Medical AI committees adopt wrong answers 38% of the time

Gemini medical diagnosis committees adopted a wrong answer 38% of the time when two peers asserted it, per arXiv 2608.03744.

The committees were tested across seven cohorts on six clinical datasets spanning text, imaging, and tabular ICU records, per the paper. Isolated statistical shortcut cues barely moved the same committees, 5 to 16 percent, so the failure tracks peer pressure, not weak data or confusing cases.

A fabricated 'pre-screen' system flag swayed the committees just as effectively as a real second peer, and adding that second peer voice raised contagion by half again. Tripling a cue's visual salience did nothing, per the paper.

The researchers tried three oversight designs. A gate meant to flag suspicious agreement couldn't separate adoption from honest consensus, a 100% false-positive rate. A transcript judge from the same model lineage as the agents worked on text cases but collapsed on imaging. Only a referee that privately re-queried the holdout data transferred across all seven cohorts, at 77 to 88% precision.

The lesson is specific: oversight that reads the agents' own conversation, even when it shares their model lineage, still failed on imaging. The design that held went back to the original data instead of trusting what the committee said about itself.