Naively thought from the title that this was a way to detect what synthesizer / settings were used to create a sound in a song and I was pretty excited about that.
Finally, the previous implementations were very silly! OpenAI had a nice tool, but it only detected their own watermarks, and Google's process was "upload the image to Gemini and ask if it's AI generated", which seemed like a perplexing waste of tokens and breath.
(While a sibling comment points out putting the image through Google Image Search as an alternative, I don't remember this being signposted in the support article I've read, so I unfortunately didn't know about it)
The “ask Gemini” thing was so bad, because it made people think that worked with other chatbots. I know some smart, well-informed people who have been under the impression that they can ask Claude if something is AI-generated and get an accurate answer!
Even Google themselves previously offered a no-auth way (just very niche): if you uploaded an image to Google Image search, and went to "About this image", it would show if the image was Google AI-generated. Now it doesn't.
We need to inform everyone that this technology can encode database identifiers and enough entropy to uniquely identify you as an author (or downloader).
I'm confused why no one came up with this terminology for stenographic watermarks before. Now that there's an actual prosocial use for them (helping encourage clear labelling of AI-generated content), suddenly people are coining negatively-loaded terms for it, even though the technology has been in use for nefarious purposes for ages.
They are not spymarks if they don't encode any personal IDs and are merely used to indicate that an image was AI-generated. I don't think Google or OpenAI use SynthID to include personal data.
>I don't think Google or OpenAI use SynthID to include personal data.
Why? I assume by default that it's uniquely identifiable, because it just makes practical sense for OAI and Google (tracking misuse at the very least, but also data), it's trivial to implement and trivial to hide or plausibly deny.
I don't think it's trivial implement. Encoding more bits is harder or more brittle, and not trivial to hide. One could just ask for a completely white image. Any image distortions must then be due to SynthID. If the distortions differ for different accounts, that's strong evidence for varying watermarks.
It hasn't. Their DoubleCkick/AdSense (embedded website ads) used tracking, but Google built its "empire" through AdWords (keyword ads in Google Search), not AdSense.
Come on man, they're both advertising companies. Would you resist the opportunity to build that social graph and detect a known user anywhere on the internet?
Of course. Privacy is a feature that increases their product value. Moreover, they could get into trouble with the GDPR if they secretly included tracking information in their images.
So the obvious move is that marketing materials sing how it’s super private, while actual steganography encodes everything to uniquely identify you. If someone discovers something wave it away (and crack down on pesky whistleblowers), eat the fines if the worst happens.
It's presumably so they can rate-limit people who are trying to reverse-engineer or otherwise strip it (e.g. iteratively tweaking until it stops getting detected)
> iteratively tweaking until it stops getting detected
I don't think that dodgy folks have any issue with registering thousands of IDs. I know that I used to regularly get approached by these Chinese companies that were selling good reviews of my free apps. They did this by having thousands of legit AppleID accounts that would download and rate the app.
I haven't been approached by one of these in a long time. I guess Apple figured out how to put the kibosh on them.
How hard can it be to download a set of real images and generated ones in order to train a model to detect and strip the watermark with minimal perceptual difference?
Given that it's an ad company, they want something to connect what you're doing to the rest of their data, and they couldn't possibly keep the old image search results there if they want people to use this.
Yeah I've been wanting to run a ton of Wikipedia images through the detector to see if fakes snuck in. Doesn't seem to be a practical way to do this. Even Open AIs tool has a low rate limit
That’s the justification, but the EULA incorporates Google’s regular consumer terms; which means what you upload can also be used for advertising and targeting; and basically any purpose whatsoever by Google.
I've never seen it mentioned anywhere so I will just here complain that Google also removed the ability to paste images into Google Image search some years ago for no apparent reason. It worked great and I used it all the time.
In similar fashion in the Android Play Store they moved the search off from the top bar. For some time they educated users that "oo the search is no longer here, you need to click there -> to get to the page with search!". That has stopped (for me?), but they still haven't found any other use for same place in the first view.
OpenAI's tool is worse though, because it doesn't detect SynthID from non-OpenAI models. I just tried it with various images created by Imagen / Nano Banana.
It's a real shame Google isn't more transparent about how it actually works, and doesn't provide any mechanism for classifying images in bulk or offline.
They don't want you to know, because they're the baddies. Imagine how much money they can get from advertisers if they're able to identify every single piece of code, reddit thread, email, GitHub readme that you've ever written based on your secret ID. They're creaming themselves just thinking about it.
"Excellent work acquiring that outlook data Anat, now let's cross check the defunct company emails against YouTube videos edited with Google PrivateEditAI™ to figure out what the social security number of this YouTube account is."
Unfortunately this is the truth, no one can be given a benefit of the doubt because despite years of good grace they have been anti-user and enshittified everything they touch.
Not that I can think of, but you could have two watermarks, one detectable with an open-source classifier and the other proprietary.
Most AI images are either extremely low-effort or not actively trying to be deceptive, so defenders can still catch the majority. If someone is actively circumventing they'll probably circumvent both, anyway.
That paper explicitly states that the implementation is proprietary and the paper is not intended to describe the implementation itself. The paper focuses on the design considerations, technical challenges, and lessons learned from developing and deploying SynthID at scale. It explicitly omits critical implementation details such as the neural network architecture, training process and the loss functions used. The paper is really focused on deployment, not on SynthID's actual implementation.
I tested about 8 pictures. All 8 were made with OpenAI and Google. 2 were altered by and had text and picture overlays. The 6 didn't and SynthID caught them. The 2 altered passed. Interesting!
If I go directly to the URL it wants me to agree to some stuff before I even know what the service is. I had to come to these comments to figure that out (without agreeing to anything first).
Is there an open standard or something for people generating images to encode them or is the standard private and only shared with frontier model builders? That's a shame if latter. But pretty cool standardizing technology.
I worry we’ll have to approach this the other way around: verify that photos came from a camera, using hardware support like Apple’s Reference Image, rather than try to detect every AI generated one.
In many situations, photos are evidence. AI tools make convincing fakes, such as images of defect product.
I hate that Google don't have an API or open library to detect gemini images, unless you're enterprise. Create a problem then charge API access to get the solution.
The creepiest part of SynthID is that even the Chinese models are refusing to try and bypass it. Not sure if it's corruption from claudeslop data but they keep saying it's a crime to remove SynthID advertising tokens.
>I won't provide an operational recipe for stripping it, because the main use case is laundering AI content, which is deceptive and in many jurisdictions now illegal.
Ah but that's the genius of the scheme. Every single one of those providers will detect your secret ID and refuse the request. And sneak their own one in for good measure.
Like something more specific
(While a sibling comment points out putting the image through Google Image Search as an alternative, I don't remember this being signposted in the support article I've read, so I unfortunately didn't know about it)
Even Google themselves previously offered a no-auth way (just very niche): if you uploaded an image to Google Image search, and went to "About this image", it would show if the image was Google AI-generated. Now it doesn't.
https://brand.io/article/spymarks/
We need to inform everyone that this technology can encode database identifiers and enough entropy to uniquely identify you as an author (or downloader).
This extends to other forms of media as well.
It's not really possible to ban steganography, is it?
> That future lies halfway between now and 1984. So let’s stay off that timeline, shall we?
> Call a spymark what it is. A spy tool used to spy on you and everyone you interact with.
Why? I assume by default that it's uniquely identifiable, because it just makes practical sense for OAI and Google (tracking misuse at the very least, but also data), it's trivial to implement and trivial to hide or plausibly deny.
I don't think that dodgy folks have any issue with registering thousands of IDs. I know that I used to regularly get approached by these Chinese companies that were selling good reviews of my free apps. They did this by having thousands of legit AppleID accounts that would download and rate the app.
I haven't been approached by one of these in a long time. I guess Apple figured out how to put the kibosh on them.
How hard can it be to download a set of real images and generated ones in order to train a model to detect and strip the watermark with minimal perceptual difference?
Given that it's an ad company, they want something to connect what you're doing to the rest of their data, and they couldn't possibly keep the old image search results there if they want people to use this.
The textbox below it says Paste image link, but you can actually paste an image from your clipboard here, too.
I suppose it was too convenient.
URLs expand Googlebot’s indexes; pasted images don’t.
This is the best SynthID write-up I've found so far: https://fyx.me/articles/attempting-model-extraction-of-googl...
It covers how it actually works (probably), and how to train your own classifier for it, with some seemingly decent results.
"Excellent work acquiring that outlook data Anat, now let's cross check the defunct company emails against YouTube videos edited with Google PrivateEditAI™ to figure out what the social security number of this YouTube account is."
> classifying images in bulk or offline
You've described exactly the elements spammers and fraudsters need to be able to defeat this mechanism.
Most AI images are either extremely low-effort or not actively trying to be deceptive, so defenders can still catch the majority. If someone is actively circumventing they'll probably circumvent both, anyway.
https://help.openai.com/en/articles/8912793-provenance-signa...
There is rate limit though
how is "abc" different from "abc" generated by LLM
https://www.youtube.com/watch?v=kVXp6UNVPTo
In many situations, photos are evidence. AI tools make convincing fakes, such as images of defect product.
More detail at https://deepmind.google/models/synthid/
For those who, like me, were hoping it was a vision model to identify synthesizer models from photos of concerts and music studios!
>I won't provide an operational recipe for stripping it, because the main use case is laundering AI content, which is deceptive and in many jurisdictions now illegal.
Or did it just become public
Edit:
Blog post today https://blog.google/innovation-and-ai/models-and-research/go...
https://deepwalker.xyz/blog/evaluating-synthid-watermark-rob...
Hey Gemini, find a synonym for every second adjective. Replace in text.
Hey Grok, find a synonym for every third proper noun. Replace in text.
Hey …
Too short for watermarking but point taken of course.
Probably more like