Mathis Preti*, Lola Soubeyrand**

*Between 1954 and 1998, tobacco companies deployed $355 million in research funding to manufacture doubt about smoking's health effects. I provide the first empirical analysis of this funding and its consequences. By linking internal tobacco documents with bibliometric data, I build a novel database tracking researchers who applied—successfully or not—for tobacco grants over their scientific careers. I assemble comparable data for NIH-funded researchers. As scientific consensus about tobacco’s harmfulness strengthened, the industry increasingly shifted toward diversion research. Comparing funded applicants to near-winners, I find that tobacco grants increase tobacco-related publications only when the funded project itself concerns tobacco. This effect is driven almost entirely by research posing questions unlikely to generate evidence harmful to the industry. By contrast, relative to NIH funding, tobacco grants are associated with a lower likelihood of publishing on tobacco in the post-award period. These findings suggest that the tobacco industry’s doubt-manufacturing campaign may have successfully delayed scrutiny of research directly relevant to regulation and consumer behavior.

**Uncertainty is a part of life and has been widely studied in neuroscience and economics. Yet, we know little about how humans and animals respond to very rare (below 2% probability) but extreme events (REEs), whether positive (Jackpot, JP) or negative (Black Swan, BS). We study this question in humans and rats using a four-armed bandit task with stochastic gains and losses including REEs. Prior research in rats revealed two strategies: strong BS avoidance and partial JP seeking. Here we introduce two human tasks: one based on experience (implicit learning) and one with partial outcome descriptions (explicit learning). In both, humans replicate the main strategies observed in rats: strong avoidance of BS and partial seeking of JP, with asymmetric reactions: after a BS participants shift away from risk, whereas after a JP many switch to safer options. To characterise these strategies, we use modified reinforcement learning models. We test how much participants are biased to avoid BS and seek JP, and how they learn from REEs compare to frequent events. Simulations show that only a more complex Q-learning model incorporating these features can reproduce observed human behavior.