alt Hacker NewsSynthID Detector
Comments
It's baffling that this requires auth. OpenAI's own tool (https://openai.com/research/verify/) doesn't require auth.
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.
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.
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.
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)
No artist wants their artwork be defined by the noise synthid ads.
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.
Here is openai:
https://help.openai.com/en/articles/8912793-provenance-signa...
There is rate limit though
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.
No option to detect generated text. SynthID also watermarks text, right?
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.
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.
This is a Google Deepmind project to “Identify AI generated media”
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!
Wouldn't it be easy to just generate a training dataset and use a model to identify and remove said watermarks?
There is a rate limit of around 10 checks in 24 hours.
(2025)? https://blog.google/innovation-and-ai/products/google-synthi...
Or did it just become public
Edit:
Blog post today https://blog.google/innovation-and-ai/models-and-research/go...
Hey OpenAi, generate a text of 100 words.
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 …
Can’t be used without signing in with a Google, Apple, or ChatGPT account.
Deepwalker cracked this one long time ago
https://deepwalker.xyz/blog/evaluating-synthid-watermark-rob...
[dead]
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!