How Our Wearables Are Crowdsourcing Health AI

Millions of people are quietly running one of the largest health studies in history - not in clinics, but on their wrists and fingers. Every ring, watch, and tracker is a tiny sensor, streaming data that, when pooled together, creates enormous datasets fueling the next wave of AI-driven health insights.

It’s funny to think how far we’ve come. Not too long ago, wearable tech meant a clunky pedometer clipped to your belt, dutifully counting steps until you got bored and tossed it in a drawer. Fast forward to 2025, and your jewelry doesn’t just judge your bedtime - it can survive a 25-hour Senate speech, log your stress while you argue policy, and even tell you that you’re not getting enough restorative sleep (as if you didn’t already know).

When Senator Cory Booker stood (and stood… and stood) for 25 hours straight to make the longest speech in Senate history, he wasn’t just breaking political records. He was also unintentionally stress-testing the Oura Ring. The poor device had no category for “talking passionately about policy without sitting down once,” so it logged his effort as three or four workouts. Booker racked up 20,000 steps, burned calories like a spin class instructor, and saw his heart rate peak at 131 bpm. Essentially, he got a week’s worth of cardio without leaving the Senate floor. The only thing missing? Sleep. The app didn’t even bother to log a zero. Just… nothing. Which is pretty much how Booker probably felt.

Meanwhile, in the land of quantified-self hobbyists, people are using the Bearable app to log their mood, migraines, and even duvet-hogging. One couple tracked how caffeine, video games, and bed-sharing impacted sleep quality using their Oura rings. Spoiler: caffeine after noon is bad, video games before bed are worse, and apparently one person always ends up with the blanket. The app itself was born on Reddit (where else?), proving that sometimes the best science starts with nerds, spreadsheets, and a stubborn refusal to sleep badly ever again.

Now, while most of us are just happy if our ring tells us we got “okay” sleep, researchers are turning wearables into engines of health prediction. Apple and collaborators recently trained a foundation model on 2.5 billion hours of wearable data from over 160,000 people. Instead of obsessing over raw sensor signals (every wiggle of an accelerometer), they focused on higher-level behavioral metrics: step counts, VO₂ max, gait steadiness, even how fast you climb stairs.

The result? A model that can help detect conditions ranging from diabetes to pregnancy, predict sleep quality, and generally act like that overly observant friend who always notices when you look tired. And when combined with sensor-level models, it gets even better. Think of it as Sherlock Holmes teaming up with Watson - but Watson is your wristwatch.

But Wait, Is AI Ready for All This? Here’s the twist: AI isn’t flawless. In medical evaluations, even GPT-5 shows both leaps and faceplants. It crushed factual recall tasks and tied for best in medical calculations (handy if you’re wondering how much Tylenol to take). But it also stumbled badly on fairness tests and structured database queries - not exactly confidence-inspiring if you want unbiased care or accurate EHR lookups.

A recent systematic review of LLMs for patient-facing use tell a similar story: progress, yes, but uneven and often outdated by the time the ink is dry. The GPT-5 MedHELM evaluation was written up by a software engineer - probably without tapping into “high reasoning” modes - yet it still offered a valuable fresh snapshot, especially since most published evaluations are already fossils by the time they appear. The AI relevance window is now shorter than the battery life of a first-gen Fitbit.

So what do we make of all this? On one hand, wearables + AI could transform health by making continuous, personalized monitoring possible. On the other, we’re still in a phase where your smart ring thinks you’re at the gym when you’re actually filibustering.

Maybe the real lesson is this: technology is only as useful as the stories we wrap around it. Cory Booker’s Oura readout became a tale of endurance politics. The Bearable app turned nightly routines into a data-driven couple’s experiment. Apple’s new foundation models show that the real gold is in behaviors, not just beats-per-minute. And AI? Well, it’s learning - sometimes faster than us, sometimes slower - but it’s definitely not boring.

Until then, keep wearing your gadgets. Because the more we wear, the more data Apple and others have to train on. But here’s the upside: that data is fueling a wave of innovation - from foundation models that can flag infections before you notice symptoms to specialized medical AIs that are laser-focused on healthcare reasoning.

REFERENCES

Erturk E, Kamran F, Abbaspourazad S, Jewell S, Sharma H, Li Y, Williamson S, Foti NJ, Futoma J. Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions. arXiv preprint arXiv:2507.00191. 2025 Jun 30. 

Irene S. Gabashvili Evaluating General-Purpose LLMs for Patient-Facing Use: Dermatology-Centered Systematic Review and Meta-Analysis medRxiv 2025.08.11.25333149; doi: https://doi.org/10.1101/2025.08.11.25333149  2025 Aug 11.