What Can You Do with a Sword? Gender Biases in Text Game Affordances

Game mechanics can be viewed in terms of affordances: possible actions offered by the environment—depending on the agent (e.g., in a fantasy role-playing game, a sword can be wielded by a knight, but probably not by a dragon). Recently, text generated by large language models (LLMs) has been used to create open-ended text-based game content. However, LLMs have been shown to generate sexist text when trained on gender-biased data. If bias manifests in educational text game affordances it could harm goal achievement. We examine binary gender biases in LIGHT, an English-language persona-based dataset for researching language grounded in a fantasy adventure world, training LLMs on LIGHT and analyzing the diversity of affordances in quests. We find male characters have a more diverse space of affordances yet are less diverse in practice (e.g., mostly wielding a sword) in original and generated quests. To gauge impact on gameplay, we create games from LIGHT quests which can be played in the TextWorld research framework. Artificial agents trained only on male games significantly outperform female, suggesting an impact of affordance biases. These findings illustrate risks in AI- or data-driven generation of serious game content where gender is involved: overlooked biases in affordances can propagate, autonomously enforcing harmful, stereotypical behaviors.