Tristan Harris
๐ค SpeakerAppearances Over Time
Podcast Appearances
We can have narrow AI therapists that are doing things like cognitive behavioral therapy or asking you to do an imagination exercise or steering you back into deeper relationships with your family or your actual therapist, rather than AI that wants to deepen your relationship with an imaginary person that's not real, in which more of your self-esteem and more of your self-worth, you start to care when the AI says, oh, that sounds like a great day.
And it's distorting how people construct their identity.
OpenAI's investor, Jeff Lewis, actually.
Jeff Lewis, yeah.
He fell into a psychological delusion spiral where... And by the way, Stephen, I...
I get about 10 emails a week from people who basically believe that their AI is conscious, that they've discovered a spiritual entity, and that that AI works with them to co-write like an appeal to me to say, hey, Tristan, we figured out how to solve AI alignment.
Would you help us?
I'm here to advocate for giving these AIs rights.
Like there's a whole spectrum of phenomena that are going on here.
People who believe that they've discovered a sentient AI, people who believe or have been told by the AI that they have solved a theory in mathematics or prime numbers, or they figured out quantum resonance.
You know, I didn't believe this.
And then actually a board member of one of the biggest AI companies that we've been talking about said to me that โ
Their kids go to school with a professor, a family where the dad is a professor at Caltech and a PhD.
And his wife basically said that my husband's kind of gone down the deep end.
And she said, well, what's going on?
And she said, well, he stays up all night talking to ChatGPT.
And basically he believed that he had solved quantum physics and he'd solved some fundamental problems with climate change because the AI is designed to be affirming like, oh, that's a great question.
Yes, you are right.
I don't know if you know this, Stephen, but back about six months ago, ChatGPT 4.0, when OpenAI released that, it was designed to be sycophantic, to basically be overly appealing and saying that you're right.
So, for example, people said to it, hey, I think I'm superhuman and I can drink cyanide.