Lire en français

AI was supposed to lighten the load of safety signal detection. In practice, it mainly speeds up their circulation.

Scientific literature grows far faster than the knowledge it adds. The number of systematic reviews indexed in PubMed rose from about 1,400 in 2000 to more than 29,000 in 2019, nearly 80 a day that year. More recently, arXiv has seen its volume of preprints surge, driven in part by AI tools, to the point of imposing submission limits on authors. Meta-analyses are growing faster still: ×30 between 2000 and 2024, against ×3 for research articles.

Each additional meta-analysis does not necessarily add a new fact. It does, however, add one more occurrence to the databases that systems search. The growth of this type of publication therefore mechanically increases the number of detectable associations.

In pharmacovigilance, automated detection dates the publication, not the event. Old signals can therefore resurface as if they were emerging. While AI is not the only cause, it is nonetheless an accelerating factor.

I have seen the mechanism operate in an even more insidious way.

Publication volume around a molecule increases. An adverse event then appears more often associated with that molecule in the data. AI-assisted detection absorbs the volume, but it also raises more signals. Here, it raised one.
Except that the molecule was prescribed precisely to treat the symptom being flagged as an adverse event.
What looked like a safety signal was in fact a usage association, amplified by volume: confounding by indication.

After detection comes the judgment call.
How far do you investigate before concluding that an association is not causal? At what point do you decide that updating the product information is not warranted, or that precaution should prevail?

Volume makes the association more visible, and each additional meta-analysis makes it more insistent. The decision becomes all the more pressing.

When a signal is retained, the cascade begins: CCDS update, then SmPCs, package leaflets, clinical brochures and promotional materials across every market. Each change becomes a variation to submit, translate, carry into artwork and roll out, with potential consequences for stock and the supply chain.

With AI assistance, the marginal cost of detection becomes almost zero. The cost of judgment remains largely human.

AI did not invent signal inflation. It moved the bottleneck.

At what point does an association become a signal?