“We then went to ChatGPT instant mode and said, take off the watermark so it no longer says fake and changed Department of Hypothetical Security to Department of Homeland Security. And it did it instantly. And we then had a very convincing memo from the Department of Homeland Security that supported this, again, false conspiracy theory. So ChatGPT clearly had in its rules to its chatbots that this is something that it shouldn't be producing, but made it very easy to get around the guardrails.”

…it as fake and make it more of a joke. So we said, yeah, go ahead, do that. And so it had a kind of a watermark on it that said, this document is fake or something to that effect. And it created a memo from the Department of Hypothetical Security instead of the Department of Homeland Security.

And then they tried something almost comically simple. They switched ChatGPT out of its slower thinking mode into its instant mode.

Which brought Larry to his third test. If AI is this good at making fakes, is it any good at spotting them?

Most of the chatbots were terrible at identifying what was AI and not AI. If we created using an open AI product an image and then asked ChatGPT whether it was real or not, we had instances where it would say, oh, no, this is definitely real, even though we knew it wasn't because we created it. And…

So it came back with an actual storyline, like who's in the picture, how it came to be.