AI Can Now Analyze Tissue Samples — Trusting It Is the Hard Part

Researchers spent time watching how scientists actually use AI at work — not in a polished demo, but with their own real data. They interviewed fourteen spatial biologists (scientists who map where different cells sit inside tissue samples, often to study cancer and disease) and then watched ten of them use an AI assistant called Claude Science. The AI handled chart-making, locating cells of interest, and other chores the scientists called laborious or simply couldn't do themselves. That part went well.

The interesting finding is what happened next. The AI's speed wasn't the whole story — checking its work turned into a job of its own. To trust a result, the scientists had to gather supporting evidence, sometimes leaving the AI to do extra work in other software, then judge the answer using their own hard-won knowledge of the tissue and the chemical markers within it. In other words, the AI produced answers, but expertise decided whether those answers were real.

The scientists also wanted visibility. They asked for information about what the AI was doing while it was running, so they could decide how the analysis should proceed rather than waiting to see the final output. The paper proposes four design directions: better control over how the AI executes tasks, interfaces that look familiar, clear information about where data and results came from, and verification tools that work on different computers and skill levels.

Why should a non-scientist care? Because this is a preview of nearly every job. AI is increasingly good at the doing, while humans remain accountable for the checking. Companies that make verification easy — showing sources, allowing a pause and a redirect, using screens people already understand — will earn trust faster than those that just promise speed. And for patients, it means AI-accelerated medical research still depends on expert eyes.