When AI multiplies hypotheses, scientific taste becomes a practical problem.

When AI multiplies hypotheses, scientific taste becomes a practical problem.

Not taste as prestige. Taste as choosing what to test next.

On 23 September, Anthropic reported a genomic search involving roughly 950 Claude agents over 21 hours. It identified a previously uncharacterized enzyme system, ART. Human laboratory work followed, but the report leaves the system's biological function unresolved.

One detail deserves attention: the team uses its judgments about which hypotheses merit testing to refine Claude's instructions.

This interests me beyond the headline.

I have written before about the cost of verification. Here the question begins one step earlier:

Which uncertainty deserves the next hour at the bench?

Imagine a laboratory with a thousand candidate explanations and capacity for ten experiments.

The ten most persuasive reports are not necessarily the ten best choices.

Nor are the ten most familiar ideas. Nor the ten most likely to produce a positive result.

A useful experiment may distinguish between competing explanations. It may reveal a measurement error. It may eliminate an attractive idea before the team spends six months pursuing it.

That is the kind of scientific taste I care about: recognizing a promising anomaly, then finding a question that reality can actually answer.

Not simply “Could this be true?” But “What observation would help us tell?”

This is not an argument that AI can propose while only humans can judge. AI can help rank candidates, challenge assumptions and design tests too.

The distinction is not human versus machine.

It is the difference between a plausible explanation and a test capable of changing what we believe.

And human taste is not automatically reliable. It can become a preference for familiar methods, fashionable topics or prestigious names.

Selection must remain challengeable. I would reserve some experimental capacity for unconventional proposals, and judge selection methods by what subsequent tests reveal.

In an engineering workshop, three explanations for an overheating motor do not require three complete rebuilds. A well-chosen measurement may separate them before anyone orders parts.

The same discipline matters in research: spend effort where it can change the next decision.

Experimental design and research prioritization are not new. AI makes their importance harder to ignore when proposals arrive faster than they can be investigated.

The aim should not be the largest collection of hypotheses.

It should be a better sequence of questions, tests and corrections.

A thousand hypotheses are not a thousand discoveries.

But one well-chosen experiment can improve the questions we ask next.