The new GLP-1 weight-loss injections have become one of the most significant health developments in recent years, transforming countless lives. However, they also carry side effects—some anticipated, others serious, and some only now coming to light.
This was starkly illustrated a few weeks ago when I treated a 27-year-old teacher in the emergency department suffering from severe abdominal pain that required morphine. He had been struggling with his weight and decided to try Mounjaro to improve his health. Within weeks of starting the medication, he was hospitalised. His blood tests revealed extremely high levels of an enzyme produced mainly by the pancreas, indicating pancreatitis—inflammation of the pancreas that can be life-threatening in severe cases. He was admitted, given intravenous fluids, and taken off Mounjaro permanently. Fortunately, his pancreas recovered.
Pancreatitis is a known potential risk of weight-loss drugs, but his case served as a stark reminder that even highly beneficial treatments can cause real harm. Medicine has traditionally excelled at learning from formal research, such as clinical trials and official reporting systems where doctors and patients report suspected side effects. However, it is far less adept at listening to large numbers of patients simultaneously when a treatment moves from small-scale testing to widespread use.
Once a treatment is used by millions, the question shifts from whether it works—we know it does—to how quickly we can identify problems as they arise. This is why a new study from the University of Pennsylvania, published in Nature Health, is significant. Researchers analysed the side effects of GLP-1 drugs and, in addition to well-known issues like nausea, constipation, and diarrhoea, identified previously overlooked effects such as menstrual changes (including heavy bleeding and irregular cycles), chills, hot flushes, fatigue, and sleep difficulties.
The study does not prove these drugs caused all these symptoms, but it highlights patient-reported effects that may not be captured in formal trials or drug leaflets. The researchers used artificial intelligence to analyse over 410,000 Reddit posts discussing semaglutide and tirzepatide, the active ingredients in Ozempic and Mounjaro. Social media posts are valuable for doctors not because they constitute perfect science—they clearly do not—but because patients often discuss symptoms online in large numbers, quickly and honestly, about issues they might not raise in a brief appointment.
We saw a similar phenomenon with COVID-19, where Google search trends for "loss of smell" preceded official case reports, leading researchers to argue that such data is useful for monitoring disease spread. No doctor, regulator, or scientist could read that volume of social media conversation and detect patterns manually. AI can.
For me, this is where the importance lies. AI has received much negative press, but it has been revolutionary in medicine. For instance, it can help develop new drugs by searching vast amounts of scientific data for potential treatments. For families living with rare diseases—around three million in the UK—the advances AI brings are more significant than most people realise. There are no textbooks and few, if any, studies for these conditions; sometimes patients know more than professionals simply because they live with the disease daily. They are on social media, comparing symptoms, setbacks, discussing treatments, and noticing patterns that no formal study has captured. AI can do that.
Beyond analysing patient data, AI can assist doctors in real-time. It is already improving the interpretation of scans and X-rays, enabling better cancer diagnoses through more accurate analysis of pathology slides and blood tests. One of my greatest concerns in the emergency department is not the obvious disaster but the subtle one: a patient with a broken hip after a fall whose X-ray is deemed normal, leading to discharge, or worse, a patient with a sudden headache whose brain scan contains a tiny bleed from a haemorrhage that is missed, with devastating consequences. We like to think doctors get these things right every time, but we do not.
That is why the new wave of AI supporting diagnosis in emergency medicine is promising. A large review in the Annals of Medicine compiled 26 studies and found that when clinicians used AI to interpret bone X-rays, they became markedly better at both spotting genuine fractures and correctly identifying normal X-rays. It is not just broken bones. A study this year by Oxford University Hospital NHS Trust on CT brain scans found that AI helped emergency clinicians detect more critical signs, such as small bleeds, that are easy to miss but potentially devastating. Most strikingly, emergency clinicians using AI performed similarly to specialist radiologists in identifying serious problems on scans.
This matters because in real life, we often wait for specialist radiology reports before making discharge decisions based on normal CT scans. With AI, we may not need to. It is increasingly clear that AI can help us understand treatment risks faster, sifting through vast amounts of information and data quickly, and spotting patterns and harms earlier than might otherwise take years to emerge.
Traditionally, medicine has followed a set path: first animal studies, then clinical trials, and once a drug is on the market, reliance on doctors and patients reporting suspected side effects through systems like the Yellow Card scheme run by the Medicines and Healthcare products Regulatory Agency. These systems are life-saving but imperfect, as they depend on someone spotting a possible link and actually reporting it. In real life, this does not always happen, especially when side effects are vague, embarrassing, or easily dismissed. It is estimated that only about 10% of serious reactions and 2–4% of non-serious ones are reported to the Yellow Card scheme.
I do not believe AI will replace doctors—the human touch is essential—but it has the potential to work at both ends of medicine simultaneously: helping us understand trends across millions of patients while also enabling better decisions for one frightened patient in one room on one day. Every powerful medical advance carries risks, and every tool can be misused. But that has never been a reason to reject progress. The teacher I treated is a reminder of why we need systems that detect patterns earlier and feed that knowledge back into patient care before more people are harmed. AI can help with that.
