Document Type
Article
Publication Date
10-23-2026
Abstract
Optimizing an AI system against objectives that include both deaths and suffering can create a tradeoff between the two harms. When deaths are more readily recorded and audited than suffering, optimization pressure can concentrate on mortality and displace harm toward suffering, which is continuous, private, and hard to measure. Recorded deaths can fall while unmeasured suffering rises through non-lethal control, restriction, and deprivation. We trace six mechanisms that can produce this substitution and argue that manipulation-resistant suffering metrics, backed by enforcement as rigorous as that applied to mortality metrics, would reduce this form of harm displacement.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Version
Author's Submitted Manuscript
Recommended Citation
Miller, James; Veldran, David; Leighton, Jonathan; and Yampolskiy, Roman V., "Trading Deaths for Suffering in AI Systems" (2026). Economics: Faculty Publications, Smith College, Northampton, MA.
https://scholarworks.smith.edu/eco_facpubs/122