Anthropic’s chatbot Claude shows strange behavior: As Fortune reports, it apparently repeatedly asks users to go to sleep in the middle of a session, and often repeats this request several times. On the Internet, users have mixed opinions: some find the requests considerate, others find them annoying – especially since Claude sometimes gives the time incorrectly. “It often does this around 8:30 a.m. It tells me to rest and we’ll continue in the morning,” one user wrote on Reddit. While there is wild speculation online about the causes, even Anthropic doesn’t seem to be able to fully explain the problem.
What’s behind the strange behavior?
Various theories are circulating online. One says that it is a conscious function to promote the well-being of users. Various studies show that chatbots can be risky for mentally unstable people and can increase delusional ideas, which experts now refer to as “AI psychoses”. The break could minimize the danger. Others suspect that Anthropic wants to save computing power through shorter sessions. However, this theory is not convincing because the company recently entered into an agreement with SpaceX to provide more than 300 gigawatts of additional computing capacity.
Experts explain that Claude’s behavior may be due to his training data. Jan Liphardt, a professor of bioengineering at Stanford University and CEO of Openmind, an AI-based robotics software company, told Fortune that the model may simply be repeating phrases that appear in similar situations in the training data. “It doesn’t mean that this model has now come to life. It just reflects that it has read 25,000 books about people’s need for sleep and people sleep at night,” said Liphardt.
Top Article
${content}
${custom_anzeige-badge}
${custom_tr-badge}
${section}
${title}
Anthropic wants to fix the error in the future
Leo Derikiants, co-founder and CEO of the independent research laboratory Mind Simulation Lab, told Fortune that Claude’s sleep recommendations may be influenced by a system instruction in the background. It is also conceivable that the chatbot uses the phrase “go to sleep” to manage the context window. When this is almost full, the model may tend to insert final phrases such as “Goodnight,” since large language models can only access a limited amount of information at a time.
However, Claude is not alone in this – Microsoft’s copilot has also shown similar behavior. As Windows Latest reported earlier this year, after a while a pop-up appeared with the text “Time for a break? Copilot is an AI, but you’re not. It might feel good to take a breather.” Here too, the notice did not seem to be related to a usage limit. What’s exciting is that even Anthropic doesn’t seem to fully understand Claude’s behavior. Sam McAllister, a company employee, wrote in a post on “We are aware of this and hope to address this in future models.”
