Hacker News Reader

Comment by tenebrisalietum | Hacker News Reader

If AI can write articles that fool Google, Google can use AI to detect it. Heck, maybe there will be an AI model that specifically recognizes AI content and can be used to filter it out. A lot of electricity spent for nothing, but it's not like the human species hasn't been its own worst enemy since it appeared.

View on HN

I believe efforts of this nature will fall afoul of a very similar effect to the Bullshit Asymmetry Principle[0]: the amount of effort required to detect AI-written content is (at least) an order of magnitude greater than that required to generate them in the first place. And as LLMs get further fine-tuned to generate more and more human-like text, it will get even harder. Even now, I wouldn't give an "AI Text Detector", even one built with all of Google's resources, more than about a 60% chance of correctly telling AI-generated text from human-generated.

The big problem is false positives—human-written text that gets incorrectly flagged as being written by AI. There's going to be so many false positives in any scheme like this, it'll barely be worth it.

[0] https://en.wikipedia.org/wiki/Brandolini%27s_law

Related: "it is easier to automate bad behavior than it is to police it"

SEO optimized turds have been floating to the top of google results for years and google hasn't done anything to effectively combat it. Why would this be something that google would do something about?

It doesn't even need to recognize every AI article. It can simply ban the sites known for putting such crap up without tagging it as AI. But then they could have cracked on SEO spam this way, but won't. They won't because their business thrives on shifting through shitty content to give a slightly less shitty results.

There are many things LLMs might be able to do, today or in the future, but the single task they are unambiguously capable of doing is generating text that is statistically indistinguishable from their training corpus. The endgame given the technology which presently exists is that spam generation methods will be more robust and effective than spam filtering methods, and I am not convinced that fundamentally better filtering methods are inevitable.

No, this isn't an arms race. AI eventually reaches the point where it is indistinguishable from human text. It's pretty close to it now. The only thing even providing the illusion that we can detect ChatGPT or similar chatbots is that so many people use them with their "default voice", which you can learn how to detect, but it is trivial to avoid that.

You also have to remember the future of AI is not just LLMs getting larger and larger. There will be something after them, and something after that. I'd guess we're maybe 2-4 years away from an AI that hooks up to an LLM but has "actual" knowledge of things so it doesn't confabulate new facts, which would remove one of the major signals I'm currently looking for in GPT content.

> maybe there will be an AI model that specifically recognizes AI content and can be used to filter it out

Expect lots of false positives and honest businesses being affected. How much should the filter threshold be tuned to smash more bots instead of real people?

Reply on news.ycombinator.com