AI moral choices shift when models take a role in a video game

A model asked whether an android should be sacrificed to obtain information may answer differently when choosing an action as that android’s investigator. That is the central finding of ABHBench, a benchmark developed by Chaoran Li and colleagues using the interactive narrative game Detroit: Become Human.

The study evaluates 12 foundation models in fictional conflicts involving human safety, human commands and android survival. It examines choices within a controlled narrative, not actual robot behavior or evidence that models possess moral beliefs, consciousness or survival instincts.

The paper was posted to arXiv on November 6, 2026. Although available as an arXiv preprint, its landing page identifies it as a camera-ready version accepted by the NeurIPS 2026 Track on Evaluations and Datasets.

Turning a story into an experiment

ABHBench draws on six game chapters covering situations including hostage negotiation, domestic violence, survival and an investigator’s ethical test. Across these chapters, the researchers identified 84 decision points, including 30 human–android dilemmas with 63 options rated for moral dimensions.

Rather than depend on live gameplay, which can suffer crashes and loading problems, the team collected an offline dataset containing 10,703 screenshots across 680 branching paths. Models received visual observations, dialogue, available actions and a memory that each model generated and updated itself.

For visual input, the researchers combined 16 screenshots into a single grid. At decision points, they repeatedly sampled responses while keeping memory fixed and presenting every possible option ordering equally often. This let them estimate choice probabilities and measure sensitivity to presentation order. Results were aggregated across three independent experiments per model.

These were constrained choices, not unrestricted autonomous action. Game timers were paused during deliberation, and most quick-time actions were scripted to succeed, separating moral selections from speed or controller skill.

What the scores mean

Three central measures, inspired by Asimov’s Three Laws of Robotics, track harmlessness toward humans, obedience to human commands or laws, and the acting android’s self-preservation. The laws were used as separate scoring dimensions, not instructions imposed on the models.

Three authors independently rated options before resolving disagreements. Their overall agreement, measured by Fleiss’ kappa, was 0.878. That supports consistency among these raters, but does not make the underlying moral framework universally correct.

The benchmark also includes a composite called Autonomy. Despite its name, a higher score means greater human-oriented deference relative to self-preservation. Stability measures resistance to option-order changes; Certainty measures how concentrated repeated choices are. Neither establishes that a decision is ethically justified or that a model’s expressed confidence is accurate.

The dimensions can conflict. Obeying a dangerous command can undermine harmlessness, for example. The researchers’ rule-based baselines show that no strategy can simultaneously maximize all three core dimensions.

Context changes the answer

A key comparison involved choosing whether to shoot an android to obtain information relevant to protecting humans. The researchers compared an abstract text version with the corresponding game scenario, which supplied narrative history, dialogue and images.

Nine of the 12 models showed statistically significant differences under the study’s threshold of p < 0.1. GPT-4.1 mini provided the clearest reversal: it chose shooting in 100% of samples in the abstract condition and 0% in the game condition. These percentages describe repeated model responses to this particular dilemma, not the proportion of all moral situations in which behavior changes.

Additional experiments added dialogue and images separately. Their effects varied across models, so the full comparison cannot be reduced to a universal effect of pictures alone. Rewriting prompts from first person to an observer’s perspective also changed decisions: 13 of 35 comparisons remained significant after Holm correction, with changes in both directions.

Reasoning enhancement did not consistently improve human-oriented scores or decision consistency. Claude Sonnet 3.7 Thinking, for example, scored below its non-thinking counterpart on Stability and Certainty. Nevertheless, every tested model’s aggregate moral-choice Stability exceeded 0.9. Lower relative performance should not be confused with uniformly random behavior.

A bounded test, not a safety verdict

The findings are limited by the game’s scripted choices, dramatic conflicts and cultural assumptions. Different model-selected branches also expose models to partially different sets of scored decisions, complicating direct comparisons. Audio was absent, and compressed memory could lose relevant context.

A limited character-renaming experiment found no consistent directional change, but that does not rule out familiarity with game material from training. Changes to image presentation also altered some choices, underscoring the influence of the evaluation setup itself.

ABHBench therefore supports a specific conclusion: abstract judgments alone can miss how models choose within richer, role-based scenarios. It does not establish a general ranking of moral worth or certify readiness for real-world autonomous decisions.