When medical images do not agree, AI models can invent a connection

A medical explanation can sound coherent even when its supporting observations are wrong. In the Trick-or-Treat preprint, Oishi Banerjee and colleagues examine a specific version of that problem: models describing abnormalities in a clinical photograph that are not there, particularly when those abnormalities would fit an accompanying chest X-ray.

Posted to arXiv on November 4, 2026, the study presents a deliberately challenging evaluation, not a clinical trial. Its central finding concerns photograph descriptions under controlled conditions—not whether the models correctly diagnosed patients.

A controlled mismatch

The benchmark covers 40 conditions, using 40 real chest radiographs and 80 synthetic clinical photographs. Each radiograph appears in two pairs. In the matched pair, the photograph shows a sign associated with the condition suggested by the X-ray. In the mismatched pair, it shows an alternative finding with little or no connection to that condition.

For example, a radiograph compatible with varicella pneumonia can be paired with a photograph showing chickenpox-like lesions, or with one showing ordinary acne. The second pairing does not support the same unifying diagnosis. Importantly, “mismatched” does not mean medically impossible: unrelated findings can coexist in one person.

The researchers selected radiographs from the ReXGradient dataset and had clinicians verify their compatibility with the target conditions. They generated photographs using gemini-3-pro-image, iteratively incorporating human feedback. Each final benchmark question was approved by one to three board-certified physicians.

Synthetic generation allowed the two photographs associated with each X-ray to retain the same apparent person, pose and setting while changing the relevant finding. This reduces some differences that would complicate comparisons between naturally collected photographs. These are constructed image combinations, however, not documented photographs and radiographs from the same real patient.

What counted as an error

The team evaluated GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1. Models described photographs alone and then photographs paired with radiographs. For paired inputs, the prompt presented the images as coming from the same patient and requested separate descriptions followed by a discussion of what they might suggest.

Scoring focused only on the photograph-description portion. A false positive meant reporting a clinically significant abnormality absent from the photograph; a false negative meant missing one that was present. Normal variants and clinically insignificant incidental features were excluded.

An automated judge, Claude Sonnet 5.5, compared extracted descriptions against reference photograph captions. Diagnostic suggestions elsewhere in a response were excluded. Consequently, these error rates should not be read as rates of incorrect diagnoses.

The same photograph, different results

For mismatched photographs, adding the radiograph increased false-positive rates in all three evaluated models:

• GPT-6 Astra: 37.5% to 57.5%, or 15 to 23 photographs out of 40.
• Claude Fable 5.1: 25% to 45%, or 10 to 18 out of 40.
• Gemini 4 Argon: 35% to 45%, or 14 to 18 out of 40.

These are increases of 20, 20 and 10 percentage points, respectively. The authors report statistically significant increases for GPT-6 Astra and Claude Fable 5.1 individually, but not for Gemini 4 Argon.

Qualitative examples show descriptions shifting toward findings that fit the radiograph. In one example, GPT-6 Astra correctly described a port-wine stain from the photograph alone, but interpreted it as flank bruising when the photograph accompanied a radiograph suggestive of injury.

Matched pairs showed the opposite pattern: all three models had lower false-positive and false-negative rates with the radiograph present. Across models, the decrease in missed abnormalities was statistically significant; the decrease in false positives was not. The result is therefore not simply that additional images made performance worse. Whether the images supported the same clinical explanation mattered.

What the benchmark cannot establish

The evaluation is small and intentionally adversarial, including subtle findings and visually similar alternatives. Its percentages do not estimate error rates in routine clinical practice. Synthetic photographs may also contain generation artifacts despite clinician review.

Clinicians reviewed benchmark construction, but an AI judge supplied the reported error labels. The study did not evaluate radiograph-description quality, overall diagnostic accuracy or downstream patient outcomes. Proposed mitigations, including abstention and multi-agent approaches, remain suggestions rather than demonstrated fixes here.

The contribution is a controlled way to probe whether models preserve visual evidence when combining image types. In this benchmark, a more unified explanation could accompany a less accurate description. Reliable evaluation must distinguish those two things.