Does this count as AI making a scientific discovery entirely on its own?
Not strictly, no. Claude completed the search, screening, literature cross-referencing, and candidate report generation — but the discovery itself couldn't be confirmed as a real, existing enzyme system until human scientists stepped in and performed biochemical and structural verification in the lab. A more accurate description is that Claude autonomously handled a search-and-first-pass-screening stage that traditionally required heavy human labor, while humans remained an indispensable part of the verification step.
That division of labor itself may be more worth tracking long-term than the specific discovery — it shows a concrete shape for human-AI collaboration, not AI replacing scientists.
Why are scientists so divided in their reactions to the same discovery?
The difference mainly comes from what each person is evaluating. Feng Zhang and Stanley Qi are praising the methodology itself — an AI Agent autonomously identifying a structurally unusual pattern in a massive database is genuinely new and worth attention as a search capability. Kevin Blake's skepticism targets the rarity and significance of the finding's actual content — he's pointing out that CRISPR-like sequences aren't uncommon in nature, and questioning whether this specific case is genuinely groundbreaking.
This kind of split is actually common in science, especially at the stage when a new method produces preliminary results: the novelty of the methodology and the substantive significance of the finding are two separate things to evaluate, and media coverage often conflates the two.
Does the scale here — 950 parallel agents, 210 million tokens — offer anything practically relevant to an ordinary developer?
For most developers, the scale itself isn't directly relevant — ordinary teams neither will nor need to deploy nearly a thousand agents for a search task. What's actually worth taking away is the underlying task-decomposition logic: breaking an open-ended search goal (find interesting reverse transcriptase examples) into parallelizable subtasks, where each agent handles a slice of the database and produces its own candidate report, followed by centralized screening and consolidation.
This "massive parallelism plus centralized screening" pattern is actually the same logic ordinary teams use with the Batch API or multiple subagents for large-scale document classification or data-cleaning tasks — just scaled down by two or three orders of magnitude.
If other labs later claim similar AI-driven discoveries, how do I judge whether it's a genuine advance at the same level?
A few concrete elements from this case work as a useful checklist: whether a preprint or technical report is publicly released for peer scrutiny (not just a press release); whether independent third-party scientists outside the research team gave public comment, and specifically whether skeptical views were presented in full rather than cherry-picking only positive quotes; and whether the team itself honestly flagged what it still doesn't know, instead of packaging a preliminary finding as an already-verified conclusion.
Anthropic handled all three elements fairly transparently here — a preprint was published, Kevin Blake's skepticism was quoted in full, and the team explicitly acknowledged not yet understanding the ART system's actual function. That's a reasonable baseline to compare future similar announcements against.
On September 23, 2026, Anthropic announced that its newly established San Francisco biology lab, running an autonomous research task on Claude, identified a previously undocumented enzyme system — named Array-Associated Reverse Transcriptases (ART) — with structural features that reminded researchers of CRISPR gene-editing systems. This is one of the first publicly documented cases of an AI agent being credited with independently completing the full search process behind a concrete biological finding, rather than merely assisting human scientists with literature review.
The instruction given to the research team's Claude deployment was relatively open-ended: search a massive DNA sequence database for "interesting reverse transcriptase examples." Claude wasn't told what specific target to look for — instead it autonomously carried out a sequence of steps: reviewing existing literature, reproducing established results, searching for family members that hadn't yet been characterized, screening candidates, and generating reports. The search deployed roughly 950 agents working in parallel, consumed 210 million tokens, and took 21 hours to complete. Only after that did human scientists take over to review the candidate results and conduct biochemical and structural verification in the lab.
The ART system Claude identified consists of three components: a reverse transcriptase enzyme that copies RNA into DNA, a partner gene, and an array of evenly spaced DNA repeat sequences — it's this repeat-array structure that gives it the CRISPR-like resemblance. The system was found primarily in bacteriophages (viruses that infect bacteria). Initial experiments showed the ART array expresses distinct short RNA sequences, but Anthropic explicitly states in its own technical report: "Although we don't yet know its function" — meaning what this system actually does, and whether it could ever be applied to gene editing, remains genuinely unknown. The company has released a preprint technical report for peer scrutiny.
Feng Zhang, one of the CRISPR pioneers at MIT and the Broad Institute, called it "an exciting example of how AI agents can contribute to biological discovery." Stanford's Stanley Qi praised the AI's "ability to recognize an unusual biological pattern." But skepticism exists too — Washington University's Kevin Blake pointed out that "CRISPR-like sequences" are actually fairly common in nature, stating plainly that "there's nothing to indicate this is a rival to CRISPR-the-technology." This mixed reaction accurately reflects where the discovery actually stands: identifying a structurally unusual pattern is a genuine scientific contribution, but there's still real, unverified distance between that and having an actual, applicable function.
If you work in biotech, drug development, or gene-editing adjacent fields, what's worth paying attention to here isn't the inflated "AI invented the next CRISPR" framing, but a more concrete signal: large-scale parallel agent search (950 agents, 21 hours) can now complete a first-pass screening and initial verification cycle against real biological databases, substantially compressing a literature-review and candidate-screening stage that traditionally required heavy human labor. For investors or industry watchers, the more useful thing to track isn't whether this specific ART system ever becomes the next gene-editing tool — it's whether biology labs like Anthropic's can keep producing peer-reviewable, concrete cases like this one going forward. That's the real indicator of whether "AI-assisted scientific discovery" is a one-off PR moment or an actually scalable research pattern.