What 400,000 Reddit Posts Reveal About GLP-1 Drug Side Effects — and What Clinical Trials Missed

On April 10, 2026, researchers at the University of Pennsylvania published a study in Nature Health that applied large language models to more than 400,000 Reddit posts spanning six years [1]. The study identified two categories of symptoms that current drug labeling does not prominently address: reproductive symptoms, including menstrual irregularities and intermenstrual bleeding, and temperature-related complaints such as chills and hot flashes [2]. The findings land at a moment when nearly 12% of American adults report having used a GLP-1 agonist [3], and when the pharmaceutical industry is projecting an additional $400 billion in US GLP-1 revenue between 2025 and 2030 [4].

The question is whether AI-powered social media analysis represents a genuine advance in drug safety monitoring — or an unreliable shortcut that generates more noise than signal.

What the Study Found

The Penn team — led by senior author Sharath Chandra Guntuku, a research associate professor in computer and information science, with first author Neil Sehgal, a doctoral student, alongside Lyle Ungar and Jena Shaw Tronieri — used GPT and Gemini to map user-described symptoms from Reddit to the Medical Dictionary for Regulatory Activities (MedDRA), the standardized terminology that clinicians and regulators use to classify adverse events [2].

Approximately 44% of the nearly 70,000 Reddit users in the sample reported at least one side effect [1]. Gastrointestinal complaints dominated: nausea appeared in 36.9% of posts reporting side effects, followed by fatigue (16.7%), vomiting (16.3%), constipation (15.3%), and diarrhea (12.6%) [1].

Source: Nature Health / Penn Engineering

Data as of Apr 10, 2026CSV

The study's more striking claims involve symptoms that fall outside well-established GLP-1 side effect profiles. Nearly 4% of users reported menstrual irregularities — a figure that would be substantially higher in a female-only sample [2]. Temperature-related complaints, including chills and hot flashes, also appeared at rates the researchers considered notable compared to their representation in prescribing labels and clinical trial reports [1].

"Some of the side effects we found, like nausea, are well known, and that shows that the method is picking up a real signal," Guntuku said. "The underreported symptoms are leads that came from patients themselves, unprompted, and clinicians could potentially pay attention to them" [2].

How These Rates Compare to Clinical Trial Data

The pivotal SURMOUNT trials for tirzepatide (Mounjaro/Zepbound) and the SUSTAIN trials for semaglutide (Ozempic/Wegovy) reported gastrointestinal adverse events in 40–70% of participants, with some studies reaching 85% [5]. Nausea was consistently the most frequent complaint across these trials, and across the SURMOUNT-1 through SURMOUNT-4 trials, GI events were mostly mild-to-moderate in severity and clustered during dose escalation periods [6].

What the clinical trials did not prominently report were reproductive symptoms or temperature-related complaints at rates comparable to what the Reddit analysis surfaced. Fatigue, which ranked second in the Reddit data at 16.7%, receives limited attention in standard prescribing information despite being one of the most commonly reported complaints in online patient communities [1].

The researchers hypothesize that GLP-1 drugs may affect the hypothalamus in ways that influence hormone regulation, which could explain both the reproductive and temperature-related symptoms [1]. But they are explicit about the limits of this claim: "We can't say that GLP-1s are actually causing these symptoms," Sehgal said. "We think that's a signal worth investigating" [2].

Methodology: How They Filtered 400,000 Posts

The study used large language models to perform what the researchers call "computational social listening" [2]. The AI models parsed Reddit posts for references to semaglutide and tirzepatide, then mapped the unstructured language — "feeling freezing all the time," "my period is all over the place" — to standardized MedDRA codes [1].

The researchers did not publicly disclose which specific subreddits were included or their precise filtering methodology for distinguishing verified GLP-1 users from casual commenters, caregivers, or users discussing the drugs secondhand [2]. The study covers posts from nearly 70,000 users, but the proportion who were confirmed patients versus spectators or critics is not broken out in available reporting on the study [1][2].

This is a meaningful gap. Prior social media pharmacovigilance research has found that 68% of social media mentions related to drug safety are irrelevant — posts discussing news articles, joking about drugs, or describing others' experiences rather than reporting personal symptoms [7]. Without rigorous user-level verification, inflated prevalence rates are a real risk.

The Track Record of Social Media Pharmacovigilance

The Penn study is part of a rapidly growing field. Academic research on GLP-1 pharmacovigilance has surged, with 583 papers published in 2025 alone — up from just 84 in 2022 [8].

Source: OpenAlex

Data as of Jan 1, 2026CSV

But the track record of social media-based drug safety surveillance is mixed. In a 2024 FDA pilot program involving six companies, false positive rates reached 97% for rarely prescribed medications — meaning the vast majority of "signals" detected in social media were noise [7]. The FDA's stated goal was to bring false positives below 15%, a target it has not yet publicly reported meeting [7].

On the other hand, social media has occasionally detected real safety problems faster than traditional reporting systems. One study found that association rule mining detected fluoxetine-induced suicidal thoughts one year before formal regulatory action, and simvastatin-induced kidney disease six years in advance [9]. Another analysis found that social media detected a diabetes drug safety signal 47 days before formal reports, though only after filtering out 68% of irrelevant mentions [7].

A 2021 scoping review in JMIR Public Health and Surveillance concluded that the ability of social media to detect drug safety signals that later resulted in FDA black box warnings, label changes, or withdrawals varied significantly by drug and by methodology, and that confirmatory studies with positive controls remained necessary to validate any social media-derived signal [10].

The FDA Reporting Lag Problem

The traditional pharmacovigilance system relies on the FDA Adverse Event Reporting System (FAERS), a database of voluntary reports from healthcare professionals and patients that is updated quarterly [11]. FAERS data have contributed to more than half of all post-market safety-related label changes [9], but the system has well-documented delays. Reports can take months to reach the FDA after a patient experiences an adverse event, and analysis of accumulated data takes additional time.

For GLP-1 drugs specifically, the FDA added delayed gastric emptying to the semaglutide label only after multiple reports of rare pulmonary aspiration cases during anesthesia — events that patients had been describing in online communities for months prior [12]. The median time-to-onset for GLP-1-related neurological adverse events reported to FAERS was 32 days, with 45% of events occurring within 30 days of starting treatment [11].

"Clinical trials are the gold standard, but by design, they are slow," Guntuku said [2]. The Penn team argues that AI-assisted social media monitoring could compress the timeline between patients experiencing symptoms and regulators becoming aware of patterns.

Historical comparisons underscore the stakes. The Vioxx recall in 2004 came after the drug had been on the market for five years and had been linked to thousands of heart attacks and strokes — safety signals that accumulated gradually through FAERS and post-market studies [9]. The thalidomide disaster of the late 1950s and early 1960s predated modern pharmacovigilance entirely [9]. Neither case would have been caught by social media analysis, but they illustrate the consequences of detection delays.

Who Is Missing from Reddit

Reddit's user base is not representative of the broader GLP-1 patient population. The average Reddit user is approximately 23 years old. As of 2026, 59.8% of users are male and 39.1% are female [13]. The platform is disproportionately American, urban, and concentrated in technology-adjacent professions [14].

GLP-1 users look different. According to RAND, women between ages 50 and 64 have the highest usage rate for GLP-1 agonists, with one in five women in that age group reporting having used them [3]. These women are underrepresented on Reddit. So are older adults with type 2 diabetes, rural populations, non-English speakers, and patients in lower income brackets — all groups with significant GLP-1 exposure.

This demographic mismatch means the Reddit data is more likely to surface side effects experienced by younger, tech-literate users and less likely to capture adverse events affecting older, sicker, or more diverse populations. Reproductive symptoms may have emerged partly because Reddit's user base skews toward women of childbearing age relative to the overall GLP-1 patient population, or they may represent a genuine underreported signal. The data cannot distinguish between these explanations.

The Market Context: Scale and Financial Exposure

GLP-1 prescriptions have more than tripled since 2020 [15]. As of September 2025, they account for 6.5 out of every 100 US prescriptions [15]. Between January 2018 and September 2025, approximately 2.45 million patients were prescribed a GLP-1 receptor agonist, generating over 12.2 million total prescriptions [15].

Source: Truveta Research

Data as of Dec 31, 2025CSV

The financial stakes for manufacturers are enormous. Eli Lilly reported $12.7 billion in Q1 2025 revenue (up 44% year-over-year), while Novo Nordisk brought in $11.8 billion (up 19%) [4]. Eli Lilly holds approximately 57% of the GLP-1 market as of mid-2025 [4]. Cumulative US GLP-1 revenue reached $71 billion by end of 2024, with projections of $400 billion more between 2025 and 2030 [4].

New label requirements or safety communications could affect that trajectory. Novo Nordisk already faces a multidistrict litigation involving over 2,190 plaintiffs related to semaglutide and retinal vein occlusion, with potential liabilities exceeding $220 million [16]. If newly identified side effects — reproductive or otherwise — trigger additional label changes, Risk Evaluation and Mitigation Strategy (REMS) programs, or further litigation, the financial exposure could expand significantly.

The FDA approved oral Wegovy in December 2025 [17], and Eli Lilly's oral weight-loss drug has also entered the market [18]. Expanded access to oral formulations is expected to further increase the patient population, amplifying both the public health significance and the commercial implications of any new safety signals.

Funding, Conflicts, and Peer Review Status

The Penn study reports no outside funding [2]. Three of the four authors — Guntuku, Sehgal, and Ungar — disclose no conflicts of interest. The fourth, Jena Shaw Tronieri, a senior research investigator at Penn's Center for Weight and Eating Disorders, reports receiving an investigator-initiated grant on behalf of the University of Pennsylvania from Novo Nordisk, as well as consulting fees from Currax Pharmaceuticals [2].

The study was published in Nature Health, a peer-reviewed journal in the Nature portfolio, meaning it underwent editorial and peer review prior to publication [1]. An earlier version of related work by some of the same researchers appeared as a preprint on arXiv in April 2024, under a different title and with a different scope [19].

The Novo Nordisk connection through Tronieri is worth noting given that the study examines side effects of drugs that Novo Nordisk manufactures, though the disclosure of no outside funding suggests the company did not finance this particular research.

The Case Against Reliability

The strongest argument that this methodology is scientifically unreliable rests on several points.

First, social media posts are self-selected. People who experience side effects are more likely to post about them than people who tolerate a medication well, creating inherent selection bias [14]. This is distinct from clinical trials, where adverse events are systematically collected from all participants regardless of severity.

Second, causality cannot be established. A Reddit user taking semaglutide who reports menstrual irregularities may be experiencing a side effect of the drug, a symptom of rapid weight loss itself, a pre-existing condition, stress, or coincidence. Without medical records, dosing information, or clinical context, the Reddit data cannot distinguish among these explanations [2].

Third, the false positive rate in prior social media pharmacovigilance work has been high. The FDA's own pilot found 97% false positives for rare drugs [7]. Even for more commonly prescribed medications, the signal-to-noise ratio remains a persistent challenge.

Fourth, formal clinical validation of social media-derived signals has been limited. The 2021 JMIR scoping review found that while social media sometimes detected signals ahead of regulatory action, the proportion of those signals that were later confirmed through rigorous clinical study varied widely and was often not systematically tracked [10].

The Case for Taking It Seriously

The counterargument is equally substantive. Clinical trials for GLP-1 drugs typically enrolled thousands of participants over 12–68 weeks [5][6]. The Reddit dataset covers nearly 70,000 users over six years — a far larger sample observed over a far longer period [1]. Clinical trials also systematically exclude populations with comorbidities, polypharmacy, or complex medical histories — exactly the populations most likely to experience unexpected drug interactions or atypical side effects.

The fact that the Reddit analysis independently identified the same well-known GLP-1 side effects (nausea, vomiting, constipation) at rates broadly consistent with clinical trial data provides a form of internal validation [2]. If the methodology were purely generating noise, it would not consistently reproduce known signals.

And the reproductive symptoms flagged by the study are biologically plausible. GLP-1 receptors are expressed in the hypothalamus, which regulates both appetite and reproductive hormones [1]. Rapid weight loss itself is known to affect menstrual cycles. Whether GLP-1 drugs have an independent effect on menstruation beyond their weight-loss mechanism is an open question — but one that formal clinical investigation could resolve.

What Happens Next

The Penn researchers have not called for label changes or regulatory action based on their findings alone. They describe their work as generating hypotheses for further study [2]. Whether the FDA or the pharmaceutical companies will pursue formal investigation of the reproductive and temperature-related signals remains to be seen.

Meanwhile, GLP-1 pharmacovigilance research continues to accelerate, with hundreds of papers published annually examining adverse events through FAERS analysis, electronic health records, and now social media [8]. The methodological question — whether AI-powered analysis of patient-generated content can reliably complement traditional drug safety systems — is likely to be answered incrementally, as more social media-derived signals are either validated or debunked through formal clinical study.

For the millions of Americans currently taking GLP-1 medications, the practical takeaway is narrower: if you are experiencing menstrual irregularities, persistent chills, or unexplained temperature fluctuations while on semaglutide or tirzepatide, the Penn study suggests these symptoms may be more common among GLP-1 users than prescribing labels currently indicate — but not that these drugs have been proven to cause them.