AI in healthcare isn't being deployed to eliminate doctors. It's being deployed because we don't have enough of them. At least that’s what a lot of AI companies are purporting.

Ambient scribes reduce documentation time. Predictive models flag deteriorating patients faster. Imaging AI catches patterns our mere human eyes miss because it’s 2 AM on an overnight shift.

Every single one of these use cases exists because the system needs doctors to do more, not because it needs fewer doctors. Let’s dive in.

You've probably already used an AI-assisted tool this week. Maybe for a radiology read, a clinical decision support alert, or an ambient scribe drafting your notes.

While the tools are made for you, a recent study showed only about 22% of studies developing clinical AI tools actually involved clinicians throughout the development process.

The other 78% was for, you guessed it, engineers building for engineers.

The FDA has authorized over 1,200 AI-powered medical devices, primarily in radiology, cardiology, and pathology. The memo argues that speed without physician involvement is creating tools that don't fit real workflows, embed biases from unrepresentative training data, and may actually make you a worse doctor over time.

Take a second to reflect on that.

Researchers call it "deskilling", where your pattern recognition and clinical instincts quietly atrophy when a machine handles the heavy lifting. Anyone have brain fog already from talking to Claude too much?

If the algorithm goes down during your overnight shift, are you sharper or duller than the attending who trained without it?

The trust problem runs in both directions too.

A Pew survey found that 60% of Americans would be uncomfortable with their provider relying on AI in their care. And physicians aren't much warmer to the idea. When clinicians see an AI tool make errors, they’re less likely to use it even if the tool is generally accurate.

So we're stuck in a loop: tools built without physician input don't earn physician trust, which tanks adoption, which means the tools never get the real-world feedback they need to improve.

I’ll save you some time from reading the full memo. The high-level recommendations are the following:

This isn’t some future problem we can hand off to the next generation of physicians, it’s a now problem. These tools are already shaping standards of care.

With proper clinician input, AI can reduce diagnostic errors, improve accuracy in high-stakes cases like cancer detection, and free up time for direct patient care.

However, the white coats need to be let into the room to help design it.

Physicians are inundated with new companies and tools each week.

To make all this easier to digest, we’d like to start writing company breakdowns so you can understand what these new tools do, how they impact your work and what’s worth your time.

Any companies you’d like to learn more about? Reply to this email and make a request.

Best,

M&H