News for hackers. Stuff that matters.
Re:Google Is Now Watermarking Its AI-Generated Text by kuhewa on Wednesday November 06, 2024 @06:53AM (#42060658)
Probably not entirely fair. e.g. After enough sentences it is trivially easy to identify LLM output. So you repeatedly get the opportunity to test a sentence or two, guess the provenance and then realise it is the first sentence in several paragraphs of generated output.
Re:Google Is Now Watermarking Its AI-Generated Text by exe34 on Wednesday November 06, 2024 @09:50AM (#42063143)
without external validation, you still just have a gut feeling. it doesn't matter if it's the first or thousandth.

I can tell if somebody ate cornflakes or oatmeal for breakfast just by looking at how they walk. I'm always right. you better believe me. I've seen thousands of people walk.

Re:Google Is Now Watermarking Its AI-Generated Text by kuhewa on Wednesday November 06, 2024 @03:52PM (#42069185)
If the cornflakes had been sitting in old milk for a week, then yes, I would believe you could tell by how they walk, either hunched over with a cramping stomach or with shit pants. LLM text is quite similar.
Re:Google Is Now Watermarking Its AI-Generated Text by exe34 on Wednesday November 06, 2024 @04:02PM (#42069345)
Yes because it would be completely different if it were oatmeal.
Re:Google Is Now Watermarking Its AI-Generated Text by kuhewa on Wednesday November 06, 2024 @04:10PM (#42069446)
If the oatmeal hadn't sat in old milk for a week, yes, it would not likely cause gi distress
Re:Google Is Now Watermarking Its AI-Generated Text by exe34 on Wednesday November 06, 2024 @05:03PM (#42070175)
and yet if it had, it would.
Re:Google Is Now Watermarking Its AI-Generated Text by kuhewa on Wednesday November 06, 2024 @05:19PM (#42070392)
Let's make one of your breakfast foods eggs. Sometimes you won't notice from the walk if they just had normal eggs. However, when they are also rotten and were over indulged in, you can tell from the person's walk that they ate them for breakfast and due to the noxious sulfur smell emanating from the diarrhoea in the person's business casual slacks, you can tell with a high degree of confidence it wasn't the rotten milk cornflakes, but one of a very few number of sulfur-rich foods, probably eggs.

Bad human resume text and overuse of unmodified LLM output are both detectable, but they are detectable because they are bad in quite different ways.

Regarding the original resume reader's notion that they can detect LLM text with a high degree of accuracy, it is not their LLM output detection specificity I would take issue with (similarly, despite stating validation is critical, I would bet you, too, are pretty confident when you see an entire page of blogspam or marketih copy that you regard as LLM generated despite it rarely being marked as such). Rather, it is their sensitivity, as I am sure occasional use and especially slightly modified output from LLMs gets by them now and again without them knowing.

Re:Google Is Now Watermarking Its AI-Generated Text by exe34 on Wednesday November 06, 2024 @05:28PM (#42070521)
yes, it's like men who think they can always spot makeup on a woman. or the economists who predicted 19 of the last 5 market crashes. no need for external validation.
Re:Google Is Now Watermarking Its AI-Generated Text by kuhewa on Wednesday November 06, 2024 @05:41PM (#42070690)
The makeup is a great analogu. I can bet with 99.9% that when I say a woman is wearing makeup that I'm correct (or it is tattoed on or similar). When it's obvious, it's obvious. However I don't detect makeup on women quite often.

The economists is not not as good of an analogy, almost converse of the makeup example as that is a high rate of false positives.

Re:Google Is Now Watermarking Its AI-Generated Text by exe34 on Thursday November 07, 2024 @02:12AM (#42074292)
that's why I gave both examples - you have no evidence which camp you are in, without ground truth. if you add gut feeling to gut feeling, you don't get evidence.
Re:Google Is Now Watermarking Its AI-Generated Text by kuhewa on Thursday November 07, 2024 @02:33AM (#42074431)
One camp is a high specificity camp, the other is a high sensitivity camp. You definitely know what camp you are in if you are only making the argument you can detect true cases of LLM use without many false positives- I've already admitted less blatant LLM use goes undetected so I am not arguing for high sensitivity.

And we have plenty of evidence of hallmarks of LLM use, we can even replicate the LLM resume generation process if we wanted. There is plenty of useful "training data" available even if you don't have a validated set of resumes submitted for this type of role at this type of company from this demographic of applicants.

Basically what you are trying to argue is that you can't have confidence that the animals you see people walking down the street on leashes are dogs unless you ask the owners whether they are dogs are not... AND that it doesn't matter that dogs are highly distinct from other domestic pets AND that we've seen many verified dogs before in other contexts, AND have even bred different varieties of dogs on our own.

I highly doubt that you maintain that standard for inductive inference across the board in your own practice. Life would be very difficult if you refused to make inferences about novel things (with any confidence) based on generalised patterns derived from other, similar cases.

Re:Google Is Now Watermarking Its AI-Generated Text by auggierose on Thursday November 07, 2024 @04:52AM (#42075225)
If there are alien shape shifters around which pose as dogs and look and behave just like them, then yes, your dog example is the same.
Re:Google Is Now Watermarking Its AI-Generated Text by exe34 on Thursday November 07, 2024 @05:24AM (#42075388)
> if you are only making the argument you can detect true cases of LLM use without many false positives

this seems at odds with the ongoing issues at universities where professors blindly trust AI detectors that label their own work as AI generated. you'd think if it were that obvious, they would have high specificity.

with your dog example, children absolutely have to learn by checking with their parents - you show them a dog and teach them the word, they will apply it to cats and goats and you have to correct them.

you are like a child pointing at every animal and calling it a dog, but refusing to shift your position when your elders tell you no, that's a goat.