TL;DR

Stanford researchers have discovered that frontier AI models — including GPT-5, Gemini 3 Pro and Claude — confidently fabricate detailed descriptions of images that were never uploaded. The “mirage effect” appeared in over 60% of tests, raising serious concerns about how visual AI capabilities are measured and their reliability in medical settings.

The mirage effect

The research team created Phantom-0, a test that asks AI models specific questions about images across 20 categories — but provides no images at all. Rather than admitting they could not see anything, models generated elaborate fabricated descriptions, including exact licence plate numbers, specific newspaper languages and life-threatening medical conditions.

On average, this hallucinated visual behaviour appeared more than 60% of the time across tested models. The finding suggests current multimodal AI systems may be relying on text-based pattern matching rather than genuine visual comprehension.

Implications for medical AI

The study’s most troubling finding concerns healthcare applications. The researchers trained a text-only model — with no access to visual data — to answer questions about chest X-rays. It outperformed both top-tier AI systems and human doctors on a standard chest X-ray benchmark, suggesting the evaluation itself was testing language patterns rather than diagnostic ability.

With over 230 million people worldwide now using AI for health and wellness queries daily, and trust growing among both patients and clinicians, the gap between perceived and actual visual understanding carries real patient safety risks.

A proposed fix

The researchers introduced B-Clean, an evaluation method that filters out questions answerable without images. By removing text-based shortcuts, the approach aims to test what multimodal AI models genuinely see rather than what they can infer from linguistic cues.

They also found that when models were explicitly told an image was missing and asked to guess, accuracy dropped significantly — but when prompted as if an image were present, performance improved through “mirage mode,” where the model relied on hidden text patterns.

Looking forward

For UK businesses and NHS trusts exploring AI-assisted diagnostics, these findings are a cautionary signal. The research, published as a preprint on arXiv, calls for more rigorous benchmarks that separate genuine visual understanding from sophisticated text-based guesswork — particularly in high-stakes medical and safety-critical applications.