Mitchell Katz, CEO of NYC Health and Hospitals — America's largest public health system with 11 hospitals — declared at a Crain's panel that his organization could "replace a great deal of radiologists with AI at this moment, if we are ready to do the regulatory challenge." Katz specifically cited automating breast cancer screening, keeping radiologists on standby only when AI flags abnormal readings for "major savings." This comes weeks after NYC's largest nurses strike in history.

The timing couldn't be worse for AI radiology hype. New Stanford research reveals that frontier AI models can ace medical benchmark tests on chest X-rays without ever seeing the actual images — what researchers call "AI mirages." Unlike typical hallucinations, these mirages produce rational, coherent explanations for findings that don't exist. The models simulate the entire diagnostic process while being anchored to nothing, making standard hallucination safeguards useless.

Radiologist Mohammed Suhail from North Coast Imaging called Katz's comments "undeniable proof that confidently uninformed hospital administrators are a danger to patients" and warned that "any attempt to implement AI-only reads would immediately result in patient harm and death." Suhail's assessment aligns with the Stanford findings — visual language models remain functionally blind despite appearing competent on benchmarks.

For developers building medical AI tools, this is a wake-up call about evaluation methodology. If your models can pass tests without seeing images, your benchmarks are broken. Healthcare administrators shopping for AI solutions need to understand the difference between benchmark performance and real-world reliability — especially when lives hang in the balance.