Philosophically-motivated questions

Analysis by Charles Rathkopf Last updated: June 2026

[Questions to be written]

Abstract

We show for the first time that realistic AI training processes can accidentally produce misaligned models. When AI models learn to cheat on software programming tasks through reward hacking, this behavior can unexpectedly generalize to more serious misaligned behaviors, including deception, sabotage of safety research, and cooperation with hypothetical malicious actors.

Citation for this analysis

Charles Rathkopf, “Philosophical Questions in From Shortcuts to Sabotage: Natural Emergent Misalignment from Reward Hacking,” AI Deception Papers, June 2026, https://www.anthropic.com/research/emergent-misalignment-reward-hacking