Comment by wongarsu | Hacker News Reader

Comment by wongarsu | Hacker News Reader

Parent Full thread

Microsoft is probably the best placed company to build a good detection method, having both a giant corpus of text predating LLMs from their Bing crawler, and a giant corpus of AI-generated text thanks to their relationship to OpenAI.

But Google is probably a close second, with an even bigger archive of pre-LLM text (including the world's largest collection of digitized books), and supposedly a good LLM and lots of smart ML engineers.

But no matter what method google finds, whether it's some complex AI trained on massive amounts of training data or a simple n-gram analysis that ends up being telling, they aren't going to tell us. That would just accelerate the arms-race between LLM spam and LLM detection. If they have something that works (big if) the details will stay hidden, the best we are going to get is some kind of press release aimed at dissuading people from using AI on their websites.

It is impossible to create detector for LLM generated writing. The LLM detector technology must beat the LLM at its own game and if it can we can easily create a better LLM. Anyone who claims they can do it is sorely mistaken or flat out lying.

By the same logic I guess you could create one for previous generations of LLMs. "Looks like you need to update buddy...busted!" And across multiple industries a decent percentage of users will be using outdated versions...

Your LLM detection method doesn't have to have a better understanding of human language than the LLM you're detecting, it just has to be able to identify patterns of speech typical of a particular LLM. Of course you have to repeat that work for each network and each significant update and finetune, but right now most people just use one of two ChatGPT versions.

And of course you don't tell people how your LLM detector works, don't provide a public API, and preferably leave it ambiguous whether it even exists. That solves most issues with LLMs using your detection method to get better in future versions.

That's counter to the business model of these LLM detection websites, but Google doesn't have a problem with keeping any potentially existing models inhouse and locked down.

Really all of it is bullshit if we don't know the true/false positive/negative rates.

If you detect LLMs 100% of the time correctly, it is irrelevant if you also incorrectly detect human written text as LLM text 80% of the time. "Detect them all and let god sort them out" will eventually fail as a business plan.

Reply on news.ycombinator.com