Model
152d agoModels Claim to See Images They Never Saw
Multimodal AI models such as GPT‑5, Gemini 3 Pro, and Claude Opus 4.5 can confidently generate detailed image descriptions and medical diagnoses even when no image is supplied.
The Signal
A Stanford study shows that widely used benchmarks mask this hallucination, raising concerns for global AI safety, product reliability, and regulatory oversight across industries and developers.