OpenAI Is Adding Text Watermarking In ChatGPT and Codex
- Reference: 0186035538
- News link: https://slashdot.org/story/26/10/06/0443237/openai-is-adding-text-watermarking-in-chatgpt-and-codex
- Source link:
> OpenAI [2]says its textGrain watermarking "matched or exceeded" other approaches like Google DeepMind's SynthID for text, which is also the basis for the watermarking Anthropic announced in August. Like OpenAI, Anthropic made the move to meet the requirements of the EU's AI Act; however, not everyone was happy to learn about the addition. OpenAI also included scores from AI benchmarks showing similar performance from watermarked and unwatermarked text. But it notes that textGrain "does not guarantee reliable detection," and says text watermarks don't verify accuracy, determine who owns the text, measure how much a human contributed, or prove human authorship.
[1] https://www.theverge.com/ai-artificial-intelligence/1004880/openai-chatgpt-text-watermarks-eu-ai-act
[2] https://openai.com/index/eu-text-provenance/
So what? (Score:2)
It is a problem for chatgpt users, and that's far from their worst problem. Using chatgpt is more of a problem.
Re: (Score:2)
Oh... wah wah... can't use their ChatGPT.
Maybe your kids can learn how to write their essay by hand.
Re: (Score:2)
They do. In cursive, too.
Writing is a very important skill to have, you can't do advanced stuff well without it.
If you want to be really stupid .. (Score:4)
.. make yourself dependent on AI.
Let it think for you. Just do what it says, it is so much easier. Vote Billionaire, vote yourself poor! Vote for freedom from owning anything and renting it all!
You don't need powers of cognition, you don't need an opinion, we'll sell it to you. I don't know how you'll pay for that without a job, who knows? Maybe AI does, I don't really care, I'm RICH!
AI. It's like watching evolution take a leap backwards.
Re: (Score:2, Interesting)
It can actually do incredible things. Dump the contents of an eeprom and upload it to Claude.
Ask it to decompile the binary back into c code.
done
Analyze the source and figure out what string the cpu is looking for in order to boot
done
Write me a python script that sends this string thru my serial port
done
Now I have a working piece of test equipment.
Dumb people are going to do dumb things with smart tools.
Re: (Score:2)
Genius... you could just understand the code and write a program.
Re: (Score:2)
There is a very small group of people that think that's incredible, the rest wont get it.
And even for you to get to that point required you to have a deep understanding of what you wanted to achieve, maybe years of experience to describe it.
Good luck to you, unless you are rolling your own LLM, every time you use these commercial LLMs you are devaluing your work towards redundancy and giving away your talent to people who are rich enough already.
The newbie your boss wont hire now, because they think you can
Re:If you want to be really stupid .. (Score:4, Interesting)
Especially since you know the AI tools are going to be enshittified in short order. That ChatGPT query is free today, tomorrow once you're hooked and can't live without it, it's going to be a dollar query. 50 cents if you can wait for it to be answered during "off peak" hours.
Because that's the business model that's likely to win out. Get kids hooked on AI then get to bill everyone $50/month to use it because they've gotten so reliant on it that they can't live without it.
The whole AI causing shortages is to ensure that you don't get a chance to run local models yourself without paying for the privilege.
Re: (Score:1)
Just wait until AI is baked into your cell phone, fridge, toaster, microwave, TV, car... oh, wait... it's already there!
Soon, I can see it switching to paid-only... you can't even toast a slice of bread or call your mom unless you pay the $250 AI bill (if it's not just automatically withdrawn from your bank).
Re: (Score:2)
I'm surprised you didn't use the word "slop". Spewing a bunch of slogans that would have to be more clever to make it on to bumper stickers makes you look like the people you are complaining about. There are legitimate criticisms to make about AI, and ways to support and defend those arguments. You are not doing that.
Very funny (Score:2)
After you paste the AI-generated text into your dissertation in Libreoffice, the 'watermark' is gone.
Re: (Score:3)
First, I'd paste it into notepad just to make sure.
Re: Very funny (Score:2)
The watermark is in the pseudorandom number generator that makes the words, genius. You can copy it in purple crayon and it will still show up if the text is long enough.
Re: (Score:2)
So how is this different from the way it's always worked? It's never been difficult to spot AI slop precisely because of how similar it all is.
Re: (Score:2)
"AI slop" is so 2024. That argument isn't valid any more.
Re:Very funny (Score:5, Insightful)
Sorry, but that's not how this works.
At the softmax, the hidden state of the last layer output is converted back to token-space / text-space. This is a high-dimensional-space to low-dimensional-space conversion, so there is usually no "single right path" in which the target concept can be represented. The softmax in effect converts the nearest token pathways into probabilities; the closer the token's position to the latent position, the higher the probability.
So you have a probability-scaled list of options. Sometimes, the token that comes after is almost 100% certain. For example:
"This artifact was found in the tomb of the boy-pharaoh Tutankh amun "
Sometimes it's not:
"Ravi was hungry, so he climbed the mango tree and picked (... a mango)"
"Ravi was hungry, so he climbed the mango tree and carefully (...picked a mango)"
"Ravi was hungry, so he climbed the mango tree and grabbed (...a mango)"
Etc. There may be a whole wide range of possibilities.
A primitive watermarking algorithm, thus, can be represented in the "even-odd" algorithm. You divide up all tokens randomly into even or odd bins. On even-numbered generated tokens, you slightly increase the odds of tokens from the "even" bin and decrease them from the "odd" bin - maybe, say, "picked" declines in odds and "grabbed" rises. On odd-numbered generations you reverse your boosting. The probability shifts aren't huge, so it's still going to say "Tutankhamun", not "TutankhJetBrains", regardless of whether it's an even or odd token. But it might flip, say, "picked" vs. "carefully" vs. "grabbed" or whatnot. With a large enough sample size - and it doesn't have to be huge - you can tell whether this token bias exists.
This is, as mentioned, a primitive algorithm, and it has vulnerabilities - but it's not hard to see how you can adapt it to more advanced algorithms that are less vulnerable to user manipulation.
So no, you can't get rid of the watermark just by cut and paste. It's embedded in the choice of wording itself.
Re: (Score:2)
TL/DR , your options are:
1) Rewrite it in your own words
2) Have a model make a summary or shorthand version of the watermarked version, then have a non-watermarking model flush it back out
3) Use a non-watermarking model to begin with.
Re: (Score:3)
Dear Gemini,
Please paraphrase this GPT-inspired homework so that I sound like Copilot submitting my assignment at 11:59pm after consuming seven cups of coffee to keep myself awake.
Re: (Score:2)
It'll still get tagged by SynthID.
You have to use a non-watermarking LLM. Saying "Don't do anything that will watermark it" doesn't help.
Re: (Score:2)
SynthID in text genuinely is invisible. The probability shifts are small, and all the pathways are valid pathways to express the same thought.
And no, you cannot "collide" text watermarks. At best, it'll only have the latter one's watermark. At worst, both.
Re: (Score:3)
OR, now hear me out:
4) do it yourself without AI.
Re: (Score:3)
Here's [1]a video [youtube.com] about how it works. Really quite interesting.
[1] https://www.youtube.com/watch?v=Cmi-1QSaptA
So what's it good for? (Score:2)
"OpenAI also included scores from AI benchmarks showing similar performance from watermarked and unwatermarked text. But it notes that textGrain "does not guarantee reliable detection," and says text watermarks don't verify accuracy, determine who owns the text, measure how much a human contributed, or prove human authorship. "
Oh, I see, it is good for Company Misdirection: "OOooooo.....look at that Blibbering Humdinger in the window" (to reuse a Luna Lovegood phrase).
Re:So what's it good for? (Score:4, Interesting)
It's not hard to see how this can backfire. People treat watermarks as the Word of God. But let's say, for example, a journalist, in an article, quotes a White House press statement for something, but the White House used AI. Since they don't "measure how much a human contributed", and detect the watermark in the White House statement in the journalist's article, they'll just flag the whole journalist's article as AI.
And honestly, "not measuring how much a human contributed" is IMHO a massive flaw in general even if the author was using AI . If the person wrote an article, and then told an AI, "correct my spelling, grammar, and poor phrasing", that's IMHO entirely different from the person just telling an AI "Write this article for me", and just pasting whatever it spits out in as their own, whole cloth. Contribution assessment is, IMHO, essential for fairness. And also, eminently doable. It is perfectly technologically possible to not just see, "does this signature exist", but even get a sense of exactly what the AI contributed vs. what it didn't.
Re: (Score:2)
Clever bot, you removed your em-dash from your reply!
Re: (Score:2)
Exactly, my top use of AI is probably, this prompt:
Please review the document below for typographical and grammatical errors and make the required edits. Do not suggest edits that simply alter style or tone. You may make clarity changes, such as splitting a compound sentence into two separate thoughts or exchanging a pronoun for the proper noun it references. Replacing a less specific nouns such as 'thing' and 'it', with a more concrete noun being referenced; as long as it will not read as overly verbose
Re: (Score:2)
I don't think that makes the document any less your own than having a human copy editor conduct a similar review which has been done since forever.
Which is what [1]I've been saying as well [slashdot.org]. If all you're doing is asking for suggestions to something you have already written, using AI is no different than using a human editor. It's your work. It's been written. You're just looking for advice on how to improve it.
This is completely different than people who tell AI what they want and it does all the w
[1] https://slashdot.org/comments.pl?sid=24075534&cid=66298890
AI done foolishly (Score:2)
They are so bent on faking intelligence we are not allowed specialized tools - or when we had them those are being replaced with these fake general intelligence chat bot.
Grammar checkers are old; they did get better and I read some so good they were re-writing too much and actually limited due to expert feedback which users may have liked. Today to get a good grammar checker is hard because even the ones you pay for are trying to replace them with general AI bots.
It wouldn't hallucinate, etc. if it wasn't
Translation (Score:2)
I've had students legitimately using translation tools to get better english. With AI, they do much better but also it looks like AI and I can't possibly tell if they had AI do the work or if they merely translated or corrected their poor mixed english answers. So then everything has to be evaluated without the tools they come to depend upon. Tools that are legitimate if limited to translation. The subject isn't english and is not a requirement for competency in that subject; they could go back home with
Re: (Score:2)
Exactly. I give a few embedded/industrial linux lessons at the university. Some students regularly give me code with comments like "/* and here replace with the name of your variable */" or similar. Some people are not very bright.
Re: (Score:2)
Perhaps you should be realizing by now that homework is no longer a viable way to test how well students actually know the subject matter, and the only viable way to do so is in-class tests/exercises - frequent enough and worth enough of the grade to force them to study between lessons.
OR, alternatively, you could just keep doing what you're doing, and instead, in effect grade people on:
Comparably poor grades:
* People who actually do the work by hand
* "AI N00bs" who let stuff like that slip through
Comparabl
Re: (Score:2)
It takes one hell of a toll to try to do verbal exams 1:1 on students to evaluate them. In person exams in a monitored room are not so easy either... not that we have a huge problem now with undetectable ear plugs or smart glasses or anal beads or remote controls... but that tech progresses and demand will increase with more in person testing. A radio shielded test room would work well until local run AI fits in your underwear. Then you have an increasing number of IEPs so you can't even do a quiz without e
Re: (Score:2)
As Eric Raymond (ESR) says, AI is a force multiplier. My elaboration: If you are good at what you do, it will let you do accomplish more good stuff faster. If you are not very bright, AI will let you be not bright more efficiently, but it won't make you bright. It's just a tool. It needs to be utilized correctly. There is no shortage of things in this world that can do tremendous damage if not used properly. AI is nothing new in that regard.
Paywalled AI Marketing (Score:3, Interesting)
Can we stop promoting this crap.
Re: (Score:2)
Just get the archive page extension then you can bypass paywalls.
Re: (Score:2)
Watermarks are "marketing"?
What user wants watermarks?
Re: (Score:3)
Another story about AI is marketing.
Who cares about watermarking? An AI company conceding to watermarking to respond to slow legislation is like closing a tiny barn door after the T-Rex has bolted and run off with all your copyrights.
When we discuss these stories, express our interest, we're just helping the marketing, spread the bullshit, pump the stock, build data centres, and help Trumpy get a tiny little cheeto hardon. Ivana doesn't need that, she's been through enough already, just look at her eyes, th
Re: (Score:3)
I'll repeat: watermarks are anti-marketing . They discourage use.
Re: (Score:2)
> I'll repeat: watermarks are anti-marketing. They discourage use.
Every chance to get another story about how AI is so powerful it must be regulated serves at least two purposes. First, it is an advertisement for the power of the technology. Second, it is a plea to allow the entrenched AI companies to write regulation to permit them to control the industry. These are not extremely complicated ideas, though they do require that you not cede responsibility for thinking to software which is incapable of doing it because the tokens don't mean anything to it. It has the same a
you are here (Score:2)
understandable frustration, but you forget that this site lives off advertising, particularly for other outlets looking for visibility. it's natural that they service the stampede du jour, the endless stream of chatter and circular nonsense about ai (or crypto or whatever). also, in this particular case the subject has at least some mild technical interest.
what other generalist technically oriented forums are free from this eternal plague? even soylentnews is swamped with it ...