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

81–90 of 189 posts

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

#81

Post by guappa »

retrac98 wrote: post #55
amonith wrote: post #46

Earlier quoted context omitted.

Eh, eventually AI will write like humans but currently most of the time it's very much apparent what was written by AI. English is my second language so it's hard for me to pinpoint the exact reason why but I guess it's more about the tone and the actual content (a.k.a bullshit) rather than grammar / choice of words. Most of the time AI slop reads like a soulless corporate ad. Probably because most of the content the…

If the AI writing is good, you’re not going to know it’s written by AI and you’ll continue to think you "can always tell" while more and more of what you read isn’t written by humans.

Yeah but to obtain good, you have to give such a specific and detailed prompt that you might just do it yourself.

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

#82

Post by meindnoch »

emiliobumachar wrote: post #9

Startup idea: online text editor that logs every keystroke and blockchains a hash of all logs every day. If you're accused of AI use, you can pull up the whole painstaking writing process and prove it's real.

Startup idea: keep humans in liquid-filled pods, connecting sensors to their central nervous system, and record every nerve impulse they generate. This way we can be 100% sure that those nerve impulses were generated by humans, and not an AI.

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

#83

Post by guappa »

gravescale wrote: post #78
immibis wrote: post #43

Earlier quoted context omitted.

If you're reading nonfiction, it means you're wasting time reading a lot more words when you could have just read the prompt.

Indeed you can losslessly "compress" an LLM's spew into just the prompt (plus any other inputs like values of random variables). But you can also compress a book's entire content into just its ISBN. It's just that books are hopefully more than just statistical mashups of existing content (some books like textbooks and encyclopaedias are kinds of mashup, though one hopes the editors have more than a statistically-base…

You can't regenerate the book from the ISBN. But you can generate the text from the prompt.

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

#84

Post by littlestymaar »

I know someone in this exact situation: used to be a copywriter for years, over the past few month got angry reviews from customers for «using AI» and decided to quit her job because getting new jobs was becoming harder and harder because of the bad reviews.

This is somewhat related to what Eliot Higgins (Bellingcat) said about generative AI:

> When a lot of people think about AI, they think, “Oh, it’s going to fool people into believing stuff that’s not true.” But what it’s really doing is giving people permission to not believe stuff that is true.

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

#85

Post by conartist6 »

Did anyone pay attention to how we made machine learning in the first place? We picked a task that only humans could do, and we made humans train a model to do it until of course it was no longer a task that only humans could do.

AI detectors will work exactly the same way. The effect of AI detectors on people is unfairness and misery, so people will be incentivized to remove the characteristics the detectors can find from AI output, and then other people will make better detectors, and the only possible outcome of this arms race is that it will no longer be possible at all to tell whether something was written by a machine or a person.

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

#86

Post by BurningFrog »

The one kind of AI detector that could work would be if the AIs store some checksum(s) of each text they write. Then you can ask it if it wrote a certain text.

With the naive version of this, you only have to change one word to get around the system.

A better version is to checksum smaller segments. Maybe a "chunk size" of 50 words is good. If you find several such chunks in a text, it's pretty clear you have a slightly altered AI text.

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

#87

Post by everdrive »

We're all at the mercy of the whims of imperfect people, but we just keep adding more ways to get things wrong. If feels like a step back, and people just can't stop inventing terrible things. The discourse is only "this new and also awful technology is just here to stay. In 5 years, someone will invent some other new horrible thing we'll all be at the mercy of, just we just have to get used to it." I don't have any better answers, but it's very discouraging.

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

#88

Post by meindnoch »

TeMPOraL wrote: post #72
emiliobumachar wrote: post #9

Startup idea: online text editor that logs every keystroke and blockchains a hash of all logs every day. If you're accused of AI use, you can pull up the whole painstaking writing process and prove it's real.

Stupid startup killing idea: an open-source script that runs LLM in the background and streams its output as input events, so the idiotic keylogger thinks it's all written by hand. Just writing this down here instantly invalidates the premise. An overkill variant to rub salt in the wounds of duped investors: make the script control a finger bot on an X/Y harness, so it literally presses the physical keys of a physica…

>An overkill variant to rub salt in the wounds of duped investors: make the script control a finger bot on an X/Y harness, so it literally presses the physical keys of a physical keyboard according to LLM output.

That's how the first automated trading firms operated in the 80s. NASDAQ required all trades to be input via physical terminals, so they build an upside down "keyboard" with linear actuators in place of the keys, that would be then placed on top of the terminal keyboard, and could input trades automatically.

https://www.npr.org/2015/04/23/401781306/we-built-a-robot-th...

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

#89

Post by cheq »

amonith wrote: post #46
bradley13 wrote: post #5

In some cases an AI will make a weird word choice. So do a lot of humans. Sometimes AIs are needlessly wordy. Um...so are a lot of humans. Rinse and repeat. AI detectors are useless. The AIs are training on human writing, so they write fundamentally like humans. How is this not obvious?

Eh, eventually AI will write like humans but currently most of the time it's very much apparent what was written by AI. English is my second language so it's hard for me to pinpoint the exact reason why but I guess it's more about the tone and the actual content (a.k.a bullshit) rather than grammar / choice of words. Most of the time AI slop reads like a soulless corporate ad. Probably because most of the content the…

"most of the content the AI was trained on was already SEO optimized bullshit mass produced on company blogs."

totally agree in the last part, I work with copywriting and I spent most of my prompts trying to double-down on the pretentious discourse.

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

#90

Post by amonith »

retrac98 wrote: post #70
amonith wrote: post #63

Earlier quoted context omitted.

Yeah but to reach that point you will probably need those "useless AI detectors" (as stated by the comment I was replying to). That was my point - we're not there yet therefore those tools can be useful.

But how do you know we’re not there yet? Not across the board, but isn’t it possible there’s a small yet growing portion of written content online that’s AI generated with no obvious tells?

I think we have a misunderstanding - I don't mind if I'm reading AI generated content as long as it doesn't look like "the typical AI content" (or SEO slop). In my point of view companies/writers might use AI detectors to continue improving the quality of their content (even if it's written by hand, those false positives might be a good thing). We're not there yet because I still see and read a lot of AI/SEO slop.

I agree with you that the "portion of written content online that’s AI generated with no obvious tells" is "small yet growing". That's exactly the thing - it's still too small to "be there yet" :)

Post reply on HN