We then provide a theoretical impossibility result indicating that as language models become more sophisticated and better at emulating human text, the performance of even the best-possible detector decreases. For a sufficiently advanced language model seeking to imitate human text, even the best-possible detector may only perform marginally better than a random classifier. Our result is general enough to capture specific scenarios such as particular writing styles, clever prompt design, or text paraphrasing.
Interesting. Now, this is just one paper. And one paper does not mean the science is settled on that topic.
The implications are certainly interesting.
I’m curious how much data would be required to successfully mimic a specific writing style (e.g. lemmy post or research paper or letter to family) for a specific person. And conversely how easy it would be to detect.
I haven’t thought about this in depth yet. But the threats that come to mind are: someone spoofing me for some reason or me using AI to “research” and write for me (school, say) so I don’t actually have to learn anything. The former makes me wonder if digital signatures will become more widely adopted. The latter probably requires a different approach to assessing the knowledge of students. I’m sure there are other threats we can think of given a little more time.
It’s mathematically impossible to detect them reliably. But I know you will know it better.
https://www.techspot.com/news/98031-reliable-detection-ai-generated-text-impossible-new-study.html
Edit: here is the actual study: https://arxiv.org/abs/2303.11156
Here is the link in that article to The study
Regarding mathematical impossibility…
Interesting. Now, this is just one paper. And one paper does not mean the science is settled on that topic.
The implications are certainly interesting.
I’m curious how much data would be required to successfully mimic a specific writing style (e.g. lemmy post or research paper or letter to family) for a specific person. And conversely how easy it would be to detect.
I haven’t thought about this in depth yet. But the threats that come to mind are: someone spoofing me for some reason or me using AI to “research” and write for me (school, say) so I don’t actually have to learn anything. The former makes me wonder if digital signatures will become more widely adopted. The latter probably requires a different approach to assessing the knowledge of students. I’m sure there are other threats we can think of given a little more time.
This is like someone disputing an article about the Wright Brothers first flight with one from 6 months earlier that says manned flight can’t happen…