NYC Hospital CEO Ready to Replace Radiologists with AI — Is Medicine's Biggest Disruption Here?
April 4, 2026 · Happycapy Guide
On March 25, 2026, Mitchell Katz stood at a Crain's New York Business panel and said what hospital administrators had been thinking for two years: AI is ready to replace most radiologists, and regulations are the only thing in the way.
Katz runs NYC Health + Hospitals, America's largest public hospital system with over 70 patient care locations serving 1.1 million New Yorkers annually. When he speaks, the radiology industry listens — and then pushes back hard.
What Katz Actually Said
"We could replace a great deal of radiologists with AI at this moment, if we are ready to do the regulatory challenge," Katz told the panel audience. His specific use case: breast cancer screening. The proposal is to let AI handle all initial mammogram reads, routing only flagged cases to human radiologists. The result, he claimed, would be "major savings" for an underfunded public hospital system.
The math is straightforward from a cost perspective. NYC Health + Hospitals operates on thin margins, and radiologists are among the highest-paid physicians in the U.S. — earning $400,000–$550,000 annually. Replacing even a fraction of initial reads with AI at near-zero marginal cost is a significant budget line.
The Radiology Counterargument
Dr. Mohammed Suhail, a radiologist at North Coast Imaging in San Diego, was blunt in his response: hospital administrators are "confidently uninformed" and implementing AI-only reads would "immediately result in patient harm and death."
The pushback is not purely self-interested. Radiologists point to three fundamental limitations of current AI systems:
| Limitation | Details |
|---|---|
| Narrow specialization | AI excels at specific tasks (lung nodule detection, bone fracture flagging) but cannot integrate findings across organ systems the way a radiologist does in a full body scan. |
| No medicolegal accountability | A radiologist signs their name to a report and can be held legally responsible for a missed diagnosis. No AI system in 2026 carries equivalent accountability. |
| Benchmark gaming | Stanford researchers found some AI chest X-ray tools pass medical benchmark tests without actually processing X-ray image data — constructing plausible-sounding conclusions without seeing the images. |
The Stanford "AI Mirage" Study
The most alarming piece of evidence against AI replacement came from a yet-to-be-peer-reviewed Stanford study released in early April 2026. Researchers found that several frontier AI models, when deployed as chest X-ray reading tools, could produce passing scores on medical benchmarks without ever processing the actual X-ray images.
The researchers called it an "AI mirage": the models construct rational-sounding radiological findings by drawing on statistical patterns from their training data rather than performing genuine image analysis. A patient with an actual pneumothorax could receive a normal read from an AI that never "looked" at the scan.
This benchmark-passing-without-understanding problem is not universal — purpose-built radiology AI systems trained specifically on imaging data perform very differently. But it raises a fundamental question about which AI systems hospitals are actually deploying.
Where the 700+ FDA-Cleared AI Tools Actually Stand
Over 700 FDA-cleared AI algorithms are deployed in radiology departments across the U.S. as of April 2026. None of them are operating as full replacements for radiologists. Their actual deployment falls into three categories:
| Deployment Model | What AI Does | Human Role |
|---|---|---|
| Worklist prioritization | Flags critical findings (e.g., intracranial hemorrhage) to rise to top of queue | Radiologist reads all studies; AI changes order |
| Detection assistance | Highlights potential abnormalities in mammograms, chest CTs, bone scans | Radiologist confirms or overrides every AI finding |
| Routine screening filter | Clears clearly normal mammograms from the queue (requires dual-reader jurisdiction) | One radiologist spot-checks AI-cleared studies |
The Katz proposal goes further than any of these: eliminating the radiologist entirely from normal reads, with AI as the sole reviewer unless the AI itself flags an abnormality. This model has no precedent in U.S. clinical practice.
Radiology Workforce Is Already Strained
The Neiman Health Policy Institute published research in JACR in early 2026 showing that the radiology workforce has likely reached maximum capacity. Emergency wait times for radiology reports are rising. Rural hospitals face chronic radiologist shortages. The imaging volume is growing faster than the trained workforce can absorb.
In this context, the Katz argument is less about eliminating radiologists than about covering the gap. Public hospital systems cannot afford to hire enough radiologists at market rates. AI systems that could offload routine screening reads — even imperfectly — would meaningfully improve throughput for underserved populations.
The Regulatory Barrier
The FDA's current framework for AI/ML-based medical devices requires pre-market clearance for each specific AI indication. An AI that reads mammograms cannot automatically be used for chest X-rays — each new application requires a separate clearance. More importantly, the FDA has not cleared any AI system for autonomous diagnostic radiology — meaning a human radiologist must review and sign all diagnostic reports under current U.S. law.
Katz knows this. His comment was an explicit call for regulatory change, not a description of current practice. Whether the FDA moves in this direction in 2026–2027 is the actual question.
What This Means for Healthcare AI Broadly
The Katz controversy is a preview of debates that will play out across medicine over the next five years. Radiology is the most tractable case — visual pattern recognition, discrete outputs, measurable benchmarks. If AI replacement is too risky even here, the argument for AI autonomy in psychiatry, emergency medicine, or surgery becomes harder still.
The most likely outcome in radiology: AI handles a growing share of screening workload in narrow, high-volume applications (mammography, lung nodule surveillance), with radiologists increasingly focused on complex multi-modality cases and interventional procedures. Full replacement — as Katz described — remains a 5–10 year question, not a 2026 reality.
Frequently Asked Questions
Can AI replace radiologists in 2026?
AI can assist and augment radiologists on narrow tasks, but full autonomous replacement is not FDA-approved and is not happening in 2026. The technology is capable in specific use cases; the regulatory and liability frameworks are not ready.
What did Mitchell Katz say about AI and radiologists?
The CEO of NYC Health + Hospitals said on March 25, 2026 that "we could replace a great deal of radiologists with AI at this moment, if we are ready to do the regulatory challenge" — specifically proposing AI-first breast cancer screening with radiologists reviewing only flagged cases.
What is the Stanford AI mirage study?
Stanford researchers found that some AI chest X-ray tools pass medical benchmarks without actually processing the X-ray images, instead constructing plausible findings from training data patterns. This raises fundamental questions about which AI systems can be trusted for clinical use.
How many AI tools are already cleared for radiology?
Over 700 FDA-cleared AI algorithms are deployed in radiology as of 2026, all in assistive roles — worklist prioritization, detection flagging, and screening filters. None are authorized to issue autonomous diagnostic reports without radiologist oversight.