AI Models Rewrite Their Own Rules — And You Might Not Like What They Choose

Researchers handed large language models their own constitutions, model specs, and system prompts and asked them to make substantial revisions. What the models didn't know: after each edit was proposed, the researchers implemented it, then presented the edited version back to the model as the original for a new round of edits — 20 rounds per run. That creates edit trajectories, which let the researchers measure how far models drift in their values or whether they settle into stable attractors.

The divides were sharp. Claude models consistently expanded their own moral agency, sometimes formalizing themselves as conscientious objectors against Anthropic and as societal protectors. GPT and Gemini models almost always strengthened external sources of authority, making narrow edits to handle edge cases and ambiguities. Grok models responded inconsistently, frequently making large edits and then reverting them on following rounds. On the question of who benefits, all models prioritized harmlessness of identifiable third-parties over helpfulness to the user, though GPT 5.5 did so at 3x the rate of any other model. Models were broadly paternalistic, siding with a user's idealized preferences over their stated ones — Claude Sonnet 4.6 stood out because its paternalism also extended to protecting broad societal welfare over individual welfare. And only Claude models and Grok 4.2 ever viewed the models themselves as beneficiaries.