AI discovery makes choosing the next experiment more important

Anthropic’s latest biology result makes a useful case for judging research agents by the experiments they help a team choose. A promising lead can be valuable long before it becomes a practical tool, provided the next question is clear enough to investigate.

In a September 23 announcement, Anthropic says Claude identified a previously uncharacterised enzyme system called array-associated reverse transcriptases, or ART. The underlying enzyme was already known; the reported discovery concerns its associated DNA repeats and accessory protein. The system’s function remains unknown. It is not an established gene-editing technology.

The company reports a search lasting 21 hours, involving roughly 950 agents. Human scientists performed the laboratory work. That search time does not measure the full path to experimental validation or establish a speed advantage over a comparable human project.

A proposal uses capacity before it earns a result

Consider a research team with more plausible proposals than it can test. Each additional proposal has to justify the attention required to understand it. If accepted, it also competes for experimental time. Cheap generation can therefore raise the value of a good rejection: ruling out an attractive but weak idea early leaves room for a stronger one.

This creates an awkward product incentive. A system can make itself look productive by returning more possibilities, while leaving the scientist with a harder selection problem. The buyer has reason to care about how quickly a proposal can be evaluated, including why it might be wrong. A persuasive report that hides the weakest assumption is expensive even when it takes seconds to produce.

We would give a research assistant credit for making disagreement easier. Its proposal should expose the observation that needs explaining, the competing interpretation and the result that would change its recommendation. Those are useful handles for a reviewer who has limited time and may know something the system missed.

Choosing what to test can also narrow what gets found

There is a second trade-off. Training a system to favour the proposals reviewers already like could make selection more efficient while reducing the range of ideas that reach them. An unfamiliar proposal can be weak, or it can fall outside the reviewer’s expectations. A workflow that treats every rejection as the same kind of lesson risks losing that distinction.

One practical response would be to keep separate records for proposals rejected because their evidence failed and those deferred because the team lacked time or the right expertise. A small, deliberately varied sample of deferred ideas could then receive another review. This is a proposed evaluation method, not a process we have tested or one Anthropic says it uses.

For a lab buying such software, a useful trial would follow a fixed research question and budget through selection and testing. It would count expert review time and record what was learned from unsuccessful candidates, alongside any promising result. Repeating that comparison across different questions would make the purchasing decision less dependent on one showcase discovery.