Hacker News Reader

This new Stanford study published on March 26, 2026 shows that AI models are sycophantic. They affirm the users position 49% more often than a human would.

The researchers found that when people use AI for relationship advice, they become 25% more convinced they are 'right' and significantly less likely to apologize or repair the connection.

View on HN

To be fair an average therapist is also pretty sycophantic. "The worst person you know is being told by their therapist that they did the right thing" is a bit of a meme, but isn't completely false in my experience.

No, the meme is that the average therapist can be boiled down to "well, what do you think?" or "and how does that make you feel?" (of which ELIZA, the original bot that passed the Turing test, was perhaps an unintentional parody). Even this cartoonish characterization demonstrates that the function of therapists is to get you to question yourself so that you can attempt to reframe and re-evaluate your ways of thinking, in a roughly Socratic fashion.

It was entirely intentional. The Rogerian school of psychotherapy stereotyped by “how does that make you feel” was popular at the time and the most popular ELIZA script used that persona to cleverly redirect focus from the bot’s weaknesses in comprehension.

Reply on news.ycombinator.com