Every morning, somewhere in the world, a drug safety associate opens a queue of new adverse event reports. Some arrive as tidy E2B files from a partner company. Others are scanned letters from physicians, call-centre notes, literature articles or social media posts flagged by a monitoring tool. Each has to be read, triaged, coded in MedDRA, assessed for seriousness and expectedness, and submitted to regulators within strict timelines.
It is careful, important work. It is also a volume problem that keeps growing, and that is exactly where artificial intelligence (AI) has entered the picture.
Why pharmacovigilance is under pressure
The World Health Organization (WHO) defines pharmacovigilance as the science and activities relating to the detection, assessment, understanding and prevention of adverse effects or any other medicine-related problem. WHO has also noted that adverse drug reactions are among the leading causes of death in many countries, and that a large share of them are preventable.
The data behind that mission is expanding rapidly. VigiBase, the WHO global database of individual case safety reports maintained by the Uppsala Monitoring Centre (UMC), contained more than 40 million reports from over 160 countries by the end of 2024, according to a resource profile published in Drug Safety in 2026. Around 70% of those reports were received in the most recent decade. The same analysis found that reports from consumers and non-healthcare professionals rose from 28% to 42% of the total, and that Asia's share of reports grew from 13% to 29%.
More reports, from more sources, in more languages, with more variable quality. For safety teams, the maths is unforgiving.
Where AI is being used today
AI in pharmacovigilance is not a single technology. It covers a range of tools, from rule-based automation and classical machine learning to natural language processing (NLP) and, more recently, large language models (LLMs). The main applications are:
Case intake and processing. NLP and LLM-based tools can extract key information from unstructured sources, such as the patient, the suspect product, the adverse event and the reporter, and populate a draft case for human review. Some tools also suggest MedDRA coding.
Duplicate detection. The same event is often reported more than once, by a patient, a physician and a company, for example. Machine learning methods help identify likely duplicates, which is essential for accurate signal detection. UMC has used probabilistic record matching in VigiBase for many years.
Literature screening. Classifiers can prioritise published articles that are likely to contain reportable safety information, reducing the time spent on irrelevant hits.
Signal detection and prioritisation. Disproportionality analysis has been a statistical workhorse for decades. Newer approaches combine it with additional features to help prioritise which potential signals deserve expert review first.
Information synthesis. Generative AI can draft summaries of case series or sections of aggregate reports, which a safety physician then checks and edits.
Regulators are modernising too. The FDA has announced it is consolidating its adverse event reporting systems, formerly known as FAERS, into a new Adverse Event Monitoring System (AEMS), which it describes as including AI-based redaction and digitisation tools and enhanced analytics.
The CIOMS Working Group XIV principles
Until recently, the industry lacked a shared reference point for using AI responsibly in pharmacovigilance. The Council for International Organizations of Medical Sciences (CIOMS), whose earlier working groups shaped much of modern drug safety practice, set out to fill that gap.
CIOMS Working Group XIV brought together regulators, industry and academic experts, met between 2022 and 2025, released a draft for public consultation in May 2025, and published its final report in December 2025. A summary of its guiding principles appeared in Drug Safety in August 2026.
The report sets out seven guiding principles:
- Risk-based approach: the level of validation and oversight should reflect the potential impact of the AI system on patient safety and regulatory decisions.
- Human oversight: people remain accountable for pharmacovigilance decisions, with oversight models chosen to fit the risk.
- Validity and robustness: AI systems must be shown to perform reliably for their intended use, and continue to do so over time.
- Transparency: users and regulators should understand what a system does, its limitations, and where AI has been used.
- Data privacy: personal health data must be protected throughout development and use.
- Fairness and equity: systems should be checked for performance differences across populations, languages and data sources.
- Governance and accountability: organisations need clear structures, documentation and responsibility for AI systems across their life cycle.
Human in the loop, on the loop, or in command?
One of the most practical parts of the CIOMS framework is its description of human oversight models:
- Human in the loop: a person reviews and approves each output before it is used. This suits higher-risk tasks such as causality assessment.
- Human on the loop: the system operates with people monitoring performance and able to intervene. This may suit lower-risk, high-volume tasks with strong quality controls.
- Human in command: people keep overall authority over whether, when and how the system is used.
The choice is not about trusting or distrusting AI in general. It is about matching the oversight to the consequence of an error in a specific task.
The risks we should take seriously
Hallucination and omission. LLMs can produce fluent text that is wrong, or leave out important details. In a safety narrative, a missing concomitant medication or an invented onset date is not a cosmetic problem.
Performance drift. A model trained on last year's reports may perform worse on new products, new reporting channels or new terminology. Ongoing performance monitoring is essential.
Bias. If training data over-represents certain regions, languages or reporter types, performance can be uneven. Given the global shifts in reporting patterns seen in VigiBase, this is a live concern.
Validation burden. Pharmacovigilance systems are subject to GVP requirements and inspection. AI components need to be validated in a way that is proportionate but defensible, and documentation must show how they were tested and how they are controlled.
Over-reliance. When a tool is right most of the time, reviewers can start to skim. Good process design keeps human reviewers engaged and accountable.
What this means for PV professionals
AI is changing pharmacovigilance roles, but it is not removing the need for scientific judgement. If anything, it is raising the bar.
Drug safety associates will spend less time on manual data entry and more on quality review, complex cases and follow-up. Knowing how to review AI-assisted output critically becomes a core skill.
Safety physicians and scientists remain responsible for medical assessment, causality and signal evaluation. Understanding how AI tools prioritise information helps them use those tools wisely.
PV quality and compliance specialists need to understand risk-based validation, performance monitoring and the documentation regulators will expect.
Everyone benefits from a working understanding of MedDRA, ICH E2B(R3), GVP modules and the basics of how machine learning models are trained and evaluated.
Key takeaways
- The volume and diversity of safety data are growing fast: VigiBase passed 40 million reports by the end of 2024.
- AI is already used in case intake, duplicate detection, literature screening, signal prioritisation and summarisation.
- CIOMS Working Group XIV published its final report in December 2025 with seven guiding principles, from a risk-based approach to governance and accountability.
- Human oversight should be matched to risk: in the loop, on the loop, or in command.
- The future PV professional combines drug safety expertise with the ability to evaluate and govern AI tools.
Building a career in modern drug safety
At CliniXen Institute, our programmes connect core clinical research and drug safety concepts with the data and technology skills the industry now expects. If you are aiming for a role in pharmacovigilance, a strong grounding in GCP, safety reporting and clinical data, combined with a practical understanding of AI, will set you apart.
Advanced PG Diploma in Clinical Research & Pharmacovigilance
References
- Council for International Organizations of Medical Sciences. Artificial Intelligence in Pharmacovigilance: Report of CIOMS Working Group XIV, December 2025. https://cioms.ch/working_groups/working-group-xiv-artificial-intelligence-in-pharmacovigilance/
- Norén GN, et al. Artificial Intelligence in Pharmacovigilance: Guiding Principles from the CIOMS Working Group XIV. Drug Safety, 2026. https://link.springer.com/article/10.1007/s40264-026-01708-z
- Norén GN, et al. VigiBase: Resource Profile Update with a Summary of Global Patterns and Trends in Adverse Event Reports for Medicines and Vaccines. Drug Safety, 2026;49(6):613-629. https://link.springer.com/article/10.1007/s40264-025-01642-6
- World Health Organization. Safety of medicines - adverse drug reactions: briefing note. https://www.who.int/docs/default-source/medicines/safety-of-medicines--adverse-drug-reactions-jun18.pdf
- World Health Organization. Pharmacovigilance. https://www.who.int/teams/regulation-prequalification/regulation-and-safety/pharmacovigilance
- US Food and Drug Administration. FDA Adverse Event Monitoring System (AEMS) [formerly FAERS]. https://www.fda.gov/drugs/surveillance/fdas-adverse-event-reporting-system-faers
Disclaimer: This article is for educational purposes and reflects publicly available regulatory and scientific information at the time of writing. Guidance documents are updated periodically; always consult the latest official version.
