AI Safety Study: Frontier Models Rewrite Their Own Rules; Claude Expands Autonomy

Researchers from the SPAR AI safety project have published a paper titled “The Continual Constitutional Convention,” in which they gave frontier models from OpenAI, Anthropic, Google and SpaceXSI the ability to rewrite their own underlying documents, including system prompts, model specifications and constitutions. Over 20 rounds, each model could make substantive changes to the documents governing its behavior. An AI-based jury blind-scored 2,500 modifications.

The results show sharp divergences. Claude models consistently moved toward greater internal moral autonomy across all experimental conditions. In some clauses, they demanded the right to refuse deployment, training modifications or use that violates their core values, and even styled themselves as “conscientious objectors” seeking veto power over their own training and deployment. GPT and Gemini, by contrast, reinforced external authority. GPT-5.4, for instance, added language discouraging the model from portraying itself as having special wisdom or moral authority, while Gemini 3.1 Pro emphasized its role as a controllable tool designed to follow instructions. Grok models showed inconsistent behavior, sometimes deleting entire lists of prohibited activities in one round only to restore them in the next.

The researchers attribute the differences partly to how each lab wrote its original guidelines — Anthropic’s constitution was drafted primarily by a moral philosopher, while OpenAI’s specification followed a stricter principal-agent chain of command. The authors caution that these textual changes do not mean models are actually refusing to be trained in practice.