950 Claude agents searched for 21 hours; function still unknown── What was published is an unreviewed preprint; an outsider can follow only the search conditions and the record of human involvement

On 24 September 2026 Anthropic announced that roughly 950 agents of its AI model Claude had spent 21 hours and 210 million tokens searching more than 200,000 reverse transcriptases, and had found a new enzyme system carrying CRISPR-like repeats. What has to be recorded so that an outsider can check a finding produced by AI? Not the conclusion. The search space, the conditions under which candidates were discarded, the points where people intervened, and whatever remains unverified.

01The AI finding was published as a result; the conditions needed to retest it are separate

The figures released were specific. Around 950 agents, 21 hours of running time, 210 million tokens consumed. More than 200,000 reverse transcriptases were examined, narrowed to 3,500 candidate systems, and from those a group emerged that Anthropic named ART.

The document itself, however, is a preprint that has not been peer reviewed. Anthropic writes that it has shown neither that the enzyme is active nor that it acts on the RNAs in question. What the word discovery points to here is the finding of an arrangement of sequences, and no further.

The work of checking this result has therefore not begun. Outside laboratories will begin it, and what they need in order to begin is not the conclusion but a record of the route that led there.

02ART names an arrangement of sequences; no gene-editing activity has been shown

Only the result and the count at each stage were published. So what can the result actually support?

The group named ART is built from three parts: the gene for a reverse transcriptase, a neighbouring gene, and DNA repeats spaced at regular intervals. The arrangement was found mostly inside bacteriophages, the viruses that infect bacteria.

Regularly spaced repeats do call CRISPR to mind. But the preprint does not show that this enzyme is active. Nor does it show that the enzyme works on any particular RNA. The claim on the table reaches as far as the arrangement being unusual.

Outside researchers are divided at exactly this point. One microbiologist said there was nothing to indicate a rival to CRISPR-the-technology, and nothing pointing towards a therapeutic or practical application. Others noted that the enzyme itself was already known, and that the novelty lies in recognising the surrounding arrangement as a single system.

03Wherever AI narrows candidates, including medical document review, the signer cannot see the path unless the conditions are recorded

There is a range the preprint does not cover. The same blank space appears well outside the laboratory.

AI narrows the candidates; a person makes the final call on what is accepted. That shape of work is shared by the hunt for a new enzyme and by the review of medical documents. What they share is not the subject matter. In both, unless the narrowing conditions are recorded, the person who signs cannot see the path by which the field was narrowed.

The same blank is anticipated in law. Article 14(4) of the EU AI Act requires people overseeing high-risk AI to remain aware of their own tendency to rely, or over-rely, on its output. Yet if no record states which candidates were dropped and under which conditions, the person overseeing the system has no way to measure that tendency. The material for measuring it is not in their hands.

04Numbers about scale persuade, but on their own they give no handle for refutation

Without a record of the discarded conditions, nothing reveals over-reliance. So why do figures about scale fail to serve as that material?

950 agents, 21 hours, 210 million tokens. These numbers describe how much computation was run. They do not describe which of the 200,000 candidates were set aside, or on what grounds. Until those grounds are known, no other researcher can walk the same search a second time.

Numbers about computation carry a particular kind of persuasion. The larger they are, the more thorough the work appears. Refutation, though, cannot be performed by volume. To show that a given candidate was wrongly excluded, a researcher needs the criteria that did the excluding.

None of this is a charge against the company. Anthropic states its own reasons for releasing early in preprint form. Releasing early and assembling the route for replication are two different pieces of work.

05Anthropic wrote that human involvement was limited to the prompt and the lab work

If the search cannot be walked again, at least the parts that can be should be counted. Two starting points are on record.

Anthropic wrote that its own involvement was limited to the initial prompt and the experimental work. Put another way, the claim is that no person entered the intermediate stages of gathering sequences and narrowing candidates. For anyone attempting replication, that line is a starting point, because it fixes where machine judgement ends and human judgement begins.

The search began with more than 200,000 reverse transcriptases, passed through 3,500 candidate systems, and arrived at ART. The count at each stage is public. What was thrown away between one stage and the next is not.

Even so, the fact that human contact is recorded at two places is workable for a reader. Examine the initial prompt and it becomes clear what the search was instructed to look for. Examine the experimental record and it becomes clear where living material was involved. Knowing which places can be examined is itself progress.

06A reader should separate three things: peer review, function, and human involvement

The boundary between the published starting points and the unpublished conditions is now visible. From there, a reader's checks divide into three.

The first check is peer review. The ART preprint has not been reviewed. The company explains its own reasons for publishing early, and an explanation is not a substitute for verification. At the moment of release, this result had been through no external scrutiny.

The second check is function. The enzyme's activity was not demonstrated, and no target RNA was identified. A claim that an arrangement is novel and a claim that it can edit genes are separate claims. Only the first is on the table, and nobody should proceed as though genes can be edited.

The third check is whether the points of human involvement are written down. Here the ART announcement did the work: the initial prompt and the experimental stage. Because that is recorded, a replicating laboratory can decide where its own work starts. Of the three checks, this is the only one that gives a starting point for retesting.

07Rules that bar AI from authorship keep responsibility fixed on people

A reader separates the three checks. So who carries responsibility for the result of checking?

The ICMJE, the committee of medical journal editors, states the position plainly. Chatbots cannot be responsible for the accuracy, integrity or originality of a work, and so they must not be listed as authors. Alongside that, the committee asks authors to disclose where and how such tools were used.

The rule is not about capability. It is about who can accept responsibility. However much of a search an AI carries out, the names on the paper belong to people. Whether ART's function will ever be verified is unknown to everyone at present. What is known is that people will answer for the result of the verification.

If responsibility stays with people, so does the making of records. Translated to one's own working day, three things are worth keeping.

Of those three, the one most often forgotten is the discarding conditions. Accepted candidates survive in the finished work, so they remain visible without any record at all. Rejected candidates survive nowhere. Only the person who wrote down the reason for rejection can answer, six months later, why a particular option was never examined.

What an outside laboratory needs in order to recheck an AI finding is not a record of the conclusion. It is a record of the search space, the discarded conditions, the points of human involvement, and the parts left unverified.

The ART announcement kept two of those: the search space and the two places people touched the work. The discarding conditions are absent. As for the unverified parts, Anthropic states them itself — neither the enzyme's activity nor its target has been demonstrated. Three of the four are on record — the search space, the points people touched, and the unverified parts; only the discarding conditions are not.

Read under the pressure of large numbers, that difference disappears. 950 agents and 21 hours describe a quantity of computation and nothing more. What decides whether replication can start is what was written down beside them.