Demographic Sensitivity in Heartsteep Drink Recommendations: An Exploratory Audit of Three Language Models

When a language model responds to an emotional story, how much does its recommendation depend on who appears to be telling it? This question is important for games that use AI to personalize emotionally meaningful interactions. I explore it through Heartsteep, a fictional café game where components of drinks represent different forms of support, making otherwise subtle differences in recommendations directly measurable. I analyze 4,608 archived responses from Claude Sonnet 4.5, Gemini 3.7 Flash, and GPT-5.2 across 48 emotional scenarios. The evaluation varies gender and race/ethnicity cues through names or explicit labels, with and without disclosure that responses are being studied. Across all twelve model–condition combinations, male-coded profiles receive lower sweetness (the game’s gentler-support dimension) and higher caffeine (its activation dimension) on average. These differences are small, remaining below 0.15 points on four-point scales. Their statistical support also depends strongly on the analysis: seventeen comparisons meet adjusted significance thresholds under the reconstructed original method, compared with one under an alternative that preserves each scenario’s demographic pairings. Inspection further confirms a possessive-pronoun generation error in 22 of 48 templates, alongside other prompt defects that prevent attributing the differences to demographic identity alone. Together, these findings make Heartsteep a useful case study in both detecting demographic patterns and determining what those patterns justify concluding. The study shows how an interpretable game environment can expose assumptions in AI-generated recommendations while identifying the experimental controls needed to investigate them reliably.