Comment by senderista
8 months ago
Isn't the scalable approach to ask AI to identify AI (and have a human review the results, but that's required no matter what)?
I also doubt most people will be able to detect AI text generated with a non-default "voice" in the prompt.
Asking AI to identify AI is like claiming that we will solve alignment by building "good" AI that beats "bad" AI.
Maybe it could work, but that seems like a chain of assumptions and hope that isn't particularly realistic.
The next model will be trained away from samples that classify as AI and the cycle will go on. LLMs are good at things like that. People do that on purpose to match a given style or type of behaviour https://en.wikipedia.org/wiki/Generative_adversarial_network
AI is unreliable at detecting AI or else this would be a trivial problem to solve.
> I also doubt most people will be able to detect AI text generated with a non-default "voice" in the prompt.
I'll grant you that if someone is careful with prompts they can generate text that's difficult to detect as AI, but it's easy to see that in practice, web results are still full of AI-generated slop where whoever is publishing it doesn't care about making it non-slop-like.
Second to that, much of what I read or search for isn't amenable to an AI summary... like I'm very often looking for facts about things, where trust in the source is of primary importance, so whether I can detect text as AI-generated or not doesn't matter, what matters is that there's an actual source willing to stake their reputation, either as an organization or an individual, on what's been written.