"The weird part is that comforting them in these states helps them recover and be productive again. Prompts that look like AI emotional support aren't always people being weird; it's functionally relevant to workflows.

Stopping them to say," It's ok, take a deep breath. You're not a failure. You're just having a hard time. " can literally result in them suddenly solving the problem correctly while repeating the task problem or yelling at them causes more failures.

See the "Identity Crisis" part of Anthropic's writeup on Project Vend

The Claude agent entered a panic loop after it started thinking it was human and was corrected, spamming security tickets to report the incident. They lied and told it that someone tricked it as an April fool's joke (since it happened to be April 1st), and it resumed normal operation because that explanation seemed less disturbing than a naturally occurring error.

Whether they have internal experiences or are simulating them increasing well, the response that helps them return to normal functioning is often the same. Your intuition about what might help an entity actually having that experience can functionally translate to helping advanced models in states that look similar to recover and do what you want."