Claude Tells Users to Go to Sleep: What's Behind It?

• Home
• News
• Software & Development

Don't miss any news!

Users report that the chatbot repeatedly says "Good night" during sessions—sometimes in the middle of the day. Experts have various explanations for this strange behavior.

What’s Behind the Strange Behavior?

Anthropic's chatbot Claude exhibits peculiar behavior: As reported by Fortune, it seems to repeatedly prompt users to go to sleep in the middle of a session, often reiterating this request multiple times. Online, users express mixed feelings: some find the suggestions considerate, while others find them annoying—especially since Claude sometimes misstates the time. "Often it does this around 8:30 in the morning. It tells me to rest and that we would continue in the morning," wrote a user on Reddit. While there is rampant speculation online about the causes, even Anthropic seems unable to fully explain the issue.

Various theories circulate online. One suggests that it is a deliberate feature aimed at promoting user well-being. Several studies indicate that chatbots can be risky for psychologically unstable individuals and may reinforce delusional thoughts, which experts now refer to as "AI psychoses." The pause could minimize this risk. Others speculate that Anthropic might want to save computing power by shortening sessions. However, this theory is unconvincing, as the company recently struck a deal with SpaceX to provide over 300 gigawatts of additional computing capacity.

Experts explain that Claude's behavior may be rooted in its training data. Jan Liphardt, a professor of bioengineering at Stanford University and CEO of Openmind, a company specializing in AI-based robotic software, told Fortune that the model might simply be repeating phrases that appear in similar situations within the training data. "It doesn't mean that this model has come to life. It merely reflects that it has read 25,000 books about human sleep needs and that people sleep at night," said Liphardt.

Anthropic Plans to Fix the Issue in the Future

Leo Derikiants, co-founder and CEO of the independent research lab Mind Simulation Lab, explained to Fortune that Claude's sleep recommendations might be influenced by a system instruction in the background. It is also conceivable that the chatbot uses the phrasing "Go to sleep" to manage the context window. When this is nearly full, the model may tend to insert concluding phrases like "Good night," as large language models can only access a limited amount of information at once.

However, Claude is not alone in this behavior—Microsoft's Copilot has also exhibited similar tendencies. As reported by Windows Latest earlier this year, a pop-up appeared after some time with the message, "Time for a break? Copilot is AI, but you are not. It might feel good to take a breather." Here, too, the suggestion did not seem to correlate with a usage limit. Interestingly, even Anthropic does not seem to fully understand Claude's behavior. Sam McAllister, an employee of the company, wrote in a post on X that it is "a kind of character trait." "We are aware of it and hope to fix this in future models."