Adding more medical images can make an AI answer less reliable, not more.
That is the warning from Trick-or-Treat, a new test of how AI handles conflicting medical images. It pairs a real chest X-ray with a synthetic clinical photograph that does not match. The test checks whether a model notices the mismatch. Instead, models often ignored the conflict and produced findings that fit the X-ray.
That matters because a second source should help a medical professional check an answer. It should not quietly make a wrong answer sound more convincing.
In mismatched pairs, clinically significant hallucinations appeared in roughly one in two cases. With the same photographs shown alone, the rate was about one in three. It was like hearing two witnesses disagree, then quietly changing one account to make them fit.
It is a useful reminder that more input does not automatically mean better reasoning. Trick-or-Treat is a research benchmark, not a test of deployed clinical systems. Its clinical photographs were synthetic. Medical AI evaluations need to check whether systems catch contradictions between images. They should not only test whether a system can describe each one.
The next question is whether future models can flag conflicting evidence instead of smoothing it over.
