A machine-learning sweep of roughly 400,000 Reddit posts has surfaced a cluster of symptoms that users of Ozempic, Wegovy, Mounjaro and Zepbound say they experienced but that rarely appear on the drugs' official labels — menstrual changes, chills, hot flashes and persistent fatigue among them. The findings, reported by a research team that used large language models to trawl years of patient chatter, do not prove the medications caused any of those problems. But researchers say the patterns point to signals that randomized trials, with their narrow questionnaires and limited follow-up, may simply never have asked about.

What the Reddit analysis actually found

The study is part of a fast-growing genre of "pharmacovigilance by social media": mining the candid, unscripted testimony of millions of patients for adverse events that formal reporting systems like the FDA's FAERS capture only sparsely. The posts analyzed spanned the major GLP-1 receptor agonists — semaglutide, sold as Ozempic and Wegovy, and tirzepatide, sold as Mounjaro and Zepbound — a class that has become one of the most commercially significant in modern medicine, with tens of millions of prescriptions written worldwide and analysts projecting a market worth well over $100 billion by the end of the decade.

According to Science Daily's account of the work, the unexpected symptoms clustered in ways that suggested a shared biological thread rather than random noise. Hormonal and thermoregulatory complaints — cycle disruption, hot flashes, chills — appeared repeatedly alongside fatigue, a combination that sits awkwardly next to the well-documented gastrointestinal profile of the drugs.

The research team was explicit about the limits of its method: the drugs cannot be said to have caused these symptoms, but the recurring patterns may reveal overlooked signals that deserve formal study.

Association is not causation — and the internet is not a clinic

That caveat is the story's spine. Reddit users are self-selected, self-diagnosing and frequently taking compounded versions of the drugs whose purity and dosing are not guaranteed. Weight loss itself alters menstrual cycles, mood and energy. Rapid caloric restriction produces chills and fatigue on its own. Any of these confounding factors could generate the same signal.

Yet the value of the approach is precisely its breadth. A trial of a few thousand participants, run for 68 weeks and focused on weight and glycemic endpoints, is powered to detect common effects, not rare or slow-building ones — and certainly not ones participants weren't prompted to report. "Real-world" evidence from forums can flag hypotheses; only controlled studies can test them.

How different outlets framed the same science

• Science Daily led with methodology and caution, emphasizing that the AI had uncovered candidate side effects worth investigating rather than confirmed ones.
• Forbes angled toward the consumer, packaging the findings as "hidden side effects to consider before treatment" — a pre-decision framing aimed at prospective patients.
• MSN treated it as a service story, pairing risks with the question of what happens when patients stop the drugs, a major concern given trial data showing that most people regain roughly two-thirds of lost weight within a year of discontinuation.
• Psychology Today took the widest swing, tying the side-effect research to a separate AI-driven discovery of a "natural Ozempic" — a framing that positions AI as both detector of problems and source of solutions.

The Stanford thread: an AI-discovered 'natural Ozempic'

That second strand is a genuinely separate piece of science. Stanford researchers, per Science Daily, used an AI screening platform to identify a naturally occurring peptide that appears to suppress appetite in a way that mimics GLP-1 drugs — but, in early experiments, without the nausea, vomiting and other gastrointestinal burden that drives many patients to quit. The work is preclinical: results in cells and animal models are a long way from a pharmacy shelf, and the history of weight-loss science is littered with promising molecules that failed in humans. Still, the discovery illustrates a broader shift in drug development, where generative and predictive models are being used to search chemical and biological space far faster than traditional screening allows.

Read together, the two stories describe a feedback loop. AI is being used to find drug candidates that sidestep known side effects, while the same technology is being used to find side effects that trials missed. Both applications run on the same raw material: enormous, messy, human-generated data.

What it means for patients

For people currently on a GLP-1 — roughly one in eight U.S. adults say they have tried one, according to polling — the practical takeaway is not alarm but documentation. Clinicians interviewed about the class consistently recommend logging symptoms, timing them against dose escalations, and raising anything persistent or disruptive with a prescriber. Regulators, meanwhile, are likely to treat the Reddit analysis as a lead rather than a verdict, the way they have handled prior social-media-derived signals: as a prompt for epidemiological follow-up.

The deeper implication is about what counts as evidence. Trials remain the gold standard, but the gap between what a trial measures and what a patient experiences is where these AI tools now live — and increasingly, where the next round of research questions will come from.