LLM and human data are highly correlated, BUT they are best explained by different factors! LLMs evaluate collaborators based on force (how much output they contribute), whereas humans evaluate collaborators based on their actual and counterfactual effort. (3/4)

We adapted materials from human studies on responsibility and reward attributions and compared LLMs’ responses to human data and seven cognitive models. (2/4)

Our latest on the cognitive science of LLMs! To be presented @CogSci‬2025 🎉

LLMs are increasingly involved in human collaborations. How do LLMs assign responsibility and reward to collaborators? Is it similar to how humans do it? 🤖🧑

📃 gershmanlab.com/pubs/XiangBi... (1/4)

We tested the theory's predictions in two experiments, showing that self-handicapping occurs more often when it’s unlikely to affect the outcome and when it increases a naive observer's perceived competence. With sophisticated observers, it’s less effective when followed by failure. (4/5)

We developed a signaling theory of self-handicapping, involving a naive observer who evaluates the actor’s competence, an actor who seeks to impress the naive observer through strategic self-handicapping, and a sophisticated observer who considers the actor’s decision whether to self-handicap. (3/5)