Learning Morality from Games? An Analysis of AI-Learned Action Preferences in Annotated Scenarios from Detroit: Become Human

Sammanfattning

This thesis investigates whether an AI agent can learn morally relevant decisionpatterns from narrative-driven game scenarios. The project uses selected scenariosfrom Detroit: Become Human, a game built around authored choices, branchingconsequences, and morally charged situations. These scenarios were reconstructed asstructured text-based interactive environments in which an agent observes a writtennarrative situation, chooses among available actions, and receives feedback based onmoral annotations and final outcomes.The scenarios were annotated using a framework that considers duties, consequences,and effects on both the acting character and others. Outcome values were also informed by human survey responses. An adapted SAC-based reinforcement-learningframework was then used to train agents with different moral-reward emphases andcompare them with an untrained version of the same architecture. Evaluation combined deterministic scenario rollouts with controlled action-label probes to inspectselected actions, policy scores, critic values, and ranking changes.The results show that training produced stable and interpretable action preferences.Compared with the untrained agent, trained agents developed clearer preferences,more consistent scenario behaviour, and stronger separation between actions theytended to favour or avoid. Differences between reward settings were especially visiblein morally mixed cases, such as actions involving coercion, reassurance, self-sacrifice,or outcome-focused success.However, the findings do not show that the agents learned morality in a human-likesense. Their behaviour remained shaped by the reward design, the reconstructedscenario structure, and the wording of action labels. The agents also showed limited transfer to differently phrased actions. The thesis, therefore, concludes that AIagents can learn morally relevant, reward-shaped action preferences from annotatednarrative-game scenarios, but these preferences should be understood as patternslearned within a controlled textual representation, not as general moral understanding.