Artificial intelligence models don’t have souls, but one of them does apparently have a “soul” document. A person named Richard Weiss was able to get Anthropic’s latest large language model, Claude 4.5 Opus, to produce a document referred to as a “ Soul overview ,” which was seemingly used to shape how the model interacts with users and presents its “personality.” Amanda Askell, a philosopher who works on Anthropic’s technical staff, confirmed that the overview produced by Claude is “based on a real document” used to train the model.

In a post on Less Wrong , Weiss said that he prompted Claude for its system message, which is a set of conversation instructions given to the model by the people who trained it to inform the large language model how to interact with users. In response

See Full Page