Baldur’s Gate 3 study: AI deception backfires, cost 0.4% ratings

A paper from Dalian Jiaotong University examined how AI deception affects player ratings in Baldur’s Gate 3, using 54 patch versions as a natural experiment. The study distinguished between designer-intended deception intensity (DDI) and players’ actual perceived deception (PDA), the latter measured via a fine-tuned BERT model on English Steam reviews posted 1–28 days after each update. Key finding: PDA has a monotonic negative effect on positive review rates—within the observed range, the net loss is roughly 0.4 percentage points. The negative quadratic term rejects the hypothesis that moderate deception is optimal.

DDI, in contrast, shows a U-shaped effect with a low inflection point. However, the upward right branch is primarily driven by bundled new content rather than the deception itself. The authors conclude that any degree of perceived deception undermines player evaluations, while the positive effect of high design deception is a content-confound artifact. The study controlled for player and version fixed effects, with five robustness checks including subsample partitioning and placebo tests.