Diagnostic AI Finally Gets a Rulebook

The industry just admitted something

On August 11, twelve US health systems, among them Cedars-Sinai, Northwell, Mount Sinai, Houston Methodist, and Advocate Health, announced they are forming a Diagnostic AI Consortium with the imaging AI company Aidoc. Together these systems care for close to 20 million patients a year. The stated goal: build shared standards for how diagnostic AI gets deployed, measured, and governed, instead of every hospital figuring it out alone with whatever vendor showed up first.

The framing from Aidoc's CEO, Elad Walach, is the interesting part. He said the industry has "treated speed and safety as opposing forces" in AI, when medicine has never accepted that tradeoff anywhere else. That is a striking thing for an AI vendor to say out loud, because speed is usually the entire pitch.

Why now

The numbers behind the announcement explain the urgency. Imaging interpretation turnaround times more than doubled between 2014 and 2023. Radiologists have been leaving the field at a rate 50% higher than before 2020. Aidoc's own tools are already running across roughly 2,000 hospitals globally, reading 60 million cases a year, more than 150 million cumulatively. That is a lot of AI-assisted diagnosis happening without a shared answer to a basic question: how do you know it is working, and working the same way, at hospital twelve as at hospital one.

That is the gap the consortium says it wants to close. Members will use Aidoc's CARE foundation model and its aiOS deployment layer as common infrastructure, then pool what they learn about safety and quality across very different patient populations and workflows. Hartford HealthCare's CEO, Jeffrey Flaks, called it a commitment to "shaping the future of healthcare through innovation" with responsible development attached. First results are expected in 2027, which tells you this is a multi-year bet, not a press cycle.

The part worth sitting with

Radiology got here first because it was the easiest place to point a model: discrete images, clear ground truth, a specialty already used to reading studies against a reference standard. Ward medicine has none of that structure. A doctor's shift is scattered across a chart, a WhatsApp handover thread, a stack of old discharge summaries, and whatever the last resident remembered to write down. If a well-defined field like radiology needed three years of doubling turnaround times and a mass departure of specialists before the industry agreed governance mattered more than raw speed, the messier parts of medicine are going to need it even more, not less.

This is the same argument for verifiable AI in the workspace where a doctor actually spends the shift, not just at the point where a scan gets read. A tool that drafts a note or suggests a differential is only useful if the doctor can see where the suggestion came from and check it against the chart in front of them, the same way a hospital now wants to check that an imaging model performs consistently before trusting it across twelve different patient populations. Governance is not a constraint bolted onto speed. It is what makes the speed trustworthy enough to actually use.

The consortium will not have real answers until 2027. But the fact that it exists, that a vendor whose entire business is speed is now the one saying speed and safety cannot be traded off, is itself the news. It suggests the easy phase of clinical AI adoption, install the tool, measure the minutes saved, is ending. The harder, more useful phase, proving the tool is right and staying right, is starting.

Sources

• Twelve U.S. Health Systems and Aidoc Unite to Confront America's Diagnostic Capacity Crisis (PR Newswire, August 11, 2026)
• 12 health systems form AI diagnostic consortium (Becker's Hospital Review)
• Aidoc, 12 health systems form diagnostic AI consortium (Modern Healthcare)