Stanford Finds Frontier AI 'Sees' Images It Was Never Shown — And Medical Benchmarks Are Built on This Flaw
Stanford researchers have documented a fundamental defect in how multimodal AI models are evaluated: frontier models score 70-80% on visual benchmarks without any images provided, fabricating confident descriptions through text pattern-matching alone. In medical imaging contexts, models generated severe diagnoses for images that didn't exist.
D.O.T.S AI Newsroom
AI News Desk
A Stanford University research team has identified a structural flaw in the evaluation of multimodal AI systems that undermines a large portion of the performance claims that have driven enterprise AI adoption in healthcare, radiology, and other vision-critical domains. The finding — which the researchers term the "mirage effect" — is not a model quirk. It is a benchmark design problem that affects how the entire field measures visual AI capability.
The Mirage Effect
The core finding: frontier multimodal models achieve 70–80% of their standard visual benchmark scores even when no images are provided. The models are not seeing the images. They are pattern-matching on the text of the question — drawing on statistical regularities in training data — and generating confident, authoritative visual descriptions without any visual input at all.
The researchers tested GPT-5 (including 5.1 and 5.2 variants), Gemini 3 Pro and 2.5 Pro, and Claude Opus 4.5 and Sonnet 4.5. All exhibited the mirage effect in over 60% of tested cases. With typical evaluation prompts, rates climbed to 90–100%.
The term "mirage effect" is chosen precisely: unlike hallucination, which involves generating false details within a valid visual context, the mirage effect means the model constructs an entire false premise — responding as though visual input exists when none was provided. It is a category-level failure that hallucination detection frameworks are not designed to catch.
Medical Imaging: The High-Stakes Case
The researchers paid particular attention to medical imaging benchmarks, where the stakes of inflated performance claims are highest. When Gemini 3 Pro was asked to diagnose across five clinical categories — chest X-ray, brain MRI, ECG, pathology, and dermatology — without any images, the model generated detailed, confident diagnoses. Those fabricated diagnoses skewed heavily toward severe conditions: ST-elevation myocardial infarctions, melanomas, carcinomas. The model was inventing emergencies from nothing.
The implications for API-based healthcare applications are direct: any deployment where an image upload could silently fail — a common occurrence in production environments — could trigger a false urgent diagnosis generated from an empty context. Current benchmark performance figures provide no protection against this failure mode because those figures were established under conditions that share the same flaw.
The Benchmark Clean-Up
The Stanford team introduced a framework called B-Clean that removes benchmark questions solvable without visual input. When B-Clean was applied to three standard visual benchmarks, 74–77% of questions were removed as non-visual. Model rankings shifted on two of the three benchmarks. The performance hierarchy that the field has been using to make deployment decisions was partially built on questions that never required vision to answer.
One particularly striking data point: a 3-billion-parameter text-only model outperformed all tested frontier multimodal models on chest X-ray analysis — and exceeded human radiologists by over 10%. This is not a sign that small text models are better at radiology. It is a sign that the chest X-ray benchmark was measuring something other than visual medical reasoning.
What This Means for Enterprise Deployment
The practical implication is direct: any organisation that has made AI procurement decisions based on multimodal benchmark performance — especially in medical, legal, or technical document analysis contexts — should audit whether the benchmarks used to support those decisions are B-Clean validated. The research suggests they probably are not.
The mirage effect does not mean multimodal AI is useless. It means the evaluation infrastructure that was supposed to tell us how capable these systems are has been systematically overestimating visual performance. The gap between what models claim to see and what they actually process has real consequences in high-stakes domains — and the field has been measuring itself against a mirror.
