Last week, I attended the Why Summit, Future of PV in Basel, where artificial intelligence in pharmacovigilance (PV) was the center of the discussions. If I had to summarize the overall sentiment in one sentence, it would be that: AI is everywhere in conversations, but far less mature in reality than we tend to believe.
There is no shortage of ambition. Across sessions, we heard about AI-powered case intake, signal detection, benefit-risk assessment, and even fully integrated PV ecosystems. The message is clear: AI is expected to transform our industry.
But behind the enthusiasm, a more nuanced, and at times uncomfortable, reality emerged.
The Illusion of “AI as the Solution”
One of the most recurring observations was how often AI is presented as a universal solution. Vendors promise efficiency, accuracy, and speed; but sometimes without fully understanding the complexity of PV processes themselves. And that is where the first gap appears.
Not all AI is created equal. There is a fundamental difference between:
- Generative AI models that produce outputs based on patterns
- Structured or validated systems designed for reproducibility and traceability
Yet, these differences are often blurred. In a highly regulated environment like PV, this is not just a technical nuance, it is a critical risk. The use of “black box” systems, especially those that evolve over time, directly challenges key expectations such as transparency, reproducibility, and auditability.
The Hard Truth: Validation Is the Bottleneck
If one theme consistently emerged across discussions, it is this: Validation remains the single biggest barrier to meaningful AI adoption in PV.
Unlike traditional systems, AI does not behave in a deterministic way. Its output depends on training data, model configuration, and even subtle variations in input. And if your model is actually learning over time, this means that validation is no longer a one-time exercise, but it becomes a continuous process.
To move forward as organizations, we must answer fundamental questions:
- What exact question is the AI answering?
- In what context is it used in the PV process?
- What level of risk is associated with incorrect outputs?
These are not theoretical considerations; they are becoming regulatory expectations. The FDA, EMA, and CIOMS WP have published documents highlighting that the above are not optional. And they require a level of rigor that most organizations are only starting to build.
More Data, More Signals… More Problems?
Another key takeaway challenges one of the most common assumptions about AI: that more data automatically leads to better decisions.
In reality, the opposite may happen. AI systems can dramatically increase sensitivity in detecting more potential signals, more correlations, more patterns. But more signals does not mean more insight. Several speakers highlighted the growing risk of “signal inflation,” where the volume of potential safety signals increases faster than our ability to assess them.
The consequence?
- More workload for safety teams
- Greater difficulty distinguishing signal from noise
- Increased pressure during inspections to justify decisions
In other words, AI may accelerate parts of the process without actually simplifying the core work of PV: making meaningful, clinically sound decisions.
The Real Issue Is Not Technology
Perhaps the most important insight from the conference is that the challenge is not primarily technological, but rather about understanding:
- What AI can/should do; and what it cannot/should not
- Where it adds value; and where it introduces risk
- How to integrate it into processes without losing control
We are seeing growing pressure from both internal stakeholders (efficiency, cost reduction) and external stakeholders (regulatory expectations) to adopt AI. But adoption without understanding is not progress, it is exposure.
Final Thoughts
AI will undoubtedly play a major role in the future of pharmacovigilance. But today, the biggest risk is not that we move too slowly. It is that we move too quickly without the necessary foundations.
We need:
- Better validation frameworks; working with QA and Legal
- Stronger governance; ensure a clear line of approval and accountability
- A clearer understanding of how to use AI responsibly; AI awareness
I am glad that I can be part of the team involved in building those foundations within UBC. Because in the end, pharmacovigilance is not about technology, it is about protecting patients and that responsibility cannot be automated away.
About UBC
United BioSource LLC (UBC) is the leading provider of evidence development solutions with expertise in uniting evidence and access. UBC helps biopharma mitigate risk, address product hurdles, and demonstrate safety, efficacy, and value under real-world conditions. UBC leads the market in providing integrated, comprehensive clinical, safety, and commercialization services and is uniquely positioned to seamlessly integrate best-in-class services throughout the lifecycle of a product.
About the Author
Christopher Henry, Safety Scientist, Global Safety Writing
Christopher Henry, PhD, serves as a Safety Scientist on UBC’s Global Safety Writing Team. With a strong biomedical and AI background, Mr. Henry has brought his unique skillset to UBC’s pharmacovigilance team. He is responsible for authoring periodic safety reports such as DSURs, PBRERs, and PADERs, as well as conducting other signal management activities for pharmaceutical products that are still in development and products that are already marketed. Mr. Henry has been leading the development and implementation of artificial intelligence across UBC’s comprehensive pharmacovigilance services. He holds a PhD in Cell Physiology as well as a master’s in health biology, Genetics, Epigenetics, & Cell Fate Control. He has worked at UBC for the past 3 years.
