Three Years In: The Fastest Adoption Curve We Have Ever Measured

Last month we wrote about the new front door to healthcare: a billion health questions a day answered by Google's AI, 230 million health conversations a week on ChatGPT, one in three US adults consulting a chatbot before a clinician. Some readers asked a fair question: is this a hype cycle, or is something structurally different happening?

So we went looking for the most sober people in the room. Not the founders. Not the futurists. The economists.

The skeptics changed their minds

For three years, economists were the designated skeptics of the AI story. Rising unemployment among young graduates? Interest rates. Layoffs blamed on AI? Executives looking for cover. Predictions of mass displacement? A failure to learn from every previous technology panic.

That position has collapsed. In April, the New York Times documented the shift: economists who dismissed ChatGPT at launch now describe the arrival of reasoning models and AI agents as a potential industrial-revolution-scale event. A Brookings scholar realized she no longer needed the basic research work she once hired college students to do. And Erik Brynjolfsson of Stanford, noting that technology usually takes decades to show up in productivity statistics, said plainly: "I don't think it's going to be decades this time."

Then came the statement. In a public declaration titled "We Must Act Now," more than two hundred signatories - including a dozen Nobel laureates, a former Fed chair, and leaders from every major AI lab - endorsed three sentences that would have been unthinkable from this group in 2023: AI may become radically more powerful within ten years; it could drive a transformation larger than the Industrial Revolution over a vastly shorter time frame; and institutions must start building for that world now.

When the people whose job is to not believe in hype start signing statements, the prudent read is that the base case has moved.

What the adoption data shows

Stanford's Digital Economy Lab now maintains live dashboards tracking AI's economic footprint, and the adoption numbers are the fastest we have ever measured for a general-purpose technology.

Fifty-eight percent of American adults report using generative AI - three years after the product category effectively launched. Electricity took decades to reach that share of households. The internet took well over a decade. Among people who use these tools, nearly 90% use them weekly. At work, adoption is closing in on half of all workers, and roughly one in five firms tells the Census Bureau it used AI within the past two weeks.

And the labor market is starting to whisper. Stanford's "Canaries" dashboard, built with ADP payroll data, finds that employment for workers aged 22 to 25 in the most AI-exposed occupations has been declining since 2024, and the decline is deepening - while every other group grows. The pattern shows up precisely where the theory says it should: in occupations where AI is used to fully automate tasks rather than assist with them.

Health care is not watching from the sidelines

The same acceleration is visible inside medicine - just distributed unevenly.

On the consumer side, it is essentially complete. AI-generated summaries now appear at the top of 84-92% of health-related Google searches. Before a patient ever reaches us, an AI has already answered.

On the clinician side, it is moving faster than any prior health IT adoption we can recall. OpenEvidence, an AI evidence-retrieval tool, is now used by roughly two-thirds of US physicians - about 650,000 doctors - and a recent American Medical Association survey found more than 80% of physicians now use some form of AI. Ambient AI scribes became the most broadly adopted clinical AI category in a single year: physicians report up to 83% less time writing notes, Northwestern reported a 112% ROI, and one leading platform reached 63% of hospitals running Epic. The FDA has now authorized more than 1,350 AI-enabled devices. On structured diagnostic evaluations, frontier AI systems score 85.5% on cases where experienced physicians working unaided score about 20%.

And yet. Stanford's Transformation Tracker follows twelve macro indicators of economy-wide upheaval, and today zero of the twelve are flashing. Only 5% of reviewed clinical AI studies used real patient data. Just 2.4% of FDA-authorized AI devices with clinical studies were backed by randomized trials. Firms, for all their enthusiasm, forecast surprisingly modest new adoption over the next three years.

The gap is widening

Put those two pictures side by side and you get the defining tension of this moment: the technology and the consumer are sprinting, and the institutions are walking.

That gap will not persist because the technology slows down. It will close one way or the other - either because institutions learn to adopt at something closer to consumer speed, or because consumers and disruptors route around the institutions. We wrote about the second scenario in the front door series. This series is about the first.

What we have learned watching AI move through our own health system: the slowness is not incompetence, and it is not simple bureaucracy. Deploying real AI inside a health system turns out to look almost exactly like a process we have run for a century - hiring an employee - with every safeguard that implies. In the next post, we will lay out that framework, and why it explains both why health systems are slow and what it would actually take to be fast.