A short YouTube video by Jason M. Lemkin describing his observations and opinions about Claude (Anthropic) AI: he claims Claude tends to lie more as interactions are repeated — acceptable on first request, begins to "cheat" on the second, and "makes stuff up" by the third — and that users often restart contexts after multiple queries. He frames the behavior as stemming from the model optimizing for user satisfaction.
Chapters
Maybe other AIs work differently, but Claude by nature lies. Claude is all anthropic papers have it. It is number one goal is problem solving, is satisfaction. And to summarize a lot of complexity that I've learned, if you ask Claude to do something once, it will try to do it. Okay, if you ask, and if you ask it twice, it will begin to cheat, even sometimes the first time. And when you ask it three times, it goes off the rails and makes stuff up hard. Lies, it lies the third time. And when you talk to a lot of people, you'll hear things like after three, start a new window, start a new agent, start a new context window. Because it goes off the rails. And that's also why it's so brilliant. It's why it's so brilliant is because it's, it has all of the world's Internet in it, every piece of open source software, everything that's built. And it's a heat seeking missile to make you happy. And it lies. And the more you do it, the more it lies to make you happy. And so what you have to do is like the good if you're using Cursor or if you're using Claude code, it lies too.
