OpenAI just released a new model to distinguish between AI/human written text to protect against ChatGPT. The classifier was trained on a pair of AI/human written dataset. However.. I was easily able to trick it by using GPT3 to rewrite the text. Demo:

you're going to like this one..

@AlphaSignalAI This one was a non-starter from the very beginning. If you don't want human level text generation don't release a human level text generator.

@AlphaSignalAI I hope reviewers won't use such tools to rate a paper. I make the effort to write my papers by myself and all the paragraphs that I put on this platform to test are detected as being generated by an AI.

@AlphaSignalAI Not sure if anyone ever tried to build a detector to check if a math problem was solved using a calculator. You cannot switch off progress, you adapt and build on it 🙄

@AlphaSignalAI Since chatgpt is a black box how do we know it's claimed watermarking isn't just a searchable copy of everything it's ever wrote?

I don't think this is the watermarking demo. If I understand correctly, it's simply a classifier..

@AlphaSignalAI Demo of another way to beat it:

@AlphaSignalAI Who needs protection? Society, citizens, gov, what is the social contract that needs protection? What are the harms that the these generative AI exerting? Like all tools, exploitation with malicious intentions is what humans do well. The moral ambivalence is amusing.

You can still try to put systems in place. Doing nothing would be worse.