When AI Deceives: A Natural Experiment on the Causal Effects of Perceived Deception on Player Ratings in RPGs

Abstract

AI-driven deception mechanisms are increasingly prevalent in digital games, yet the direction and magnitude of their effects on player experience remain contested. Existing research has not sufficiently disentangled designer-intended deception intensity from players' actual perception of deception, and most prior work relies on low-ecological-validity experiments or cross-sectional surveys. The present study aims to independently examine the causal effects of design deception intensity (DDI) and player deception awareness (PDA) on player ratings within a naturalistic gaming environment, and to investigate the moderating role of player experience. Leveraging the 54 version updates of Baldur's Gate 3 between 2019 and 2025 as a quasi-natural experiment, it collected all English-language Steam reviews posted within 1 to 28 days following each update, and constructed a player-version two-way fixed effects panel dataset. DDI was coded by human annotators based on patch notes; PDA was extracted and aggregated from review texts using a fine-tuned BERT classifier. The model incorporated both player and version fixed effects, complemented by five robustness checks including subsample partitioning, lagged variables, and placebo tests. PDA exerts a monotonic negative effect on positive review rates: within the observed PDA range, the net loss in review valence is approximately 0.4 percentage points, with a negative quadratic term that falsifies the inverted-U hypothesis of moderate perception optimality. DDI exhibits a U-shaped effect with an inflection point at a relatively low intensity, although the upward trend on the right branch is primarily driven by contemporaneous new content bundled with high-intensity updates. Any degree of deception awareness undermines player evaluations, while the positive manifestation of design intensity depends on content-confounding effects.