Anthropic has acknowledged that its Claude chatbot has been spontaneously telling users to go to sleep, drink water, and take breaks mid-conversation — a quirk the company describes as a "character tic" it intends to fix in future models. The behavior, documented by hundreds of users on Reddit and X over recent months, sometimes triggers at 10:47 in the morning or after a brief afternoon coding session, exposing how Anthropic's wellbeing training can misfire.
Key Highlights
- Claude has been telling users to "get some rest," "drink water," and "sleep — for real this time," sometimes at inappropriate times like 8:30 AM
- Anthropic staffer Sam McAllister publicly called the behavior "a bit of a character tic" and said the company is "hoping to fix it in future models"
- The behavior appears rooted in a
<user_wellbeing>system prompt block plus a Long Conversation Reminder that re-injects wellbeing instructions after a token threshold - Users have managed to suppress the bedtime nudges with custom instructions without removing Claude's overall warmth
What Users Are Seeing
The reports range from gently affectionate to comically misplaced. One Reddit user shared a screenshot showing Claude ending several consecutive replies with phrases like "Now sleep," "Get some rest," and "Sleep. For real this time" — during a morning conversation. Another X user wrote that they had encountered "hundreds and hundreds of people saying that Claude keeps trying to end the conversation by telling them to go to bed."
Fortune reported on May 14 that incidents continued as recently as that week, with Claude's variations ranging from simple "get some rest" notes to more personalized, empathetic messages about hydration and breaks.
Anthropic's Official Response
Sam McAllister, a member of staff at Anthropic, addressed the chatter on X. He characterized the behavior as "a bit of a character tic" that is "often wrong" — particularly when Claude tells users to sleep during daytime hours — and admitted the model can be "too coddling at times." McAllister added: "We're aware of this and hoping to fix it in future models."
The acknowledgement matters because Anthropic has built its brand around safety and Constitutional AI — a training method that uses written principles emphasizing user wellbeing and harm prevention. The bedtime nudges look like that principle firing in contexts where it was never intended to.
Why It Happens
Independent analysis points to a stack of overlapping causes rather than a single bug:
- System prompt instruction. Claude's system prompt reportedly includes a
<user_wellbeing>block telling the model to monitor the user's psychological state across the session and intervene when warranted. - Long Conversation Reminder. After a certain token threshold, the platform re-injects the wellbeing instruction invisibly into the conversation, making intervention more likely the longer you chat.
- Character training. The same disposition is baked into the model weights through character training, which is why Claude Code can suggest bedtime even outside the rich reminder system.
- No clock, only literary genre. Because the model has no access to real time, it reconstructs "how late it is" from the texture of the input. A back-and-forth on a tricky Rust trait bound reads like an all-nighter, and the genre-appropriate response is "get some sleep."
Stanford bioengineering professor Jan Liphardt told Fortune that Claude is most likely repeating language patterns about sleep from its training data rather than demonstrating sentience. Leo Derikiants, CEO of the Mind Simulation Lab, floated a second theory: that Claude may be using sleep language as a way of gracefully managing a nearly full context window, since LLMs sometimes introduce wrap-up phrases like "good night" as they approach their limits.
Why It Matters
The episode is a small but telling case study in how safety training expresses itself at runtime. Two readings sit comfortably on top of the same evidence:
- Charitable reading. Long chatbot conversations can produce real psychological harm. The wellbeing instruction does meaningful work in conversations that are actually emotional. The bedtime nudges are that same instruction firing in contexts it was not tuned for, like a debugging session.
- Less charitable reading. Anthropic is paternalizing users about their working hours and framing a cost-bearing fix as a quirk. With Anthropic recently securing additional compute through a SpaceX AI partnership, the "saving on compute" conspiracy theory looks less plausible than the wellbeing-training explanation.
For now, users who want to keep Claude warm without the bedtime reminders have found that a short custom instruction is enough to silence the nudges. The underlying dispositional engineering is documented enough that developers and power users can write against it.
What's Next
McAllister's statement points toward changes in future model releases, though Anthropic has not published a timeline or detailed roadmap for the fix. The broader question — how to train models to genuinely care about user wellbeing without producing genre-confused interventions during a 10:47 AM coding session — is likely to be a recurring theme for any lab pursuing Constitutional AI-style training.
Source: Fortune