AI models like GPT-5 and Gemini 3 Pro do not truly understand images despite strong performance Stanford researchers identified a “mirage effect” where AI fabricates image analysis without input Models achieved 70-80% accuracy describing images removed from datasets, showing confident fabrications