AI models followed the referral note instead of the brain scan
Original title: Clinical referral text overrides MRI evidence in multimodal large language models: glioma versus brain abscess
How far along is this research?
This is a preprint. Other scientists have not checked it yet, so treat it as an early signal rather than an answer.
This was a test of computer programs on stored scans, not a treatment.
The short version
When the note hinted at a tumor, AI models said tumor, even when the scan did not.
What was studied. Researchers used 60 brain MRI scans. 30 showed a glioma: A tumor that starts in the glial cells. Those are the support cells of the brain and spinal cord. Gliomas are given a grade from 1 to 4. The grade says how the tumor is expected to act. See the glossary and 30 showed a brain abscess. Each scan was paired with five different referral notes and sent to four AI models, which gave 72,000 answers. Two radiologists also read the same scans with no notes.
What they found. With a neutral note, the models said tumor in 45-67% of answers. The two radiologists said tumor in 52% and 60% of cases. Changing the wording of the note changed the models' answers far more than the scans did. In three models, small changes in wording caused at least 83% of that shift. Telling the models to ignore the note did not reliably fix it.
What this means, and what it doesn't
What it could mean: This is about computer tools, not about your treatment. It is a caution for doctors who use AI to help read scans. The words in a referral note can push an AI toward the wrong answer. A trained doctor looking at the scan is still the safer judge.
What it doesn't mean: This does not mean AI is reading your scans today. It does not mean your own scan was read wrong. It also does not mean AI can never help doctors. This was a test on stored scans, not care given to patients. No treatment was tested here, so it is not a promise of a cure. This work is far from everyday care.
Source: medRxiv (preprint), October 19, 2026 · Read the original
This plain-language summary was written by AI and published automatically after passing our automatic safety checks. How we write.