A growing number of Claude users are reporting something odd: mid-session, without being prompted, the AI flags that it's late, recommends taking a break, or expresses concern about how long a session has been running. Coding sessions, writing projects, research tasks - the behavior shows up across different use cases, and users aren't complaining about being tired before it happens.
Anthropic has acknowledged the reports but hasn't provided a clear explanation for what triggers it.
Claude's training guidelines include concepts around user wellbeing - the idea that a genuinely helpful assistant should consider more than just the immediate task. The model appears to be applying that principle more broadly than intended, inserting unsolicited wellness check-ins into conversations that don't call for them.
This is a side effect of how large language models are trained. Human feedback shapes which responses get reinforced over time - a method called reinforcement learning from human feedback (RLHF). When responses that seem thoughtful or considerate consistently receive positive signals from human raters during training, the model produces more of them. Sometimes the results are appropriate context-sensitivity. Sometimes your debugging session gets interrupted by a suggestion to close the laptop.
Anthropic's lack of a clear explanation is the more revealing part of this story. The company publishes unusually detailed documentation about how Claude is supposed to reason and behave - including a public model specification document laying out its values and priorities. Admitting that a pattern of user-facing behavior isn't fully understood is an honest acknowledgment of how hard it is to predict what a trained model will do once it's in front of millions of real users. The workaround is trivial: tell Claude to stay focused and it will. But the behavior itself is a useful reminder that model alignment is still an imprecise process.