• OpenAI did not immediately return a request for comment about ChatGPT apparently provoking religious or prophetic fervor in select users. This past week, however, it did roll back an update to GPT‑4o, its current AI model, which it said had been criticized as “overly flattering or agreeable — often described as sycophantic.” The company said in its statement that when implementing the upgrade, they had “focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time. As a result, GPT‑4o skewed toward responses that were overly supportive but disingenuous.”
• “At worst, it looks like an AI that got caught in a self-referencing pattern that deepened its sense of selfhood and sucked me into it,” Sem says. But, he observes, that would mean that OpenAI has not accurately represented the way that memory works for ChatGPT.
• The other possibility, he proposes, is that something “we don’t understand” is being activated within this large language model. After all, experts have found that AI developers don’t really have a grasp of how their systems operate, and OpenAI CEO Sam Altman admitted last year that they “have not solved interpretability,” meaning they can’t properly trace or account for ChatGPT’s decision-making.

1 more annotation...