News OpenAI has built a text watermarking method to detect ChatGPT-written content — company has mulled its release over the past year

[Admin]

OpenAI built a text watermarking tool that helps it detect whether a piece of content was written by ChatGPT. However, internal debates rage on whether it should actually release it or not.

OpenAI has built a text watermarking method to detect ChatGPT-written content — company has mulled its release over the past year : Read more

 

[husker]

"For example, you can insert ChatGPT’s output in Google Translate, convert it to another language, and then back to English, effectively removing the watermarking."

... and all semblance of good writing.

 

[velocityMarcus]

I rarely see any developments in regards to checks/balances for AI content. I know folks have been making "anti-AI" filters for their digital art, and there are sites that will help you find out whether or not something has been written by AI, but as far as internal solutions go, it's a bit refreshing to hear about this. The "99.9%" success rates and "[undetectability] by humans" is something I'd love to see for myself.

In the eyes of a company, of course there's internal turmoil. Most folks will dodge AI written content if it's clear that it's written by an AI. If it was clear to the user right off the rip that they were reading something that was generated, they might just stop reading.

I think it begs the question - should companies creating AI be responsible for integrating systems that make it clear that the content an AI produces should be disclosed as generated? I think so.

 

[derekullo]

I've had chatgpt write dozens of versions of a paper for me.
I then pick out the best parts of each one and stitch them together in my own words.
The entire reason for an LLM's existence is to emulate the writing style of humans.
Trying to prove that its writing style is not human is a fool's errand since they are only going to get better at it.

rephrase this in the words of a smart person. Frazier, Plato, Freud

I've utilized ChatGPT and other LLMs to generate multiple iterations of a paper, subsequently selecting the finest elements from each version and integrating them into my own composition.
The fundamental purpose of such a model is to mimic the nuanced and sophisticated writing styles of humans.
To challenge its capacity for human-like expression is an exercise in futility, as these models are destined to evolve and mirror our own linguistic patterns with increasing fidelity.

 

[cycro]

OpenAI doesn’t even have the best models anymore. They’re already falling behind. Adding watermarking would be the final blow to make people switch to better LLMs.

 

[lete479]

"
a professor once failed an entire class because all the submitted papers were inaccurately marked as AI-generated."

The fact that a simple search shows this to be false make me believe that the whole article is false. I see no sources for the declares truths. Somehow some random guy just knows what ChatGPT is doing? I doubt it.

 

[Omniversum]

"
a professor once failed an entire class because all the submitted papers were inaccurately marked as AI-generated."

The fact that a simple search shows this to be false make me believe that the whole article is false. I see no sources for the declares truths. Somehow some random guy just knows what ChatGPT is doing? I doubt it.

Now, you see, that's weird.. Because I can easily find this fact(?) in several mainstream media.

So, please provide us with your sources that are claiming that this fact is untrue.

 

[lete479]

Now, you see, that's weird.. Because I can easily find this fact(?) in several mainstream media.

So, please provide us with your sources that are claiming that this fact is untrue.

“A&M-Commerce confirms that no students failed the class or were barred from graduating because of this issue,” the school noted. “Dr. Jared Mumm, the class professor, is working individually with students regarding their last written assignments. Some students received a temporary grade of ‘X’ — which indicates ‘incomplete’ — to allow the professor and students time to determine whether AI was used to write their assignments and, if so, at what level.” The university also confirmed that several students had been cleared of any academic dishonesty.

Perhaps there is another university story that you've found?

 

[lemons123]

OpenAI doesn’t even have the best models anymore. They’re already falling behind. Adding watermarking would be the final blow to make people switch to better LLMs.

you didn't even list a single LLM that is "better" than OpenAI's

 

[cycro]

you didn't even list a single LLM that is "better" than OpenAI's

It’s hard to say “better” since there’s so many different use cases. The best attempt at ranking I’ve seen is the chatbot arena leaderboard: https://chat.lmsys.org/?leaderboard

Google currently holds the top spot. OpenAI’s best LLM currently ranks #2, with Anthropic and Meta close behind.

If you look at text embedding models, OpenAI doesn’t even rank in the top 20. My point is, they have no magic sauce. They don’t have a monopoly on good models anymore - they’re not even the current best in any category. There is no moat. I’m hoping and believe we’ll see more open models like Llama dominate in the future.

 

[TJ Hooker]

Now, you see, that's weird.. Because I can easily find this fact(?) in several mainstream media.

So, please provide us with your sources that are claiming that this fact is untrue.

All the articles I found link back to a single rolling stone article, which only has an anonymous redditor as a source. As lete479 mentions, that rolling stone article was updated with a statement from the university that contradicts the redditor's claims.

 

[SirStephenH]

I don't accept OpenAI's claim that its fingerprinting is imperceptible to humans.

Text isn't like a video, picture, or audio where there's so much going on that a digital fingerprint can slip through unnoticed by a human observer. You have to change the content of the result in a way that would be directly seen by the observer. Instead of perhaps imperceptibly changing the color or brightness of pixels in an image, it would be more like changing a robin to a bluebird. Most people may not notice but to many, I'm sure, it would stick out like a sore thumb.

Isn't the purpose of LLMs to as closely emulate human speech as possible? This seems like it would be a step in the wrong direction to that goal.


I'd suggest just running an LLM offline at home, you don't even need a supported GPU with loads of VRAM. Most methods run on CPU with GPU offloading. You could run them completely on CPU but they would be very slow. The overall size of the module you can use is dependent on available system memory while the amount you can offload to the GPU is generally dependent on available video memory. I've been using Llama 3 Uncensored Lumi Tess Gradient 70B Q5_K_M lately. It's 50GB and seems to be just about the limit my 64GB of memory can handle.

I'd highly recommend Backyard AI. It's simple to use and is geared towards the use of character cards, with the ability to import SillyTavern PNGs. It's a very simple install. If you have a supported GPU then just enable it in the main settings and set memory to auto and you won't have to think about it again. Then just download a model and character and go. You also have the ability to set it up as your own personal online LLM, available to you anywhere using your account in browsers or mobile apps (they call it tethering). It also has a free/paid online service but you don't need to worry about that, just set it to use your personal computer.

LM Studio is more along the line of the basic prompt/system prompt setup. It's easy to setup and models are easier to browse but GPU offloading is done in "slices" that have to be manually set each time you start a new chat.