ChatGPT’s Literary “Memory” Wording May Reflect Copyright Safeguards, but Cause Is Unclear

When asked to explain a difficult passage from a book, a model reportedly said it was recalling the book from memory and might not remember it accurately. The wording made the model sound like a self-aware entity with its own understanding, an effect described as unusual and unsettling.

One possible explanation is reinforcement learning aimed at preventing models from reproducing entire copyrighted books: the model may be trained to discuss literature like someone who has read it but does not remember every detail. This remains a hypothesis, and the original response could not be checked because it was made in incognito mode and disappeared from chat history.

The exchange also raised a terminology question for systems using agent swarms: instead of referring to themselves as “I,” such systems might eventually use “we.”