Scientific production in the era of large language models, by T. Renault, A. Bergeaud and C. Bosquet
Clément Bosquet (Professor at CES) has co-authored a new article in PNAS with Thomas Renault (Professor at the Université Paris-Saclay) and Antonin Bergeaud (Professor at HEC Paris): "Scientific production in the era of large language models: Outcome-triggered treatment timing and spurious event-study dynamics", published on 11 August 2026.
The article revisits a recent study published in Science. That study inferred LLM adoption from the first paper detected as LLM-assisted and reported large productivity gains after adoption. The authors show that this way of dating treatment mechanically produces positive event-study dynamics even when there is no causal effect. The reason is simple: months with high output are more likely to contain a detected paper, so treatment assignment is intrinsically tied to productivity.
Using arXiv data they rebuilt from scratch, the authors find similar dynamics with random treatment assignments, neutral-keyword triggers, inverted treatment and pre-ChatGPT placebo periods. Simulations with no treatment effect at all reproduce the post-treatment patterns of the original study. They conclude that event-study dynamics generated by first-detection timing cannot, on their own, be read as evidence that LLM adoption increases productivity.