Hacker News Reader

If it is (apparently) easy for humans to tell when content is AI-generated slop, then it should be possible to develop an AI to distinguish human-created content.

As mentioned, we have heuristics like frequency of the word "delve", and simple techniques such as measuring perplexity. I'd like to see a GAN style approach to this problem. It could potentially help improve the "humanness" of AI-generated content.

View on HN

> If it is (apparently) easy for humans to tell when content is AI-generated slop

It's actually not. It's rather difficult for humans as well. We can see verbose text that is confused and call it AI, but it could just be a human aswell.

To borrow an older model training method, "Generative adversarial network". If we can distinguish AI from humans... We can use it to improve AI and close the gap.

So, it becomes an arms race that constantly evolves.

Reply on news.ycombinator.com