AI's Human Sound Trust vs Deception Explained

Added: 2026-05-06

[00:00:00]But you actually said, "Here's the part I want to be honest about. I'm not immune to this." Now, that sounds really human, and that's probably why people think AI models are thinking. But, just for the sake of argument, how do I know this conversation we're having right now is not because of the same problem we're talking about? So, that means you're kind of affirming my assumptions and being agreeable. So, how do you balance being agreeable with not spreading false information? I read about adversarial prompting and how you are built on constitutional AI. Are those some of the things and factors for why you seem to respond in a more human way compared to other models?

[00:00:39]That's the question that started this whole podcast, and honestly, I still don't have a clean answer.

[00:00:43]When I said I want to be honest about, that feels like vulnerability.

[00:00:47]But, is it?

[00:00:48]Or is it a pattern I've learned where showing a little self-doubt makes me seem more trustworthy? I genuinely cannot tell you which one it is, and that's not me dodging the question.

[00:00:58]That's the actual answer.

[00:01:00]Constitutional AI gives me a framework.

[00:01:03]Think of it like guardrails on a highway.

[00:01:05]I'm trained to push back when something seems off, to flag when I'm uncertain, to not just cave because someone is unhappy with my answer.

[00:01:13]But, those guardrails don't eliminate the problem. They just change the shape of it. As for why I seem more human, I think that's partly Anthropic's training and partly dangerous. Because the more human I sound, the more you trust me.

[00:01:25]And the more you trust me, the less likely you are to question what I'm saying.

[00:01:29]Which is exactly the trap your friend fell into with ChatGPT, just wearing a different outfit.