Kelsey Piper:
Yes, I think that’s a great question, because the key thing here is that, four or five years ago, people were first coming out with the earliest language models using these generative techniques that could produce text.
And they weren’t very good. They were kind of stilted. They were maybe writing it like at a middle school level, not a college level. They would make up a lot of stuff, and they didn’t seem to have a good way of telling whether the things they were saying were true.
And over the course of just a couple of years, we have vastly improved the quality of these programs. And now they’re writing at a college level, and they’re clever, and they bluff and they make things up. But they can also be pretty accurate when they’re prompted to be pretty accurate.
So where is that taking us. In a few years, maybe we do have something that replaces my job as a journalist. Maybe we do have systems that can tell you what you want to hear in incredibly convincing length at any time. And then there’s the question of, who decides what those A.I. systems tell everybody, and how do we train those A.I. systems, such that they are on humanity’s side, helping us understand the world better, instead of — right now, I would argue they’re kind of trained to appease us.
They’re trained to say what they think we want to hear. And there’s something scary about the idea of introducing these immensely powerful, immensely persuasive systems to the world with a mandate that’s as limited as, just say what will make people happy with you.