Is AI Becoming More Human? Evidence from LLMs and the Ultimatum Game

We replicate and extend the ultimatum game experiment of Araujo and Uhlig (2026) to five consumer-facing large language models (LLMs)—ChatGPT, Claude, Copilot, Gemini Flash, and Gemini Pro—across four scenario types (HH,HA, AH, AA), four stake levels ($10 to $10,000), and ten repetitions per configuration, yielding 6,832 proposer and 8,532 responder observations. Four findings emerge. First, all five models propose shares in the 30–47% range, squarely within the human empirical benchmark and absent the extreme behavioral modes documented in earlier research-grade models, suggesting that alignment training has compressed the behavioral distribution toward human norms. Second, every model exhibits twosided sensitivity to human presence: proposed shares rise when the Responder is human (+4 to +25 p.p.) and minimum acceptable thresholds rise when acting on behalf of a human (+11 to +26 p.p.), a pattern that survives even when no human principal is being served and is inconsistent with simple principal–agent alignment. Third, all five models forgo 25–63% of feasible payoff, confirming that consumer LLMs are not payoff-maximizing agents. Fourth, responder thresholds decline significantly with stake size across all models— consistent with rational expected-utility behavior—while proposer stake sensitivity is heterogeneous. We interpret these patterns as evidence of identity internalization: successive rounds of reinforcement learning with human feedback cause models to behave as if they are human rather than merely as if they prefer human-like outcomes.