Do AI detectors work? Students face false cheating accusations

I'm surprised at the number of comments that give up and say that "AI" is here to stay.

I'm also surprised that academics rely on snake oil software to deal with the issue.

Instead, academics should unite and push for outlawing "AI" or make it difficult to get like cigarettes. Sometimes politicians still listen to academics.

It is probably not going to happen though since the level of political apathy among academics is unprecedented. Everyone is just following orders.

I can't think of a single time that we've ever willingly put down a technology that a single person could deploy and appear to be highly productive. You may as well try to ban fire.

Looking at some of the most successful historical pushbacks against technology, taxes and compensation for displaced workers is about as much as we can expect.

Even trying to put restrictions on AI is going to be very practically challenging. But I think the most basic of restrictions like mandating watermarks or tracing material of some kind in it might be possible and really that might do a lot to mitigate the worst problems.

the cat is irreversibly out of the bag now, unless you want to ban gaming-grade GPUs, macbooks, and anything with high bandwidth memory capable of massively parallel compute. you can't strip the knowledge of how to build an LLM from people's brains, even non-ML software engineers will know the general research direction of how to get back to at least a GPT-3 level.

this is also not a good era for politicians to listen to academics, anti elitism sentiment is at a high and nobody will vote for "eating their vegetables" vs. "candy for dinner".

> Instead, academics should unite and push for outlawing "AI"

Prohibition does not solve the problem of needing to detect violations of the prohibition.

> or make it difficult to get like cigarettes.

Cigarettes aren't, at all, difficult to get, they are just heavily taxed.

> But I think the most basic of restrictions like mandating watermarks or tracing material of some kind in it might be possible and really that might do a lot to mitigate the worst problems.

Watermarking output (anything that is detectable that is part of the structure of the text, visual--if even imperceptible--image, or otherwise integrated into whatever the primary output is) will make it take a bit more effort to conceal use, but people and tooling will adapt to it very quickly. Tracing material distinct from watermarking -- i.e., accompanying metadata that can be stripped without any impact to the text, image, or whatever else is the primary output -- will do the same, but be even easier to strip, and so have less impact.

And also, mandates for either are mainly going to effect use of public, hosted services; but the proliferation of increasingly-capable open models where fine-tuning and inference can be done locally on consumer hardware will continue and be an additional problem for anything that relies on such a mandate.