▲ 110 points
back
83 comments
> Large language models generate text one word (token) at a time. Each word is assigned a probability score, based on how likely it is to be generated next. So for a sentence like “My favourite tropical fruits are mango and…”, the word “bananas” would have a higher probability score than the word “airplanes”.
> SynthID adjusts these probability scores to generate a watermark. It's not noticeable to the human eye, and doesn’t affect the quality of the output.
I think they need to be clearer about the constraints involved here. If I ask What is the capital of France? Just the answer, no extra information.” then there’s no room to vary the probability without harming the quality of the output. So clearly there is a lower bound beyond which this becomes ineffective. And presumably the longer the text, the more resilient it is to alterations. So what are the constraints?
I also think that this is self-interest dressed up as altruism. There’s always going to be generative AI that doesn’t include watermarks, so a watermarking scheme cannot tell you if something is genuine. It is, however, useful for determining that something came from a specific provider, which could be valuable to Google in all sorts of ways.
This might be enforced in some trivial way, e.g. by requiring AI models to answer with at least a sentence. The constraints may not be fully published and the obscurity might make it more efficient, if only temporarily.
Printer tracking dots[1] is one prior solution like this; annoying, largely unknown, workarounds exist, still - surprisingly efficient.
For answers like that, it probably wouldn't matter whether it was AI-generated or not. It becomes more relevant with long-form generated content
Security and surveillance products don’t have to be perfect to be useful enough to some.
Choosing the slightly less probable output is changing the quality of the output if it weren't LLMs wouldn't work by processing a large amount of data to get these probabilities as accurate as possible.
> It is, however, useful for determining that something came from a specific provider, which could be valuable to Google in all sorts of ways.
Oh crap, knowing Google it probably means they will put articles generated using their AI higher among the search results.
1. One-sample detection is impossible. These detection methods work at the distributional level—more like a two-sample test in statistics—which means you need to collect a large amount of generated text from the same model to make the test significant. Detecting based on a short piece of generated text is theoretically impossible. For example, imagine two different Gaussian distributions: you can never be 100% certain whether a single sample comes from one Gaussian or the other, since both share the same support.
2. Adding watermarks may reduce the ability of an LLM, which is why I don’t think they will be widely adopted.
3. Consider this simple task: ask an LLM to repeat exactly what you said. Is the resulting text authored by you, or by the AI?
For images/video/audio, removing such a watermark is very simple. By adding noise to the generated image and then using an open-source diffusion model to denoise it, the watermark can be broken. Or in an autoregressive model, use an open-sourced model to do generation with "teacher forcing" loll.
Would you really use Google products to write your email if you knew that they were watermarking it like this?
I think this technology is gonna quickly get eliminated from the marketplace, cause people aren’t willing to use AI for many common tasks that are watermarked this way. It’s ultimately gonna cause Google to lose share.
This technology has a basic use dilemma problem where widely publishing it’s ability and existence will cause your AI to stop being used in some applications
People use much more invasive tools than that. If it works, they don’t care.
I think it’s weirder that they’re clamoring to give people tools to detect AI while clamoring to present AI-generated content as perfectly normal— no different than if the user had typed it in themselves.
Have you weighed factors that would push in the other direction? This requires a synthesis, and it requires breaking out of the tendency to only think about factors that support one narrative.
To the extent watermarking technology builds trust and confidence in a product, this is a factor that moves against your prediction.
Talk is cheap. People sometimes make predictions just as easily as they generate words.
While I want to believe this is true, the experience of being human over the past decade or so suggests otherwise. I think, overall, most people would not even begin this line of inquiry; not to mention care once the thought is considered.
Could anybody explain how this isn't easily circumvented by using a competitor's model?
Also, if everything in the future has some touch of AI inside, for example cameras using AI to slightly improve the perceived picture quality, then "made with AI" won't be a categorization that anybody lifts an eyebrow about.
> Could anybody explain how this isn't easily circumvented by using a competitor's model?
If the problem is "kids are using AI to cheat on their schoolwork and it's bad PR / politicians want us to do something" then competitors' models aren't your problem.
On the other hand, if the problem is "social media is flooded with undetectable, super-realistic bots pushing zany, divisive political opinions, we need to save the free world from our own creation" then yes, your competitors' models very much are part of the problem too.
By lobbying regulators to force your competitors to add watermarks too.
I wonder if this will survive distillation. I vaguely recall that most open models answer “ I am chat gpt” when asked who they are, as they’re heavily trained on openai outputs. If the version of chatgpt used to generate the training data had a watermark, a sufficiently powerful function approximator would just learn the watermark.
If you see the mark, you'd know at least that you aren't dealing with a purely mechanic rendering of whatever-it-is.
> Could anybody explain how this isn't easily circumvented by using a competitor's model?
Almost all the big hosted AI providers are publicly working on watermarking for at least media (text is more of a mixed bag); ultimately, its probably a regulatory play—the big providers expect that the combination of legitimate concerns and their own active fearmongering, combined with them demonstrating watermarking, will result in mandates for commercial AI generation services to include watermarking. This may even be part of the regulatory play to restrict availability and non-research use of open models.
I feel this is not the scalable/right way to approach this. The right way would be for human creators to apply their own digital signatures to the original pieces they created (specialised chips on camera/in software to inject hidden pixel patterns that are verifiable). If a piece of work lacks such signature, it should be considered AI-generated by default.
> If a piece of work lacks such signature, it should be considered AI-generated by default.
That sounds like a nightmare to me.
Then you just point the special camera at a screen showing the AI content.
That seems like a horrible blow to anonymity and psuedonymity that would also empower identity thieves.
I guess this is the start of a new arms race on making generated content pass these checks undetected and detecting them anyway.
Its not really an arms race; any gen AI system that doesn't explicitly incorporate a watermarking tool like this won't be detectable by tools that read the watermarks.
There is a kind of arms race that has existed for a while for non-watermarked content, except that the detection tools are pretty much Magic 8-ball level of reliability, so there's not a lot of effort on the counter-detection side.
I really hope SynthID becomes a widely adopted standard - at the very least, Google should implement it across its own products like NotebookLM.
The problem is becoming urgent: more and more so-called “podcasts” are entirely fake, generated by NotebookLM and pushed to every major platform purely to farm backlinks and run blackhat SEO campaigns.
Beyond SynthID or similar watermarking standards, we also need models trained specifically [0] to detect AI-generated audio. Otherwise, the damage compounds - people might waste 30 minutes listening to a meaningless AI-generated podcast, or worse, absorb and believe misleading or outright harmful information.
[0] 15,000+ ai generated fake podcasts https://www.kaggle.com/datasets/listennotes/ai-generated-fak...
Given there is "misleading or outright harmful" information generated by humans, why is it more pressing that we track such content generated by AI?
Hey I made an open source version of this last week (albeit for different purposes). Check it out at: https://github.com/sutt/innocuous
There's lot of room for contributions here, and I think "fingerprinting layer" is an under-valued part of the LLM stack, not being explored by enough entrants.
OpenAI has been doing something similar for generated images using C2PA [0]
It is easy to alter by just saving to a different format or basic cropping.
I would love to see how SynthID is fixing this issue.
https://help.openai.com/en/articles/8912793-c2pa-in-chatgpt-...
No this is very different. C2PA is just some extra metadata, it doens't watermark the image.
You can never be sure something has been generated by a model embedding one of these anyway, so it’s pretty moot.
Create the problem, sell the solution.
The whole NFT thing could be used to mark content, pictures and hashes of texts for example.
Would be nice if you guys open source the detection code, similar to the way C2PA is open
That's like asking for Adobe to open source their C2PA signing keys.
AI watermarking is adversarial, and anyone who generates a watermarked output either doesn't care, or wants the watermarked removed.
C2PA is cooperative: publishers want the signatures intact, so that the audience has trust in the publisher.
By "adversarial" and "cooperative", I mean in relation to the primary content distributor. There's an adversarial aspect to C2PA, too: bad actors want leaked keys so they can produce fake video and images with metadata attesting that they're real.
A lot of people have a large incentive to disrupt the AI watermark. Leaked C2PA keys will be a problem, but probably a minor one. C2PA is merely an additional assurance, beyond the reputation and representation of the publishing entity, of the origin of a piece of media.
There is a repo: https://github.com/google-deepmind/synthid-text
the beginning of walled garden “AI” tools has been interesting to follow
Also, what will happen if you cut and paste some part or the whole image inside another bigger one, like traditional photo editing?
And if I scan the image or take a picture of the image on display.
"The watermarks are embedded across Google’s generative AI consumer products, and are imperceptible to humans."
I'd love to see the data behind this claim, especially on the audio side.
I wonder what exactly would prevent a developer from removing the signature from a generated file. One could remove arbitrary segments that signal that it is AI generated.
i find the premise to be an invalid one personally - why is the property that a works from an AI model must be identified/identifiable?
The first good use of blockchain comes to mind.
If I slightly edit plain text watermarked with it, will the watermark identification be robust?
I am not sure that text watermarking will be accurate, I foresee plenty of false positives.
I wonder if, conversely, authentic media can be falsely watermarked as AI-generated.
For image, what happen if I screenshot it? Will the watermark survive?
looks like the same as anti-virus companies in the 80s? Write virus, Write anti-virus and profit!
It only works across Google shit.