Note that watermarking (yes, including text) is a requirement[1] of the EU AI Act, and goes into effect in August 2026, so I suspect we'll see a lot more work in this space in the near future.
[1] Specifically, "...synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated", see https://artificialintelligenceact.eu/article/50/
Well, consider the case with murder: they're not demanding that people proactively implement a system to prevent it from happening, are they? You're just not allowed to do it, in the sense that the system will attempt to find you, prove your guilt, and punish you after the fact.
You can trivially enforce that at the AI provider level, which covers 99% of the problem the law is designed to address.
Of course it doesn't cover the issue of foreign state psyop operations but the fact that enforcing laws against organized crime and adversary state actors is hard isn't specific to AI.
Are you not aware of open-weights models and local generation? I think the vast majority of deepfake content is being genned in basements on RTX cards, not on public providers. People already have all this content, and have archives of it, and can run it airgapped. Cat is out of bag.
I genuinely feel that in this AI world we need the inverse. That every analogue or digital photo taken by traditional means of photography will need to be signed by a certificate, so anyone can verify its authenticity.
Years ago, I worked at Apple at the same time as Ian Goodfellow. This was before ChatGPT (I'd say around 2019).
I had the chance to chat with him, and what I remember most was his concern that GANs would eventually be able to generate images indistinguishable from reality, and that this would create a misinformation problem. He argued for exactly what you’re mentioning: chips that embed cryptographic proof that a photo was captured by a camera and haven't been modified.
And how do you fix the analog hole? Because if you can point your "verified" camera at a sufficiently high-resolution screen, we're worse off than when we started.
Yes, I’m more worried about the false confidence such technology could create. Implement an authenticity mechanism and it will be treated as truth. Powerful people will have the means to spoof photographic evidence.
That is presumably a very expensive endeavor. We already have hardware that attempts to mitigate this and while I think it's possible for the government it's certainly not trivial.
I'm sure Apple would love that too. More seriously, would that also mean all editing tools would need to re-sign a photo that was previously signed by the original sensor. How do we distinguish an edit that's misleading vs just changing levels? It's an interesting area for sure, but this inverse approach seems much trickier.
To be fair, I think just signing details about the way an image was assembled makes sense. Deciding on fake vs real doesn't have to be done at time of capture. We store things like the aperture size, sensitivity, camera name/model, etc in the EXIF data, including details about the image processing pipeline seems like a logical step. (With a signature verification scheme... and I guess also trying to embed that in the actual bitmap data)
There is no original image to recover, since we can't capture and describe every photon, so it's not a "fake vs real" image signature... that would be a UI choice the image viewer client would make based on the pipeline data in the image.
The text watermarking is the more interesting problem here. Image watermarking is fairly tractable - you can embed a robust signal in spatial or frequency domains. Text watermarking works by biasing token selection at generation time, and detection is a statistical test over that distribution.
Which means short texts are basically useless. A 50-token reply has too little signal for the test to reach any confidence. The original SynthID text paper puts minimum viable detection at a few hundred tokens - so for most real-world cases (emails, short posts, one-liners) it just doesn't work.
The other thing: paraphrase attacks break it. Ask any other model to rewrite watermarked text and the watermark is gone, because you're now sampling from a different distribution. EU compliance built on top of this feels genuinely fragile for anything other than long-form content from controlled providers.
Long-form content from controlled providers is by far the lion's share of what needs this regulation, at least at the moment. Perfect is the enemy of good enough. Or at least of better than the status-quo.
This is great, but there is no way for me to verify if groups or nation states can pay for a special contract where they do not have to have their outputs watermarked.
These sorts of tools will only be able to positively identify a subset of genAI content. But I suspect that people will use it to 'prove' something is not genAI.
In a sense, the identifier company can be an arbiter of the truth. Powerful.
Training people on a half-solution like this might do more harm than good.
It will just be an arms race if we try to prove "not genAI." Detectors will improve, genAI will improve without marking (opensource and state actors will have unmarked genAI even if we mandate it).
Marking real from lense through digital life is more practical. But then what do we do with all the existing hardware that doesn't mark real and media that preexisited this problem.
I agree. A mechanism to voluntarily attach a certificate metadata about the media record from the device seems like a better idea. That still can be spoofed, though.
In the end, society has always existed on human chains of trust. Community. As long as there are human societies, we need human reputation.
You could take a picture or video with your phone of a screen or projection of an altered media and thereby capture a watermarked "verified" image or video.
None of these schemes for validation of digital media will work. You need a web of trust, repeated trustworthy behavior by an actor demonstrating fidelity.
You need people and institutions you can trust, who have the capability of slogging through the ever more turbulent and murky sea of slop and using correlating evidence and scientific skepticism and all the cognitive tools available to get at reality. Such people and institutions exist. You can also successfully proxy validation of sources by identifying people or groups good at identifying primary sources.
When people and institutions defect, as many legacy media, platforms, talking heads, and others have, you need to ruthlessly cut them out of your information feed. When or if they correct their mistake, just follow tit for tat, and perhaps they can eventually earn back their place in the de-facto web of trust.
Google's stamp of approval means less than nothing to me; it's a countersignal, indicating I need to put even more effort than otherwise to confirm the truthfulness of any claims accompanied by their watermark.
It is actively harmful to society. Slap SynthID on some of the photographic evidence from the unreleased Epstein files and instantly de-legitimize it. Launder a SynthID image through a watermark free model and it's legit again. The fact that it exists at all can't be interpreted in any other way than malice.
It's security through obscurity. I'm sure with the technical details or even just sufficient access to a predictive oracle you could break this.
But I suppose it ads friction so better than nothing.
Watermarking text without affecting it is an interesting seemingly weird idea. Does it work any better than (with knowledge of the model used to produce said text), just observing the perplexity is low because its "on policy" generated text.
This technology could be used to copyrights as well.
>The watermark doesn’t change the image or video quality. It’s added the moment content is created, and designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression.
But does it survive if you use another generative image model to replicate the image?
It doesn't. I don't have a link for you right now but there was a post on reddit recently showing that SynthID is removed from images by passing the image through a diffusion model for a single step at low denoise. The output image is identical to the input image (to the human eye).
Reposting a comment I made on an earlier thread on this.
We need to be super careful with how legislation around this is passed and implemented. As it currently stands, I can totally see this as a backdoor to surveillance and government overreach.
If social media platforms are required by law to categorize content as AI generated, this means they need to check with the public "AI generation" providers. And since there is no agreed upon (public) standard for imperceptible watermarks hashing that means the content (image, video, audio) in its entirety needs to be uploaded to the various providers to check if it's AI generated.
Yes, it sounds crazy, but that's the plan; imagine every image you post on Facebook/X/Reddit/Whatsapp/whatever gets uploaded to Google / Microsoft / OpenAI / UnnamedGovernmentEntity / etc. to "check if it's AI". That's what the current law in Korea and the upcoming laws in California and EU (for August 2026) require :(
I've been looking into this. There seems to be some mostly-repeating 2D pattern in the LSB of the generated images. The magnitude of the noise seems to be larger in the pure black image vs pure white image. My main goal is to doctor a real image to flag as positive for SynthID, but I imagine if you smoothed out the LSB, you might be able to make images (especially very bright images) no longer flag as SynthID? Of course, it's possible there's also noise in here from the image-generation process...
Gemini really doesn't like generating pure-white images but you can ask it to generate a "photograph of a pure-white image with a black border" and then crop it. So far I've just been looking at pure images and gradients, it's possible that more complex images have SynthID embedded in a more complicated way (e.g. a specific pattern in an embedding space).
I just tried this idea, and it looks like it isn't that simple.
> "Generate a pure white image."
It refused no matter how I phrased it ¯\_(ツ)_/¯
> "Generate a pure black image."
It did give me one. In a new chat, I asked Gemini to detect SynthID with "@synthid". It responded with:
> The image contains too little information to make a diagnosis regarding whether it was created with Google AI. It is primarily a solid black field, and such content typically lacks the necessary data for SynthID to provide a definitive result.
Further research: Does a gradient trigger SynthID? IDK, I have to get back to work.
Note that watermarking (yes, including text) is a requirement[1] of the EU AI Act, and goes into effect in August 2026, so I suspect we'll see a lot more work in this space in the near future.
[1] Specifically, "...synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated", see https://artificialintelligenceact.eu/article/50/
EU really like unenforceable regulations, doesn't it?
Yeah, they also outlawed murder. And stealing. And bribing officials. All universally unenforceable. Weird...
Well, consider the case with murder: they're not demanding that people proactively implement a system to prevent it from happening, are they? You're just not allowed to do it, in the sense that the system will attempt to find you, prove your guilt, and punish you after the fact.
> they're not demanding that people proactively implement a system to prevent it from happening, are they?
What do you think a "background check" is?
It's regulation for providers, so it's easy to enforce.
What do you mean? There is nothing unenforceable about this.
How would you prove that something was generated by AI yet did not include a watermark?
You generate it with that particular AI and look for the watermark :/
You can trivially enforce that at the AI provider level, which covers 99% of the problem the law is designed to address.
Of course it doesn't cover the issue of foreign state psyop operations but the fact that enforcing laws against organized crime and adversary state actors is hard isn't specific to AI.
Are you not aware of open-weights models and local generation? I think the vast majority of deepfake content is being genned in basements on RTX cards, not on public providers. People already have all this content, and have archives of it, and can run it airgapped. Cat is out of bag.
It's nice that they explain the "what" (...it is doing) but not the "why". Who is going to use it and for what reasons?
Also, if it's essentially a sort of metadata, can't the output generated image be replicated (e.g. screenshot) and thus stripped of any such data?
I genuinely feel that in this AI world we need the inverse. That every analogue or digital photo taken by traditional means of photography will need to be signed by a certificate, so anyone can verify its authenticity.
This already exists: https://c2pa.org , https://en.wikipedia.org/wiki/Content_Authenticity_Initiativ... . Support by camera makers is - spotty.
Years ago, I worked at Apple at the same time as Ian Goodfellow. This was before ChatGPT (I'd say around 2019).
I had the chance to chat with him, and what I remember most was his concern that GANs would eventually be able to generate images indistinguishable from reality, and that this would create a misinformation problem. He argued for exactly what you’re mentioning: chips that embed cryptographic proof that a photo was captured by a camera and haven't been modified.
And how do you fix the analog hole? Because if you can point your "verified" camera at a sufficiently high-resolution screen, we're worse off than when we started.
Yes, I’m more worried about the false confidence such technology could create. Implement an authenticity mechanism and it will be treated as truth. Powerful people will have the means to spoof photographic evidence.
Or just extract the certificate from the hardware you own.
That is presumably a very expensive endeavor. We already have hardware that attempts to mitigate this and while I think it's possible for the government it's certainly not trivial.
This is a "solved" problem. Vendors whose keys are extractable get their licenses revoked. The verifier checks the certificate against a CRL.
Depth sensor information.
I'm sure Apple would love that too. More seriously, would that also mean all editing tools would need to re-sign a photo that was previously signed by the original sensor. How do we distinguish an edit that's misleading vs just changing levels? It's an interesting area for sure, but this inverse approach seems much trickier.
You'd have to provide both images, and let the end user determine whether they think it's misleading.
Some cameras support this, but usually only for raw.
Note that your cell phone camera is using gen AI techniques to counteract sensor noise.
Was that famous person in the background really there, or a hallucination filling in static?
Who knows at this point? So, the signatures you proposed need to have some nuance around what they’re asserting.
To be fair, I think just signing details about the way an image was assembled makes sense. Deciding on fake vs real doesn't have to be done at time of capture. We store things like the aperture size, sensitivity, camera name/model, etc in the EXIF data, including details about the image processing pipeline seems like a logical step. (With a signature verification scheme... and I guess also trying to embed that in the actual bitmap data)
There is no original image to recover, since we can't capture and describe every photon, so it's not a "fake vs real" image signature... that would be a UI choice the image viewer client would make based on the pipeline data in the image.
As a synthesizer collector with serious GAS I find this particular name very offensive.
The text watermarking is the more interesting problem here. Image watermarking is fairly tractable - you can embed a robust signal in spatial or frequency domains. Text watermarking works by biasing token selection at generation time, and detection is a statistical test over that distribution.
Which means short texts are basically useless. A 50-token reply has too little signal for the test to reach any confidence. The original SynthID text paper puts minimum viable detection at a few hundred tokens - so for most real-world cases (emails, short posts, one-liners) it just doesn't work.
The other thing: paraphrase attacks break it. Ask any other model to rewrite watermarked text and the watermark is gone, because you're now sampling from a different distribution. EU compliance built on top of this feels genuinely fragile for anything other than long-form content from controlled providers.
Long-form content from controlled providers is by far the lion's share of what needs this regulation, at least at the moment. Perfect is the enemy of good enough. Or at least of better than the status-quo.
Link to the paper please?
The act doesn't explicitly require watermarking, does it?
haha "you" say this, when your comment was written by an LLM! it's watermarked!
Is there a paper for this?
How about a database of verified non-AI images?
I'm thinking of historical images, where there aren't a huge number of existing images and no more will ever be created.
If I see something labeled "Street scene in Paris, 1905". I want to know if it is legit.
This is great, but there is no way for me to verify if groups or nation states can pay for a special contract where they do not have to have their outputs watermarked.
These sorts of tools will only be able to positively identify a subset of genAI content. But I suspect that people will use it to 'prove' something is not genAI.
In a sense, the identifier company can be an arbiter of the truth. Powerful.
Training people on a half-solution like this might do more harm than good.
It will just be an arms race if we try to prove "not genAI." Detectors will improve, genAI will improve without marking (opensource and state actors will have unmarked genAI even if we mandate it).
Marking real from lense through digital life is more practical. But then what do we do with all the existing hardware that doesn't mark real and media that preexisited this problem.
I agree. A mechanism to voluntarily attach a certificate metadata about the media record from the device seems like a better idea. That still can be spoofed, though.
In the end, society has always existed on human chains of trust. Community. As long as there are human societies, we need human reputation.
You could take a picture or video with your phone of a screen or projection of an altered media and thereby capture a watermarked "verified" image or video.
None of these schemes for validation of digital media will work. You need a web of trust, repeated trustworthy behavior by an actor demonstrating fidelity.
You need people and institutions you can trust, who have the capability of slogging through the ever more turbulent and murky sea of slop and using correlating evidence and scientific skepticism and all the cognitive tools available to get at reality. Such people and institutions exist. You can also successfully proxy validation of sources by identifying people or groups good at identifying primary sources.
When people and institutions defect, as many legacy media, platforms, talking heads, and others have, you need to ruthlessly cut them out of your information feed. When or if they correct their mistake, just follow tit for tat, and perhaps they can eventually earn back their place in the de-facto web of trust.
Google's stamp of approval means less than nothing to me; it's a countersignal, indicating I need to put even more effort than otherwise to confirm the truthfulness of any claims accompanied by their watermark.
It is actively harmful to society. Slap SynthID on some of the photographic evidence from the unreleased Epstein files and instantly de-legitimize it. Launder a SynthID image through a watermark free model and it's legit again. The fact that it exists at all can't be interpreted in any other way than malice.
It's security through obscurity. I'm sure with the technical details or even just sufficient access to a predictive oracle you could break this.
But I suppose it ads friction so better than nothing.
Watermarking text without affecting it is an interesting seemingly weird idea. Does it work any better than (with knowledge of the model used to produce said text), just observing the perplexity is low because its "on policy" generated text.
You'd be surprised what dog owners do sometimes.
Seems like this really just validates whether a piece of AI content was generated by Google, not AI generated in general
What incentive do open models have to adopt this?
This is from 2025. Did something new happen? What am I missing here?
This technology could be used to copyrights as well.
>The watermark doesn’t change the image or video quality. It’s added the moment content is created, and designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression.
But does it survive if you use another generative image model to replicate the image?
It doesn't. I don't have a link for you right now but there was a post on reddit recently showing that SynthID is removed from images by passing the image through a diffusion model for a single step at low denoise. The output image is identical to the input image (to the human eye).
> This technology could be used to copyrights as well.
That's been a thing for a while: https://en.wikipedia.org/wiki/Digital_watermarking
Extremely doubtful, due to the way that embedding and diffusion works. I would be utterly floored if they had achieved that.
Reposting a comment I made on an earlier thread on this.
We need to be super careful with how legislation around this is passed and implemented. As it currently stands, I can totally see this as a backdoor to surveillance and government overreach.
If social media platforms are required by law to categorize content as AI generated, this means they need to check with the public "AI generation" providers. And since there is no agreed upon (public) standard for imperceptible watermarks hashing that means the content (image, video, audio) in its entirety needs to be uploaded to the various providers to check if it's AI generated.
Yes, it sounds crazy, but that's the plan; imagine every image you post on Facebook/X/Reddit/Whatsapp/whatever gets uploaded to Google / Microsoft / OpenAI / UnnamedGovernmentEntity / etc. to "check if it's AI". That's what the current law in Korea and the upcoming laws in California and EU (for August 2026) require :(
Looks like there's a lot more info here, at least about the text version.
https://ai.google.dev/responsible/docs/safeguards/synthid
something new here OP?
Some previous discussion:
https://news.ycombinator.com/item?id=45071677
I wonder how it stands up to feature analysis.
"Generate a pure white image." "Generate a pure black image." Channel diff, extract steganographic signature for analysis.
I've been looking into this. There seems to be some mostly-repeating 2D pattern in the LSB of the generated images. The magnitude of the noise seems to be larger in the pure black image vs pure white image. My main goal is to doctor a real image to flag as positive for SynthID, but I imagine if you smoothed out the LSB, you might be able to make images (especially very bright images) no longer flag as SynthID? Of course, it's possible there's also noise in here from the image-generation process...
Gemini really doesn't like generating pure-white images but you can ask it to generate a "photograph of a pure-white image with a black border" and then crop it. So far I've just been looking at pure images and gradients, it's possible that more complex images have SynthID embedded in a more complicated way (e.g. a specific pattern in an embedding space).
I just tried this idea, and it looks like it isn't that simple.
> "Generate a pure white image."
It refused no matter how I phrased it ¯\_(ツ)_/¯
> "Generate a pure black image."
It did give me one. In a new chat, I asked Gemini to detect SynthID with "@synthid". It responded with:
> The image contains too little information to make a diagnosis regarding whether it was created with Google AI. It is primarily a solid black field, and such content typically lacks the necessary data for SynthID to provide a definitive result.
Further research: Does a gradient trigger SynthID? IDK, I have to get back to work.