AI models cannot actually 'see' images like X-ray and MRI reports that users share while talking to them, a new study found. It said chatbots like Gemini 2.5 Pro, GPT 5.2 and Claude Sonnet 4.5 can guess and respond to image-based questions 100% of the time based on "mirages". Mirages are AI models describing visual features in image inputs' absence.