The key points in 30 seconds ⚡
Researchers used AI to analyse around 400,000 messages posted online by people taking drugs from the GLP-1 analogue family, prescribed for diabetes and obesity. They identified symptoms reported unexpectedly, including changes to the menstrual cycle, chills, hot flushes and fatigue. The researchers themselves state that they cannot claim these treatments are the cause.
Researchers used AI to analyse around 400,000 messages posted online by people taking drugs from the GLP-1 analogue family, prescribed for diabetes and obesity. They identified symptoms reported unexpectedly, including changes to the menstrual cycle, chills, hot flushes and fatigue. The researchers themselves state that they cannot claim these treatments are the cause.
This morning we reported on clinicians' reservations about medical AI. Here is a different use, where the AI decides nothing but listens.
Why this type of study is useful
The clinical trials that precede a drug's authorisation have a known limitation: they cover a few thousand people, over a limited period, using strict selection criteria.
Rare, late-onset effects, or effects specific to certain populations, can go unnoticed. They only emerge once millions of people take the treatment in real-world conditions.
Existing reporting systems depend on what patients and doctors officially report, which remains a small fraction of what is experienced. Many people talk about their symptoms with other patients, online, far more than with their doctor.
Analysing what people write spontaneously gives access to that reservoir, and it is a task where AI contributes something real: reading four hundred thousand messages, spotting regularities, grouping together descriptions phrased in a thousand different ways.
What the researchers cannot conclude
The way this result is presented deserves credit, because it is exemplary. The researchers state explicitly that they cannot establish a causal link.
The reasons are numerous, and they apply to any study of this kind.
Nobody is verified. We do not know whether the person is actually taking the treatment, at what dose, or whether they have other health problems.
The population is biased. Those who write online do not resemble patients as a whole. Those who have unpleasant effects write more than those who do not.
There is no comparison group. Fatigue and hot flushes are common in the general population. Without knowing how many people not taking this treatment report the same symptoms, nothing can be concluded.
What this kind of study really produces 🔍
The right way to read this result is not "these treatments cause these symptoms". It is "here are signals that warrant rigorous study". This kind of analysis is a hypothesis generator, not an instrument of proof. It shows where to look, and that is already a great deal: without it, these signals would remain scattered across forums that nobody reads systematically. It is exactly the split we described with regard to mathematical research: the machine proposes, a rigorous method verifies.
The right way to read this result is not "these treatments cause these symptoms". It is "here are signals that warrant rigorous study". This kind of analysis is a hypothesis generator, not an instrument of proof. It shows where to look, and that is already a great deal: without it, these signals would remain scattered across forums that nobody reads systematically. It is exactly the split we described with regard to mathematical research: the machine proposes, a rigorous method verifies.
The question of confidentiality
One point deserves to be raised, even if it does not call the value of the study into question.
People who describe their symptoms on a forum to exchange with other patients did not consent to their posts being used for research. The data is public, but posting for one's peers and being analysed by researchers do not rest on the same expectation.
Health research strictly regulates the use of patient data. Health data published spontaneously online occupies a grey area that these methods suddenly make exploitable at scale.
What to take away
This use of AI in medicine is probably one of the most defensible: it decides nothing, it spots signals that existing systems do not capture, and it passes them on to methods capable of confirming or ruling them out.
If you are taking one of these treatments and you recognise one of these symptoms, the right response is not to conclude anything, but to talk to your doctor or pharmacist about it, and to report it through the official channels. That is precisely what will make it possible to turn a hypothesis into knowledge.