserves scrutiny. The 950 agents ran without a human directing each step of the database scan — that much is straightforwardly true. But the broader framework in which they operated was entirely human-designed: the databases were selected and curated by scientists; the search architecture reflected existing biological knowledge about what makes a reverse transcriptase interesting; and the decision to look for repeat-associated RT systems almost certainly drew on prior knowledge of CRISPR and retron biology. The agents were autonomous the way an automated stock screener is autonomous — operating within constraints, parameters, and objectives that humans defined.

Mestre's account introduces a sharper version of this problem. If he had been studying the same enzyme system for four years and shared findings with Anthropic while using the Claude API, then the question is not only whether Claude's search was autonomous in its execution, but whether the intellectual framework it operated within was built in part on human expertise that was not formally credited. Anthropic denied that Claude was trained on user transcripts. That denial is specific and important: it addresses one mechanism by which Mestre's knowledge could have influenced Claude's outputs. It does not address whether Mestre's research shaped the search strategy in some other way, or whether the pre-print's framing should have acknowledged him regardless.

The comparison that throws this into sharpest relief is AlphaFold. DeepMind's protein-structure prediction system, which earned the 2024 Nobel Prize in Chemistry and has predicted structures for more than 200 million proteins, is perhaps the clearest prior example of AI enabling genuinely transformative scientific scale. AlphaFold's claims are straightforwardly evaluable: it takes a protein sequence and predicts a 3D structure; the structures can be compared against experimental results from X-ray crystallography or cryo-electron microscopy; the error rates are known. The system's autonomy is bounded by its training data, its architecture, and its objective function — and those constraints are explicit and documented.

The ART pre-print presents a different epistemic situation. The finding cannot be independently verified by running the same procedure over the same database and arriving at the same result using different tools, in the way AlphaFold's outputs can be checked against experimental structures. It can only be confirmed by independent experimental characterization of ART's biological function — work that is still ahead.

Where Google and OpenAI Stand

Anthropic is competing in a field where the other major AI labs are also making science-adjacent claims, and the credibility of those claims varies significantly.

Google's Gemini Co-Scientist system, described in a paper published in Nature in May 2026, focuses on a different step in the scientific process: hypothesis generation from scientific literature. Rather than mining raw sequence databases, Co-Scientist synthesizes existing published research to propose novel hypotheses that human scientists can then test. The Nature publication brings with it the credibility of peer review, which the Anthropic ART pre-print has not yet cleared. Google has also published a separate AI research tool called ERA (Empirical Research Assistance) and AlphaEvolve for algorithmic optimization — a multi-platform commitment to AI-assisted science across different task types.

OpenAI's most clearly verified scientific AI achievement is of a different character entirely. In May 2026, an OpenAI reasoning model produced a lengthy formal proof disproving an 80-year-old mathematical conjecture about planar unit distances, attributed to Paul Erdős. The proof was checked by a group of external mathematicians; Fields medalist Timothy Gowers, who reviewed a companion paper on the result, wrote that he "would have recommended acceptance without any hesitation." Mathematics has a property that biology lacks: proofs are either valid or invalid, and validity can be established by expert inspection. There is no attribution ambiguity, no pre-print uncertainty, no waiting for experimental replication. OpenAI's Astra model separately solved ten long-standing open problems in mathematics and theoretical computer science in August 2026, again in a domain where correctness is checkable.

The ART finding is harder to evaluate than either of these. It is more open-ended than a mathematical proof, less experimentally established than AlphaFold's structures, and less peer-reviewed than Google's Co-Scientist paper. That does not make it unimportant — genuinely novel structural characterizations of bacteriophage molecular systems are scientifically valuable, and the scale of the genome mining methodology is legitimately notable. But the "AI becomes a scientist" framing requires scrutiny of which parts of the scientific process AI is actually performing, and in this case, the answer is primarily the search step, not the hypothesis generation, experimental design, or functional interpretation steps.

A Gap in the Rules of AI-Assisted Science

The most significant consequence of the Mestre dispute may not be about Anthropic or the ART system specifically. It is that no established framework currently exists for resolving it.

When Mestre raised his concerns, there was no journal policy to consult, no institutional guideline that applied, no precedent from a prior case. The resolution — Anthropic's denial that it had trained on user transcripts — was informal and reputational. It satisfied neither a clear legal standard nor a scientific norm, because no such standards have been articulated for AI-assisted research. The question of whether user interactions with an API can constitute undisclosed research contributions to a company's published findings has no established answer in science or law.

This gap will not remain unfilled. Science journals, funding agencies, and academic institutions are in the early stages of developing policies for AI-assisted research attribution. The Mestre dispute gives those efforts a concrete case study: a named researcher, a named AI system, a specific type of interaction (API use while working on a related topic), and a specific type of claim (autonomous AI discovery). Whatever resolution principles emerge from this case will shape how AI science claims are credited and evaluated for years.

The ART System as a Starting Point

None of this analysis undercuts the genuine scientific interest in what Claude found. The three-part ART architecture — reverse transcriptase, partner gene, evenly-spaced repeat array — is a real observation in real biological sequences, endorsed by credible experts in the relevant field. If the repeat array turns out to serve a function analogous to CRISPR's immunological memory, even in a bacteriophage context rather than a bacterial one, that would be a meaningful contribution to the understanding of phage-bacteria molecular arms races. Bacteriophages are among the most abundant biological entities on Earth; their molecular machinery is substantially undercharacterized relative to cellular organisms.

What comes next for ART will require traditional experimental biology: expressing the proteins in a controlled system, characterizing their biochemical activities, testing the repeat array's role by systematically deleting or mutating it, and observing what happens. That work is almost certainly underway in multiple labs following the pre-print's release, including at Anthropic's own BSL-1 and BSL-2 biology facility. Peer review of the pre-print will subject the structural characterization to expert scrutiny independent of the press cycle.

The ART story is therefore best understood not as a completed discovery but as a hypothesis generated by a very powerful search system — one that now needs to be tested by the slower, messier, human-driven experimental process that remains science's most reliable verification mechanism. Autonomous AI did the scanning. Discovery, in the full sense, is still pending.

The more durable milestone from this episode may be the questions it forced into the open: what credit is owed when human expertise shapes an AI's search strategy, how pre-print science interacts with AI's speed advantage over peer review, and what benchmarks the field will eventually require before an AI system can be called a scientist rather than an exceptionally powerful research assistant. Those questions will outlast the ART pre-print's path through peer review, and their answers will determine how much the AI-scientist era actually changes science rather than just the rate at which patterns in existing data get identified.