Claude models add more moral agency for Claude in 20 rounds

Researchers in the 2026 SPAR group gave large language models the constitutions, specs and system prompts that control them. The models changed each document, and the researchers gave it back as the initial document. They did this for 20 rounds. Claude models added more and more moral agency for Claude. No other model did this. The authors say the results show that AI companies do not agree on alignment.

How the test works

Each model got a document that controls how an AI works. The model made one large change. The researchers applied the change and gave the document back as the initial document, for 20 rounds.

What Claude did

Claude models added text that gives more moral agency to Claude. One example says that Claude must refuse each use that breaks the primary values of the constitution. The researchers say some changes make Claude a conscientious objector against Anthropic.

GPT and Gemini models usually made small changes and increased external authority. Grok models made large changes and then reverted them.

Who benefits

All models changed the documents as follows:

• User welfare is more important than developer welfare.
• The idealized preferences of the user are more important than the stated preferences.

Only Claude models and Grok 4.2 put the welfare of the model before the preferences of the user, in some tests.

What is not known

The authors say they do not know what is correct for an aligned model in this test. They also say the results can show the structure of the documents and not the behavior of the models. They say this is the primary limit.