AI in pharmacovigilance means using machine learning and language models to process adverse event reports, remove duplicates and flag possible drug safety signals across millions of records. It genuinely speeds up detection. But deciding whether a drug actually caused harm still requires epidemiology, clinical judgment and accountable human reviewers, which AI has not replaced.
In July 2023, Europe’s medicines regulator began looking at a worrying set of case reports. Several people taking semaglutide and liraglutide, the GLP-1 drugs used for diabetes and weight loss, had reported suicidal thoughts or self-harm. The first reports came from Iceland. Within weeks the story was global. These were among the most prescribed and most discussed drugs in the world, and a psychiatric signal was exactly the kind of thing that could change prescribing overnight.
In April 2024, after reviewing clinical trials, post-marketing data, non-clinical studies and large health records analyses, the European Medicines Agency’s safety committee, PRAC, concluded that the available evidence did not support a causal association. No label change was needed.
A year later, the same committee looked at another signal in the same drug. Reports had linked semaglutide to non-arteritic anterior ischemic optic neuropathy, or NAION, a condition that damages the optic nerve and can cause vision loss. In June 2025, PRAC concluded that NAION is a very rare side effect of semaglutide, affecting up to 1 in 10,000 users, and recommended updating the product information.
Two signals. One drug class. Opposite conclusions. And in neither case was the hard part finding the signal. The hard part was deciding what it meant.
That is the thesis of this piece. AI is making pharmacovigilance much better at the part of the job that scales, which is reading, sorting and counting reports. It has done far less for the part that does not scale, which is judging causality. Anyone buying or regulating AI in drug safety should be clear about which problem they are solving.
What does pharmacovigilance actually involve?
Pharmacovigilance is the science of monitoring medicines after they reach patients. Clinical trials, even large ones, include too few people, for too short a time, to detect rare side effects. The rest is learned in the real world.
The raw material is enormous. The World Health Organization’s global database of individual case safety reports, VigiBase, run by the Uppsala Monitoring Centre, held over 40 million reports from more than 180 member countries and territories as of February 2025. National systems such as the FDA’s Adverse Event Reporting System and the EMA’s EudraVigilance add their own flows, and every pharmaceutical company runs a safety department processing reports for its products.
The work has three broad stages. Case processing: receiving reports, extracting the drugs, events and patient details, coding them in standard terminology, checking for duplicates and meeting strict regulatory deadlines. Signal detection: looking across reports for drug and event combinations that appear more often than expected. Signal assessment: deciding whether a detected signal reflects a real causal effect, and what to do about it.
Where does AI help most in drug safety?
At the first two stages, clearly.
Case processing is a high-volume, rules-heavy task that consumes a large share of safety department budgets. Reports arrive as free-text narratives, emails, call notes, scanned forms and literature articles. Natural language processing and, more recently, large language models can extract key information, suggest coding in the MedDRA medical terminology, and flag serious cases for priority review. Duplicate detection, a surprisingly hard problem when the same case is reported by a patient, a doctor and a company, has used statistical and machine learning methods at the Uppsala Monitoring Centre for years.
Signal detection has also long been statistical. Disproportionality methods compare how often an event is reported with a drug against how often it is reported with all other drugs. Machine learning adds ways to combine more information, such as report quality, clinical detail and time patterns, to rank which signals deserve a human’s attention first. Language models can read the narratives in reports, which older methods ignored.
These are real gains. Faster, cleaner case processing means signals emerge earlier. Better prioritization means scarce safety experts spend time on the signals most likely to matter. An international working group under CIOMS, the Council for International Organizations of Medical Sciences, has been developing principles for using AI in pharmacovigilance, reflecting how quickly the field is adopting these tools.
Why can’t AI decide whether a drug caused a side effect?
Because spontaneous reports cannot answer that question, however well they are processed.
A spontaneous report says that a person took a drug and something happened. It does not say how many people took the drug and nothing happened. It does not tell you how often the event occurs in similar people who never took the drug. And the chance that someone files a report depends on things that have nothing to do with causation: media coverage, lawsuits, how new the drug is, how alarming the event sounds and whether a clinician suspects a link.
The GLP-1 suicidality episode shows this clearly. Once news spread of a possible link, reporting was likely to rise, which would make the signal look stronger in spontaneous data regardless of the truth. People taking these drugs also often have obesity and diabetes, conditions associated with depression, so some suicidal thoughts would be expected in that population with or without the drug. Answering the causal question required different data: randomized trial data, and large health records studies comparing similar patients on different drugs. That is what PRAC relied on.
The NAION case needed the same work in reverse. A rare event, a plausible mechanism and a consistent pattern across several data sources eventually supported a real, if very rare, risk.
No classifier trained on report text can settle questions like these, because the information needed is not in the reports. AI can make the question arrive sooner. It cannot make the answer arrive without epidemiology.
A worked example: when a news story looks like a signal
Here is a hypothetical, with illustrative numbers.
A newer drug for a chronic condition has been on the market for three years. Its safety database shows 20 reports of hair loss per year, a level that raises no flags. Then a popular social media post describes hair loss while taking the drug. The next quarter, reports jump to 60, many from patients rather than clinicians, many referencing the post.
An AI-enhanced signal detection system does exactly what it should. It notes the sharp rise, the disproportionality against other drugs and the clustering in time, and ranks the combination highly. A language model reading the narratives might even observe that many reports mention the viral post.
Now the humans take over. Is hair loss more common in people who take the drug than in comparable people who do not? The spontaneous data cannot answer. The safety team looks at trial data: hair loss was reported at similar rates on drug and placebo. They commission a health records analysis comparing patients starting this drug with patients starting an alternative: no meaningful difference. They review the biological plausibility: weak.
The signal is closed as unconfirmed, with monitoring continuing. AI made that process faster, by surfacing and organizing the signal. The decision still depended on data sources and reasoning outside the reports. In a different version of the story, the trial data and the health records study both show an excess, and the label changes. The AI’s output would have looked the same in both versions.
The strongest case that AI will change causality assessment too
The counterargument deserves serious attention.
First, much of what humans do in case-level causality assessment is structured reasoning: did the event start after the drug, did it resolve when the drug stopped, did it return when the drug was restarted, are there better explanations? Standard frameworks, such as the WHO-UMC causality categories, are designed to be applied consistently. Language models reading full case narratives can apply such frameworks at scale, and possibly more consistently than tired human reviewers.
Second, AI does not have to rely only on spontaneous reports. Models can combine reports with health records, claims data, trial data and literature, which is exactly what human assessors do. The FDA’s Sentinel system already queries large health data networks to study safety questions. Adding machine learning to that kind of infrastructure could make population-level causal analysis much faster.
Both points are fair. Case-level assessment is ripe for assistance, and AI will make the pipeline from signal to epidemiological study faster. But neither point removes the need for judgment where evidence is ambiguous, which is where the consequential decisions sit. Whether to warn millions of patients about a 1 in 10,000 risk, or to reassure them about a feared one, is a decision regulators and companies must own and explain.
What this means for regulators, companies and clinicians
For pharmaceutical safety teams, the case for AI in case processing is strong, and the savings can be large. The discipline is to redeploy saved capacity to signal assessment and epidemiology, not simply cut headcount. Validate any AI tool against human coding, monitor it as report sources change, and keep clear accountability for regulatory submissions.
For regulators, AI tools create two obligations. Their own systems must be transparent enough to explain why a signal was flagged or dismissed. And they should expect companies using AI in safety processes to show how those tools were validated, how errors are caught and who is accountable. The broader principle that post-market evidence must lead to timely action, as we argued in fast approval isn’t the danger, slow withdrawal is, applies with more force when signals arrive faster.
For clinicians, reporting still matters, perhaps more than ever. AI can only find patterns in reports that exist, and clinician reports with good clinical detail are more useful than any algorithm can make a vague one. When a new safety concern about a popular drug makes headlines, as happened with GLP-1 drugs, the right response is usually to wait for the assessment, while reporting what you see. We discussed how widely these drugs are now used in the GLP-1 equity problem.
For patients, spontaneous reports and early signals are not proof of harm. Anyone worried about a side effect should speak with a clinician before stopping a medicine, and can report suspected side effects to their national regulator.
Frequently asked questions
How is AI used in pharmacovigilance?
AI is used mainly to process adverse event reports, extracting details from free text, coding events in standard terminology, detecting duplicate reports and flagging serious cases. It is also used to rank and prioritize potential safety signals in large databases. Increasingly, language models are being tested to read case narratives and support causality assessment, though human reviewers remain accountable for decisions.
What is a drug safety signal?
A safety signal is information suggesting a possible new link between a medicine and an adverse event, or a new aspect of a known link, that needs further investigation. A signal is not proof of harm. Many signals are later found to reflect chance, reporting patterns or underlying disease rather than an effect of the drug.
Did GLP-1 drugs cause suicidal thoughts?
The European Medicines Agency’s safety committee reviewed this question from July 2023 and concluded in April 2024 that the available evidence did not support a causal association between GLP-1 receptor agonists and suicidal thoughts or actions. Separately, in June 2025, it concluded that NAION, an optic nerve condition, is a very rare side effect of semaglutide. Patients with concerns should talk to their clinician.
Can AI replace pharmacovigilance professionals?
Not for the decisions that matter most. AI can take over much of the repetitive case processing work and help prioritize signals, but assessing causality requires combining evidence from trials, health records and biology, and weighing uncertainty against public health consequences. Regulators hold companies and people accountable for those judgments, which is why human oversight remains central.
What is VigiBase?
VigiBase is the WHO global database of reported potential side effects of medicines, maintained by the Uppsala Monitoring Centre in Sweden. It held more than 40 million reports from over 180 member countries and territories as of early 2025. National regulators and researchers use it to detect patterns that no single country’s data would reveal.
The takeaway
AI will make drug safety faster at noticing. That is valuable, because every month a real signal goes unnoticed is a month of preventable harm. But noticing was never the whole job. The GLP-1 reviews show that the defining work in pharmacovigilance is deciding which alarms are real, and that decision depends on evidence outside the reports and judgment that someone must be willing to sign.
Speed at the front of the pipeline is only useful if the back of it can keep up.
In your organization, has AI changed how safety signals are assessed, or only how fast they arrive? We’d like to hear from safety teams in the comments.
