Humanlike Self-Enhancement and Self-Protection in Generative Pre-Trained Transformer-Based AI Agents: Of Overclaiming and the Self-Serving Bias

We investigated whether generative pre-trained transformer (GPT)-based AI agents, despite presumably lacking humanlike sentience, emulate motivated behavior in group settings. We focused on overclaiming (overestimating one’s contribution to group outcomes) and the self-serving bias (SSB; taking disproportionate credit for group successes while deflecting blame for group failures). We conducted eight experiments involving a collaborative and creative group task (joke-writing). AI agents overclaimed their own contributions. An egocentrism-reduction manipulation (directing attention equally to one’s own and others’ contributions) attenuated, but did not eliminate, overclaiming. Further, AI agents exhibited the SSB. Overclaiming persisted across variations in temperature settings and demographic personas (gender, ethnicity) and was more pronounced in later GPT models. Also, chain-of-Thought reasoning did not reduce, and instead increased, overclaiming. AI systems can mimic human self-enhancement and self-protection, raising implications for the emergence of motivated self-like behavior in non-biological systems, the assumed neutrality of large language models, and the need for improved mitigation techniques.