Researcher who shared drafts with Claude questions Anthropic's enzyme-discovery claim

Anthropic's claim covers a larger system around the enzymes it named array-associated reverse transcriptases, or ARTs ^[1]. That system is the enzyme, a run of repeated DNA sequences and a possible partner protein, and the repeats produce small RNA molecules ^[9]. Some of the evidence came from reanalysing data that another research team published in 2022 ^[10]. Reverse transcriptases copy RNA into DNA. Labs already use them in research and in some COVID tests ^[19].

Mestre's team calls the enzymes "jumbotrons" ^[3]. His patent puts a date on earlier work. The explainer's author is a University of Manchester researcher writing on phys.org. In that author's assessment, the patent documents prior study of jumbo-phage reverse transcriptases but does not on its own show that his team had characterised the full system Anthropic describes ^[7].

Anthropic presented the find as AI initiating biological discovery ^[2]. The best evidence on that point is in its own report. In 10 additional search runs, the agents missed the repeat array ^[13]. Some models recognised it when they were handed the relevant sequences directly ^[14]. Spotting a pattern you are given is a different task from finding it in a search ^[14]. The Manchester author writes that this limits claims of a consistently autonomous process but does not show the biology is wrong ^[15]. The report has not been peer reviewed ^[16].

The biology is also unfinished. Human scientists ran the laboratory experiments. They have not yet shown that the enzyme is active, that it copies the associated RNAs into DNA, or what the system does inside the virus ^[11]. The system looks like CRISPR, and that has drawn attention. According to the explainer, whether ARTs can edit genes is still an open question for experiment ^[12].

On the drafts, there is no established evidence that Mestre's unpublished material was used, and the available reporting cannot independently resolve the question ^[21]. I think this part should worry labs more than the credit dispute does. The answer depends on what happened inside a commercial system, and nothing published so far lets an outside party settle it ^[21].

The Manchester author uses AI to turn biological questions into code, then builds a documented analysis that other scientists can check and repeat ^[20]. Researchers should record the model and instructions used whenever AI is essential to a method, the author writes ^[17]. Before sharing unpublished research with an AI tool, they should consult their university's guidance and seek training on responsible use ^[18].