Yet, clinicians also saw promise:

Belief that GenAI improves accuracy: 4.30

Institution-customized GenAI viewed even more favorably: 4.96

(7-point scale)

Ratings of the care experience also dropped — from 4.48 ⭐️ to 3.08 ⭐️ (5-star scale).

Framing GenAI as a verification tool helped (clinical skill 4.99, competence 4.94), but the gap remained.

The numbers are striking.

In our randomized experiment of 276 clinicians, a physician who used GenAI as a decision aid tool was rated far lower:

Clinical skill: 3.79 vs 5.93

Overall competence: 3.71 vs 5.99

(7-point scale)

Safety gap

@fda.gov-cleared AI devices from publicly traded firms are recalled far more often: up to 30 × compared with those from private firms (14.4% vs 1.3% of cleared devices)

Development and commercialization models corrects with patient risk and should guide oversight.

(5/7)

Tech under the hood

Deep learning now powers half of new @fda.gov-cleared devices.

Transparency is improving, yet 62% of all devices still give little or no detail about how their AI works.

(4/7)

AI Clearance curve

Average @fda.gov clearances jumped from 1.4 devices per year in 1995-2014 to 146 per year in 2020-24—a 100-fold surge.

Total count went from 27 in the first 20 years to 729 in the last five.

In-house development drives nearly all growth.

(3/7)

Who is building what?

69% of @fda.gov-cleared AI device manufacturers are private, but public firms make more devices per company.

General Radiology leads with 32%, followed by cardiovascular (18%) and neuropsychiatry (15%).

A booming yet scattered market.

(2/7)

Honored to lead an incredible team of @jhu.edu medical students on the first full look at how @fda.gov-cleared medical AI devices are developed and commercialized, now in @ai.nejm.org.

We tracked 950 AI medical devices. The results may surprise you: bit.ly/fdaai25

(1/7)

5/5 As a top exporter leading in diagnostic and lab reagents, a prolonged trade war could render the global scientific supply chain more fragile, costly, and unreliable.​

We may be at the onset of a tariff-induced chaos period.​

As I told @celestebiever.bsky.social of @nature.com, “These aren’t luxury items. They’re the core infrastructure of modern science.”

2/5 These tariffs hit essentials—from basic labware to advanced instruments—just as research institutions face severe financial strain. NIH funding is down 60% in Q1, and indirect cost recovery is under fire.

This isn’t belt-tightening; it could be a breaking point.

6/ Now enter AI, and the web of trust becomes even more complex.

Patients need to trust that their doctors are using AI wisely.
Doctors need to trust that AI supports—not replaces—their judgment.

And both must trust the system that’s deploying these tools.

4/ At its core, trust is about navigating uncertainty.
It’s about being vulnerable—relying on others with the hope that they’ll act competently and in good faith.

In the paper, we unpack that idea across different levels.

1/ Can we trust AI to deliver health care?
And can AI trust us?

www.nature.com/articles/s44...

So excited to share our new perspective in @natureportfolio.nature.com Digital Medicine, written with the brilliant Madeline Sagona, Mario Macis, & Michael Darden.