AI finds GLP-1 side effects that 72-week clinical trials systematically missed.

The April 2026 study — what the AI actually found

In April 2026, researchers at the University of Pennsylvania published a study in Nature Health that did something Phase 3 clinical trials cannot: they listened to everyone. Using natural language processing (NLP) — a branch of artificial intelligence that reads and classifies text at scale — the team analyzed 410,198 Reddit posts spanning May 2019 through June 2025, capturing self-reported experiences from 67,008 people using semaglutide (Ozempic, Wegovy) or tirzepatide (Mounjaro, Zepbound) [1].

The headline number: 43.5% of users reported at least one adverse effect. The well-documented signals appeared as expected — nausea (36.9%), fatigue (16.7%), vomiting (16.3%), constipation (15.3%), and diarrhea (12.6%). These track reasonably with what clinical trials report, even if absolute rates diverge between trial populations and real-world users.

Then came the signals that are not in current prescribing information for either drug.

Menstrual irregularities appeared repeatedly and consistently — altered cycle length, skipped periods, unexpected breakthrough bleeding — in premenopausal women across both drug classes. Temperature dysregulation also emerged as a recurring pattern: cold flushes, persistent chills, hot waves outside any normal menopause context. Neither signal carries a warning, a mention, or a recommended monitoring protocol in either semaglutide's or tirzepatide's current labelling [1].

The authors were explicit: these were signals "not well captured in current labelling or trials" and warranted prospective investigation. That investigation has not yet been initiated or mandated.

67,000 real users told the AI what the trials couldn't detect. The signal was hiding in the populations that never made it into SUSTAIN or SURMOUNT.

The two signals Phase 3 never captured

Menstrual irregularities. The GLP-1 receptor (GLP-1R) is expressed in hypothalamic nuclei that regulate both energy balance and the hypothalamic-pituitary-gonadal (HPG) axis. GLP-1 receptor agonists act centrally on appetite-regulating circuits that overlap substantially with the circuits governing gonadotropin-releasing hormone (GnRH) and luteinizing hormone (LH) pulsatility.

Rapid weight loss of any cause is known to disrupt menstrual cycles via leptin-mediated effects on HPG signaling — this is why extreme caloric restriction and eating disorders cause amenorrhea. But the Penn researchers noted that the menstrual signal appeared in users who had not yet lost substantial weight, suggesting a direct pharmacological effect on hypothalamic GnRH pulsatility rather than a downstream weight-loss artifact [1, 2].

The practical implications are concrete. Women not trying to become pregnant who are relying on cycle regularity as part of family planning may need to update their contraceptive approach. Women who are trying to conceive may face a signal the current label does not prompt their prescriber to discuss. Neither group receives that conversation today because neither the labelling nor prescriber guidance has been updated to reflect it.

Temperature dysregulation. GLP-1 receptors are present in the hypothalamic thermoregulatory center and in brown adipose tissue (BAT). Animal data have consistently shown that GLP-1R activation influences thermogenesis and energy expenditure through both central and peripheral mechanisms — in part, why GLP-1 agonists modestly increase resting metabolic rate beyond the effects of weight loss alone [3].

The real-world pattern — persistent chills in some users, hot flushes in others — suggests heterogeneous receptor expression or dose-dependent effects that standardized trial designs, enrolling carefully selected participants at fixed titration schedules, were not positioned to capture. The mechanism is plausible. The signal simply wasn't visible in trials because trials weren't designed to surface it.

Why 72-week trials structurally cannot see this

The major Phase 3 programs for semaglutide and tirzepatide — SUSTAIN, STEP, SURMOUNT, SURPASS — were designed primarily to answer efficacy questions: how much weight does the drug produce, and does it improve glycemic control? They were powered on primary endpoints like percent body weight change from baseline at 68–72 weeks [4, 5].

The structural limitations that follow are layered.

Enrollment selection systematically excludes the most informative patients. Pivotal trials exclude participants with a long list of comorbidities, prior hormonal conditions, irregular menstrual cycles, and other baseline deviations. This removes the populations most likely to exhibit or report hormonal adverse effects. The people running the real-world experiment — 15 million users with every kind of medical history — were never in the trial [6].

Structured adverse event collection versus free-form reporting. Phase 3 trials collect adverse events through standardized case report forms and MedDRA (Medical Dictionary for Regulatory Activities) coding. Participants are prompted about pre-specified events. If "menstrual irregularity" is not on the checklist, and if participants don't spontaneously volunteer it during a 15-minute research appointment, it doesn't enter the dataset. The NLP study worked in reverse: it let 67,000 users describe their experience in their own words, then classified the themes that emerged organically.

Seventy-two weeks is not four years. Temperature dysregulation and hormonal effects may emerge slowly with sustained receptor engagement, or may be most prominent after users reach steady-state at maintenance doses. A trial with a 72-week primary endpoint cannot detect effects that accumulate or emerge at 18, 24, or 36 months.

Statistical power against rare signals. Even the largest Phase 3 trials enrolled 600–2,000 participants for most primary efficacy endpoints — the SELECT cardiovascular trial being the large-scale exception at 17,604. A signal affecting 4–6% of users needs a much larger dataset to separate cleanly from background noise. Sixty-seven thousand users providing free-text accounts of their experience generates a fundamentally different signal environment [1].

None of this means Phase 3 trials were conducted improperly. They were designed for regulatory approval — not comprehensive post-market surveillance. The problem is that no equally rigorous system was built to fill the surveillance gap after approval. AI pharmacovigilance has moved into that vacuum.

Real-world adverse signals already documented

The menstrual and temperature findings are new additions to a growing list of GLP-1 adverse signals that post-market surveillance and pharmacovigilance databases have surfaced since Ozempic entered mass use in 2021.

Gastroparesis and severe gastric dysmotility. GLP-1 agonists slow gastric emptying — this is documented in labelling. But the degree of slowing in some real-world users significantly exceeds trial documentation, leading to gastroparesis cases severe enough to interfere with anesthesia (undigested stomach contents during elective surgical procedures) and to cause malnutrition in outlier patients. The FDA updated its gastroparesis language in 2023 following post-market case accumulation [7].

Ophthalmic events. A 2026 pharmacovigilance analysis across FDA FAERS (2005–2025) and WHO VigiBase found strong reporting signals for nonarteritic anterior ischemic optic neuropathy (NAION) with semaglutide — reporting odds ratios of 40.18 in VigiBase and 31.46 in FAERS, representing thousands of documented cases. The authors concluded that "ocular adverse events are incompletely characterized" in current prescribing information for GLP-1 agents [8].

Suicidality signals. EMA and FDA investigations were initiated in 2023 after reports of suicidal ideation filtered through national pharmacovigilance databases. Neither agency confirmed a causal relationship in their 2024 assessments, but neither formally closed the signal with a clean exoneration. GLP-1 receptors in the central nervous system's reward and mood-regulating circuitry make the mechanism biologically plausible, and the question remains open [9].

Lean mass and facial volume loss. As documented in the GLP-1 weight regain analysis, studies consistently show that 20–40% of weight lost on GLP-1 agonists comes from lean mass — disproportionate relative to most caloric restriction models at equivalent deficits. The facial hollowing effect ("Ozempic face") is a documented physiological finding, not aesthetic overreaction. It is not prominently warned in labelling [4].

Thyroid and pancreatic signals. Animal studies for semaglutide produced C-cell thyroid tumors in rodents, which is why the drug carries a black-box warning for patients with personal or family history of medullary thyroid carcinoma. Pancreatitis reports have accumulated in FAERS across multiple GLP-1 agents. The incidence relative to background risk remains debated, but has not resolved with larger datasets [11, 12].

Fifteen million users and a labelling gap

By early 2026, an estimated 15 million Americans were prescribed a GLP-1 receptor agonist, with global figures above 20 million when European, East Asian, and Gulf markets are included [6]. The scale makes this a pharmacovigilance challenge unlike most others in pharmaceutical history.

Pharmacovigilance has historically operated through passive systems: FDA's FAERS database, WHO's VigiBase, and national systems like the UK's Yellow Card. These depend on healthcare providers or patients voluntarily submitting reports when they notice something unexpected. The structural problem with passive surveillance for a drug taken by tens of millions: a 3% adverse event rate across 15 million users is 450,000 people — dispersed across every primary care clinic in the country, each reporting to nobody.

NLP pharmacovigilance bypasses the passive reporting bottleneck by analyzing where patients actually talk about their experiences. It works for a specific reason: patients are more likely to describe reproductive side effects to an anonymous online community than to submit a MedWatch form, or even mention the issue in a quarterly 15-minute follow-up appointment. The signal lives in data already generated. The missing piece is the analytical layer to surface it.

The Penn researchers called explicitly for "regulatory re-evaluation and prospective investigation" of the signals they found [1]. As of this writing, no prospective study of menstrual effects on GLP-1 agonists has been registered by either Novo Nordisk or Eli Lilly.

Passive pharmacovigilance was built for drugs prescribed to thousands. GLP-1s are now prescribed to tens of millions. The math no longer works.

What this means practically

The right response is not to stop taking GLP-1 drugs. These are drugs with demonstrated efficacy for weight loss, glycemic control, and cardiovascular outcomes that older treatments couldn't match. The SELECT trial data on cardiovascular mortality reduction in people without diabetes are real and clinically meaningful [5]. The neurological signal across Parkinson's research is emerging with credibility. The class works.

The question is whether people using these drugs have an accurate picture of the full adverse effect profile. Based on what the current literature and the new AI pharmacovigilance data show, a few concrete points follow.

• Women of reproductive age should discuss menstrual effects proactively. Don't wait for a labelling update. If cycle irregularities appear on a GLP-1 agonist, document the timing, correlate it with dosing, and raise it with your prescriber. This signal needs prospective data, and clinical documentation contributes to it.
• Pre-operative disclosure is now standard of care. Anesthesiologists increasingly ask about GLP-1 use specifically because of gastroparesis risk. For any elective procedure, disclose GLP-1 use with the same priority as anticoagulants.
• Lean mass monitoring is underused. Most prescribers track weight and HbA1c. Few order DEXA scans or serially measure body composition. For users motivated primarily by body composition outcomes, this data matters. Current evidence supports protein intake above 1.2–1.6 g/kg and resistance training as protective against the muscle-loss fraction of weight lost on GLP-1 agents [13].
• Report what you experience. FDA MedWatch (FDA.gov/medwatch) accepts consumer reports directly. Five minutes to file. The pharmacovigilance signal depends on accumulation — and AI systems can only analyze what has been recorded.

The broader structural takeaway: Phase 3 trials are built to satisfy regulatory approval thresholds, not to build a comprehensive safety map of real-world population use at scale. They were never designed to do that, and regulatory frameworks don't require it. AI pharmacovigilance fills the gap that trials were never designed to provide.

For a drug class that keeps expanding in patient volume, indication breadth, and now oral formulations reaching entirely new demographics — alongside a next-generation incretin pipeline enrolling tens of thousands more — that gap matters more with every passing quarter.

References

1. Sehgal NKR, Tronieri JS, Ungar L, et al. Self-reported side effects of semaglutide and tirzepatide in online communities. Nature Health. 2026. doi:10.1038/s44360-026-00108-y
2. Rosenfield RL, Ehrmann DA. The Pathogenesis of Polycystic Ovary Syndrome (PCOS). Endocr Rev. 2016;37(5):467-520. doi:10.1210/er.2015-1104
3. Lockie SH, Stark R. GLP-1 receptor agonists and thermogenesis: central mechanisms and brown adipose tissue engagement. Peptides. 2020;125:170198. doi:10.1016/j.peptides.2019.170198
4. Wilding JPH, Batterham RL, Calanna S, et al. Once-weekly semaglutide in adults with overweight or obesity. N Engl J Med. 2021;384(11):989-1002. doi:10.1056/NEJMoa2032183
5. Lincoff AM, Brown-Frandsen K, Colhoun HM, et al. Semaglutide and cardiovascular outcomes in obesity without diabetes (SELECT). N Engl J Med. 2023;389(24):2221-2232. doi:10.1056/NEJMoa2307563
6. KFF. Poll: 1 in 8 adults say they are currently taking a GLP-1 drug for weight loss, diabetes, or another condition. Kaiser Family Foundation; 2025. Available at kff.org/public-opinion/
7. U.S. Food and Drug Administration. FDA Drug Safety Communication: GLP-1 receptor agonists — updated information regarding risk of gastroparesis. FDA; 2023. fda.gov/safety.
8. Cheng X, Jiang Z, Li G, et al. Pharmacovigilance analysis of ophthalmic adverse events associated with GLP-1 receptor agonists. J Endocrinol Investig. 2026;49(2):425-433. doi:10.1007/s40618-025-02712-3. PMID:41021211
9. European Medicines Agency. Assessment report: GLP-1 receptor agonists — suicidal and self-injurious thoughts. EMA/CHMP. 2024. ema.europa.eu.
10. Blundell J, Finlayson G, Axelsen M, et al. Effects of once-weekly semaglutide on appetite, energy intake, energy expenditure, and gastric emptying. Diabetes Obes Metab. 2017;19(9):1242-1251. doi:10.1111/dom.12932
11. Ruggiero R, Longo M, Mascolo A, et al. Real-world safety comparison of liraglutide and semaglutide in weight management: European pharmacovigilance data. Eur J Pharmacol. 2025;985:177051. PMID:40716637
12. Yang Z, Lv Y, Yu M, et al. GLP-1 receptor agonist-associated tumor adverse events: a real-world FAERS analysis 2004-2021. Front Pharmacol. 2022;13:1008908. doi:10.3389/fphar.2022.1008908. PMID:36386208
13. Eisa N, Barood O. Lean mass changes with incretin therapy versus lifestyle intervention: a systematic review and meta-analysis of randomised controlled trials. Diabetes Obes Metab. 2026;28(6):4818-4827. doi:10.1111/dom.70666. PMID:41877354