Beyond the Trolley Problem: Dostoevsky's Crime and Punishment as a Framework for AI Ethics
Year: 2026
Type: conference-paper
Abstract: Autonomous systems in transport, medicine, and defence increasingly select actions in situations where every available option harms someone, yet many decision architectures scalarise heterogeneous harms into one objective and return a confident optimum. This paper asks what design requirements follow when that confidence is itself a failure mode. Using a qualitative interdisciplinary method—hermeneutic reading of Fyodor Dostoevsky's Crime and Punishment, conceptual mapping to AI decision contexts, and targeted thematic synthesis of peer-reviewed AI ethics literature—we develop three design claims. First, consequentialist language can operate as post-hoc rationalisation; therefore, the deployed objective should be audited against its stated ethical justification. Second, some individual-level harms should be encoded as aggregation-invariant constraints rather than weighted terms. Third, responsibility gaps should be treated as design defects requiring logged and attributable human decision points. We consolidate these claims into a moral confidence threshold: an abstention rule, related to selective prediction and safe reinforcement learning, that blocks autonomous execution when the best available action exceeds a deliberatively set individual-harm ceiling or when competing actions are morally indistinguishable. The mechanism is a conceptual design proposal, not a validated result; a simulator-based validation protocol is specified. The framework's Christian-personalist basis is a stated limitation requiring cross-tradition testing.
Authors: Guzaliya Haydarova, Gulnoza Salimova
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7224851124