A Hypothesis Is Still a Hypothesis
As AI makes candidate discoveries cheap, the harder work moves to testing them and keeping track of what the tests teach you.
By Yaoyu Yang ·
On September 23, 2026, Anthropic published the first result from its new molecular biology group: an enzyme system that Claude agents flagged while searching more than 200,000 reverse transcriptases, the enzymes that copy RNA into DNA. Roughly 950 agents spent 21 hours on the search and flagged about 3,500 candidate systems. One of them sat next to something unusual, a partner gene and a long array of evenly spaced DNA repeats. Anthropic calls the system ART, for array-associated reverse transcriptases, and the layout looks a lot like CRISPR.
The ART result is worth reading closely, because it shows fairly clearly where AI already helps in biology and where the work still sits with the lab.
WHAT HAS THE LAB SHOWN ABOUT THE ART DISCOVERY SO FAR?
So far, the lab has confirmed that the ART repeat array is transcribed, and the system's function is still unknown. Anthropic's scientists found that the array is read out as a set of distinct short RNAs. In published infection data from a Staphylococcus phage, those RNAs made up as much as 8% of the phage's RNA 15 minutes after infection, which suggests the array is doing something during infection.
According to the preprint, the team hasn't yet shown that the ART enzyme is active, or that it works on those RNAs, and Anthropic says the work to find the system's function is ongoing. A week after a preprint, that's a reasonable place to be. It's also where most hypotheses sit for a long time: supported by a pattern, plausible, and waiting on experiments.
Until a hypothesis holds up in reproducible wet-lab experiments, biology usually hasn't given you an answer yet. The ART search took 21 hours. Finding out what ART does will probably take a lot longer, because it means expressing proteins, designing assays, running controls and repeating the runs that don't work the first time.
WHY DOES CHEAP AI SEARCH MOVE THE BOTTLENECK TO THE LAB?
AI search can multiply the number of plausible candidates a team has, while the number of experiments a team can run grows much more slowly. Each experiment needs reagents, instrument time and a scientist's attention, and some fail for reasons that have nothing to do with the hypothesis.
As searches like the one behind ART become routine, we think the pressure in AI-assisted discovery moves to three places. The first is deciding what gets tested, which means ranking candidates against what the team already knows, including experiments that failed. The second is running experiments consistently enough that a negative result tells you something. The third is feeding results back, so the next search can start from what the lab has learned, including its own unpublished data.
Many R&D teams already find that third step hard, with or without AI in the loop. Results end up split across a spreadsheet, a slide deck and someone's ELN entry, and the link back to the question that prompted the experiment gets lost along the way. I once watched an upstream vendor swap the order of two Excel sheets, and ten steps later that swap had turned into an error worth $10 million, partly because nothing connected the final reading back to where it came from. When candidates start arriving by the hundred, gaps like that tend to get more expensive.
WHAT DOES THE ART DISPUTE MEAN FOR UNPUBLISHED RESEARCH?
The ART dispute raises a practical question for any lab using AI tools: where its unpublished data goes, and who can check. After the announcement, Mario Rodriguez Mestre, a computational biologist at the University of Copenhagen, said his group has studied the same system, which they call Jumbotron, for about four years without publishing it, and that he had shared analyses from that work with Claude while using it for research. Anthropic has said that Claude was not trained on user conversations and that its in-house molecular biology team can't access them.
We don't have any reason to doubt either account. What's missing, as of late September 2026, is a way for someone outside the two parties to check the boundary between private research going into a model and discoveries coming out of it. For most R&D teams, this comes down to a few questions to ask any vendor: where their unpublished data lives, who can see it, and whether it's used to train anything.
Here's how Cypher AI answers them. Cypher AI doesn't use customer data to train models, its own or any third party's. Teams that want more separation can run a dedicated deployment in their own AWS account, and admins choose which models the assistant can use. Cypher AI is SOC 2 Type II, and the details are at trust.cypherbio.ai.
HOW CAN AN R&D TEAM KEEP UP WITH AI-GENERATED HYPOTHESES?
An R&D team can keep up by keeping each result, failed runs included, connected to the question and the samples behind it, so the next round of hypotheses starts from everything the lab has learned. That's the part of the loop we built Cypher AI for.
Cypher AI is digital infrastructure for life-science R&D: an ELN, a LIMS, data analysis and a services Marketplace in one platform, with an AI assistant that works across them. The assistant can help plan an experiment and write the analysis, and it shows a plan for approval before it adds anything to the LIMS. Each step a sample goes through is recorded, so lineage builds as the work happens, and a step can be amended later with every change tracked. Analyses run as Python you can read, so a result can be traced back to the samples and code that produced it.
To try this on your own work, start on the Free plan at cypherbio.ai and book a free 15-minute onboarding call from the home screen once you're in. The LIMS is part of the Pro plan, at $100 per user per month. For R&D organizations that need their own cloud, SSO or a forward-deployed engineer working alongside the team, the Enterprise plan covers that, and you can book a conversation through the demo link on the site.