How Far Does the Trolley Problem Go in AI Ethics Evaluation? Limits of a Canonical Benchmark and the Risks of Its Misuse

This paper examines both the genuine utility and structural limitations of the trolley problem as an AI ethics evaluation instrument. While effective for surfacing value priorities under forced choice, strong trolley-problem performance is not a safety certificate and may produce false assurance. We identify three structural limitations and argue for evaluation frameworks that extend beyond forced-choice benchmarks.