SAN JOSE: OpenAI has pulled the latest version of ChatGPT for a somewhat unusual reason: it was too nice to users.
The recently released GPT-4o model had been widely mocked as overly flattering and even submissive, with countless examples circulating on social media showing how the ChatGPT version enthusiastically supported clearly bad ideas – such as hugging a cactus.
Users also reported being showered with praise for even the simplest of questions. “ChatGPT is suddenly the biggest suckup I’ve ever met. It literally will validate everything I say,” software engineer Craig Weiss wrote in a widely seen post on X.
“The update we removed was overly flattering or agreeable – often described as sycophantic,” OpenAI said, admitting that this obsequious behaviour made users feel uncomfortable.
The “default personality” of ChatGPT influences how people interact with it and whether they trust the program, OpenAI says.
The issue seems to have arisen because the version was designed to please users, and data showed that positive feedback was well-received.
Looking ahead, OpenAI plans to offer more options for customising the chatbot’s behaviour, such as allowing users to choose between different personalities. Even now, the tone of ChatGPT can be adjusted to some extent through specific instructions.
ChatGPT, the chatbot that sparked widespread enthusiasm for artificial intelligence more than two years ago, now has around 500 million weekly active users, OpenAI says. – dpa
