The better AI gets, the more I think human experience matters.

In a recent experiment, I asked several AI agents to inspect a PBMC single-cell UMAP containing suspicious cell annotations.

The interesting part was not that the models lacked biological knowledge.

In fact, once I pointed to a specific problem, they generally knew exactly what to say.

They could explain the relevant markers, the expected relationships between cell populations, and why a particular annotation looked suspicious.

The knowledge was already there.

What they did not always do was identify which problem deserved attention first.

That distinction has made me think more about what expertise actually means.

Expertise is obviously partly about knowledge. But it is also about recognizing when something does not quite make sense.

An experienced scientist develops expectations about a system:

  • which cell populations should be close to one another,
  • which markers should or should not appear together,
  • which results are biologically plausible,
  • which patterns deserve another look,
  • and which apparently clean result may actually hide a problem.

These expectations act as a kind of prior.

When something violates that prior, an expert may not immediately know the answer. But they know that the result deserves investigation.

That is different from simply retrieving the correct biological fact after being asked the right question.

It suggests that one important challenge for scientific AI is not only knowledge acquisition, but also relevance selection.

More knowledge is useful.

Better retrieval is useful.

More tools are useful.

But more available knowledge also creates more possible explanations, more competing signals, and a larger space of things the system could pay attention to.

The difficult question becomes:

What matters in this situation?

This may be one reason human experience remains valuable even as AI becomes increasingly capable.

Humans do not necessarily know more than AI.

But years of working with experiments, messy datasets, failed analyses, strange results, and biological exceptions help scientists develop judgment about when an answer deserves another look.

AI is rapidly reducing the cost of doing an analysis.

It has not reduced the cost of knowing whether the analysis is right.

That gap may become one of the central challenges for scientific AI.