Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AI Detectors Get It Wrong. Writers Are Being Fired Anyway

gizmodo.com

Post reply on HN
Print view View original

181–189 of 189 posts

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#181

Post by TeMPOraL »

Lalabadie wrote: post #170
TeMPOraL wrote: post #58

Earlier quoted context omitted.

LLMs don't average , they learn the distribution , from which you then sample (or the UI does it for you). Because of that, they don't write in a single style that's a blend of many human styles - they can write in any and all of the human styles they saw in training , as well as blend them to create styles entirely "out of distribution". And it's up to your prompt (and sampling parameters) which style will be used.

True, a better choice of words than average would have been: a very likely completion to a starting state/input.

Perhaps, but in the context of this thread, what's important is that the space of possible completions is encompassing every writing style imaginable and then some, and the starting state/input can be used to direct the model to arbitrary points in that space. Simple example template:

  Please write , .
Where = "as if you were a pirate", or "be extremely succinct", or "in the style of drunk Shakespeare", or "in Iambic pentameter", or "in style mimicking the text I'm pasting below", etc.

There's no way those "AI detectors" could determine whether the text was written by AI from text itself, as it's trivial to make LLM output have any style imaginable.

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#182

Post by TeMPOraL »

dartos wrote: post #168
TeMPOraL wrote: post #151

Earlier quoted context omitted.

The word you're looking for is "any", not "average". As in, it can write like any human, any way you want it to. Not just as some "average human".

The word “average” implies a use of statistics, which is why I think it’s good to use, even though it’s not precise.

People aren't as dumb as you seem to imply here; "average" isn't accurate enough. "Random", or even "statistical", would be less confusing.

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#183

Post by gravescale »

guappa wrote: post #180
gravescale wrote: post #91

Earlier quoted context omitted.

You can go and fetch the book from a book store using the information. Fundamentally there's not much difference between that and "fetching" the output from some model using the matching prompt. In both cases there some kind of static store of latent information that can be accessed unambiguously using a (usually) shorter input. I'm not saying the value of the returned information is equivalent, of course. But being…

You realise that you can't fetch a new isbn without altering the archive, while this is not the case for every new prompt that you come up with?

I don't understand the distinction. If the book archive is electronic, like many in fact are, why can you not get a copy of the book with a given ISBN without altering anything? Even if it's not electronic, does the acquisition of a book by an individual meaningfully change the overall disposition of available information? If you took the last one in your local Waterstones, I can still get one elsewhere.

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#184

Post by guappa »

gravescale wrote: post #183
guappa wrote: post #180

Earlier quoted context omitted.

You realise that you can't fetch a new isbn without altering the archive, while this is not the case for every new prompt that you come up with?

I don't understand the distinction. If the book archive is electronic, like many in fact are, why can you not get a copy of the book with a given ISBN without altering anything? Even if it's not electronic, does the acquisition of a book by an individual meaningfully change the overall disposition of available information? If you took the last one in your local Waterstones, I can still get one elsewhere.

> If the book archive is electronic, like many in fact are, why can you not get a copy of the book with a given ISBN without altering anything?

Because new books are written?

It feels to me that you are set on insisting that a prompt and an ISBN are the same, and no amount of logic will move you from there.

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#185

Post by jgalt212 »

luma wrote: post #157
jgalt212 wrote: post #64

Earlier quoted context omitted.

> OpenAI announced that they had started on an AI text detector and then gave up as the problem appears to be unsolvable. Making a reliable LLM also appears to be unsolvable, but we still work at it and still use the current wonky iterations. My comment is even if there is no perfect AI detectors, a lot of these tools are good enough for a "first pass"--coincidentally the same use case many effective LLM practitioner…

Sure it could maybe be kinda right, but what is the cost of a false positive? If you have, say, a 10% false positive rate, and there are theoretical reasons to think you’ll never get that anywhere close to zero, then what use case does this serve? Hey student, there’s 90% chance you cheated, well no I’m that 10%. What now? Again, OAI cancelled work on this believing it not to be solvable with a high degree of confide…

> What is the use case for a low confidence AI detector?

What's the use case for LLMs in general if you always have to double-check their work?

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#186

Post by gravescale »

guappa wrote: post #184
gravescale wrote: post #183

Earlier quoted context omitted.

I don't understand the distinction. If the book archive is electronic, like many in fact are, why can you not get a copy of the book with a given ISBN without altering anything? Even if it's not electronic, does the acquisition of a book by an individual meaningfully change the overall disposition of available information? If you took the last one in your local Waterstones, I can still get one elsewhere.

> If the book archive is electronic, like many in fact are, why can you not get a copy of the book with a given ISBN without altering anything? Because new books are written? It feels to me that you are set on insisting that a prompt and an ISBN are the same, and no amount of logic will move you from there.

Models can be trained more and fine tuned, though, if we're going to stick to the analogy. But in the context of the analogy, the LLM won't be materially updated between two prompts in roughly the way that telling you that the answer you seek is in a book with a specific ISBN isn't materially affected by someone publishing a new book at that moment.

You are quite right that you're not convincing me of your original thesis that that a prompt contains the entire content of the reply in a way that some other reference to an entity in some other pool of information to doesn't. That's not the same as saying "ISBNs and LLM prompts are the same thing", which is a strawman. It's saying that they're both unambiguous (assuming determininism) pointers to information.

Of course no-one is disagreeing that a reply from a deterministic LLM would add no information to the global system (you, an LLM's model, a prompt) than just the prompt would. But I still think the same is true for the content of a book not adding to the system of (you, a book store, an ISBN).

In fact, since random numbers don't contain new information if you know the distribution, one can even extend it to non-deterministic LLMs: the reply still adds no information to the system. The analogy would then be that the book store gives you at random a book from the same Dewey code as the ISBN you asked for. Which still doesn't increase the information in the system.

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#187

Post by luma »

jgalt212 wrote: post #185
luma wrote: post #157

Earlier quoted context omitted.

Sure it could maybe be kinda right, but what is the cost of a false positive? If you have, say, a 10% false positive rate, and there are theoretical reasons to think you’ll never get that anywhere close to zero, then what use case does this serve? Hey student, there’s 90% chance you cheated, well no I’m that 10%. What now? Again, OAI cancelled work on this believing it not to be solvable with a high degree of confide…

> What is the use case for a low confidence AI detector? What's the use case for LLMs in general if you always have to double-check their work?

“The entire current AI market”.

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#188

Post by cesarb »

brookst wrote: post #167
cesarb wrote: post #47

Earlier quoted context omitted.

> Conceptually it seems like the average of all human texts would be distinct from any users because it would blend word choices and idioms across regions That's only true in the aggregate. Within a single answer, LLMs will try to generate a word choice which is more likely _given the preceding word choices in that answer_, which should reduce the blending of idioms. > where most of us are trained and reinforced in a…

While I agree that many people visit two regions of the same country, I think few of us would display word choice patterns reflecting the US, England, and Australia within a single piece. Could it happen? Sure. But LLMs won't have the bias towards likely combinations, except inasmuch as that's represented in training data.

> I think few of us would display word choice patterns reflecting the US, England, and Australia within a single piece.

Someone who learned English mostly through books and the Internet could very well have such a mixture, since unlike native speakers of English, they don't have a strong bias towards one region or the other. You could even say that our "training data" (books and the Internet) for the English language was the same as these LLMs.

Re: AI Detectors Get It Wrong. Writers Are Being Fired Anyway

#189

Post by dartos »

TeMPOraL wrote: post #182
dartos wrote: post #168

Earlier quoted context omitted.

The word “average” implies a use of statistics, which is why I think it’s good to use, even though it’s not precise.

People aren't as dumb as you seem to imply here; "average" isn't accurate enough. "Random", or even "statistical", would be less confusing.

We can keep bikeshedding, sure.

I don’t think people are dumb, just that the vast majority of people dont have knowledge of statistics and AI.

“Random” isn’t good either. Obviously the text gpt and other LLMs generate isn’t random.

Statistical works too, sure.

Post reply on HN