I've spent a lot of ink in this newsletter criticizing AI: the water costs, the rising bills, the forced adoption, the industry's coordinated messaging. I stand by all of it. But criticism without balance isn't honesty; it's just a different kind of bias.

So here's the other half of the truth: precision AI, aimed at specific medical and scientific problems, is already improving healthcare in ways that matter for real patients. Not someday. Now.

Consumer AI being force-fed to us through every app is one thing. Precision AI, models built for a defined scientific job and applied by domain experts, is another thing entirely. When AI reads an ECG to catch heart disease early, or maps tumor cells humans can't see, or, as this week's headline shows, surfaces medication side effects patients are experiencing but no trial ever captured, that's not hype. That's help.

You can challenge AI's real harms and still refuse to pretend it isn't doing real good.

Researchers at the University of Pennsylvania used AI to analyze more than 400,000 Reddit posts from nearly 70,000 users discussing GLP-1 drugs, the wildly popular weight-loss and diabetes medications semaglutide and tirzepatide. The study, published in Nature Health, used large language models to map how everyday people describe their symptoms onto the standardized medical vocabulary clinicians use. In doing so, it surfaced patient-reported side effects that may not be fully captured in clinical trials or drug labeling.

Two categories stood out: reproductive symptoms, including irregular menstrual cycles and unexpected bleeding, and temperature-related complaints like chills and hot flashes. Fatigue also ranked as the second most common complaint, despite reaching reporting thresholds in relatively few clinical trials.

Clinical trials are the gold standard, but by design, they're slow, and they're built to catch the most dangerous side effects, not necessarily the ones patients care most about day to day. As one researcher put it, online patient communities work like a neighborhood grapevine: people living with these medications swap notes in real time, sharing experiences that rarely make it into a doctor's visit or an official report. AI makes it possible to listen to that grapevine at scale and with a level of standardization that was simply impossible before.

There's a striking irony here worth pausing on. Back in July, I covered research showing that AI "deep research" tools could be poisoned by manipulated Reddit posts, that user-generated content was a vulnerability. This week's story is a mirror image: that same messy, unofficial, user-generated content, read carefully by AI, becomes an early-warning system for drug safety. Same platform. Opposite lesson. The tool isn't good or bad. The application is everything.

Watch for this approach to expand beyond Reddit and beyond English, which the researchers name as their next step, and watch for the obvious caution: Reddit users skew younger, more male, and US-based, so the signal isn't representative on its own. The real promise is as a complement to clinical trials, a fast early-warning layer, not a replacement for rigorous study. Watch also whether regulators start treating this kind of AI-assisted pharmacovigilance as a legitimate input.

This one lands close to home, because so many of our patients are on these medications, and OT is often where the day-to-day functional impact of a drug's side effects actually shows up. Fatigue, temperature dysregulation, menstrual changes—these aren't abstract data points to us. They affect activity tolerance, participation, energy management, and daily routines. They're what a patient is likely to mention to their OTP during a 45-60min session and not usually to their physician in a 15min visit.

That makes us part of this grapevine, whether we realize it or not. When a patient tells you their new medication is wiping them out by 2pm, you're hearing the same symptom this study captured, one patient at a time. Understanding that AI can now aggregate those concerns at scale helps us see our own observations as data worth documenting and escalating. And it's a reminder of something we've said all year: the patient's lived experience is a form of evidence. AI is finally getting good enough to listen to it. We've been listening all along.