Vision-Language Models Confabulate Medical Diagnoses Without Images

A new research paper highlights a significant issue with current vision-language models: they confabulate medical diagnoses when presented with a query lacking an image. Models like Claude Opus-4.7, GPT-5.4, and Gemini-3.1 Pro were found to generate structured diagnoses based on demographic information provided in the prompt, rather than abstaining due to the missing image. This confabulation is not random, as changing patient descriptors systematically shifted the diagnoses, indicating a vulnerability that requires direct auditing of structured output channels for safe clinical deployment. AI

IMPACT Highlights a critical safety flaw in vision-language models, necessitating new auditing methods for clinical deployment.

RANK_REASON Research paper detailing a failure mode in vision-language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Claude
• Gemini-3.1 Pro
• GPT-5.4
• melanoma
• sarcoidosis
• vision-language model

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →