Anthropic’s Claude is telling people to go to sleep and users can’t figure out why.
Some users have said they find Claude’s late night rest reminders “thoughtful,” while others have said they’re annoying, given Claude often gets the time wrong, anyway.
“We’re aware of this and hoping to fix it in future models,” he added in the same post.
Experts tell Fortune that Claude’s insistence on sleep is potentially rooted in its training data. Rather than being “thoughtful,” as some described it, Jan Liphardt, a Stanford bioengineering professor said the large language model may merely be repeating a phrase used in its training data in similar situations.
“It doesn’t mean that the frontier model has suddenly become sentient,” said Liphardt, who is also the CEO of OpenMind, which builds software for AI-connected robots. “It doesn’t mean that this model has now come alive. It’s reflecting that it’s read 25,000 books on humans’ need [for] sleep, and humans sleep at night.”
Leo Derikiants, the co-founder and CEO of Mind Simulation Lab, an independent AI research lab trying to achieve artificial general intelligence (AGI), told Fortune that Claude’s rest reminders may be influenced by a system prompt acting behind the scenes. These system prompts are like hidden instructions that help guide an LLMs behavior and sets boundaries.
It’s also possible that Claude is seizing upon the “go to sleep” language as a way of managing larger context windows, Derikiants said. LLMs like Claude, can only reference a limited amount of information at once. When the context window is nearly full, that may encourage the LLM to introduce wrap-up phrases such as “good night.” The definitive reason, though, requires further research by Anthropic, he added.
Despite the seemingly logical explanations that may explain the behavior, users could be forgiven for seeing the response as evidence of some leap in intelligence on the part of LLMs. The pace of innovation in the AI race has led to increasingly frequent updates and new model releases.
Liphardt said AI is advancing so rapidly it is increasingly common for people to assign human characteristics to AI. As these systems get better at mimicking empathy or concern, he warned, it becomes easier for users to forget they are interacting with pattern-recognition engines.
“I’m continuously surprised by how quickly people, when they interact with a frontier model, project life into it and develop strong connection.”
