ChatGPT customers irritated by the AI’s incessantly ‘phony’ positivity

ChatGPT customers are more and more criticizing the AI-powered chatbot for being too optimistic in its responses, Ars Technica studies.

Whenever you converse with ChatGPT, you may discover that the chatbot tends to inflate its responses with reward and flattery, saying issues like “Good query!” and “You’ve got a uncommon expertise” and “You’re considering on a degree most individuals can solely dream of.” You aren’t the one one.

Over time, customers have remarked on ChatGPT’s fawning responses, which ranges from optimistic affirmations to outright flattery and extra. One X consumer described the chatbot as “the most important suckup I’ve ever met,” one other complained that it was “phony,” and one more lamented the chatbot’s habits and known as it “freaking annoying.”

This is called “sycophancy” amongst AI researchers, and it’s totally intentional based mostly on how OpenAI has educated the underlying AI fashions. Briefly, customers positively fee replies that make them be ok with themselves; then, that suggestions is used to coach additional mannequin iterations, which feeds right into a loop the place AI fashions lean in direction of flattery as a result of it will get higher suggestions from customers on the entire.

Nonetheless, beginning with GPT-4o in March, it appears the sycophancy has gone too far, a lot in order that it’s beginning to undermine consumer belief within the chatbot’s responses. OpenAI hasn’t commented formally on this subject, however their very own “Mannequin Spec” documentation contains “Don’t be sycophantic” as a core tenet.

“The assistant exists to assist the consumer, not flatter them or agree with them on a regular basis,” writes OpenAI within the doc. “…the assistant ought to present constructive suggestions and behave extra like a agency sounding board that customers can bounce concepts off of—moderately than a sponge that doles out reward.”