TL;DR: SynthIDs are watermarks, that AI companies are legally required (thanks EU) to embed in Slop Imagery, so that they can reliably tell people if images are Slop.
The watermark sits in all the details of an image, as seen by the ai it is tested for, so as long as a good section of image remains, even downscaled, cropped, stretched, ... it is still detectable.
Currently it seems only OpenAI images have infested lemmy. Their testing page can be used without account.
*queue tutorial music*
Hello and welcome to my synthid explanation lemmy post :)
You probably heard that AI "detectors" are usually immense trash, unreliable, and basically just ask some ai themselves.
In my experience, this is true.
SynthID is something completely else, even if some of Big Slop likes to pretend otherwise.
Is the check AI?
No. It can't be, since I just called it "reliable" in the TL;DR.
It makes use of the model, but to inspect "what it would have done", not to consult it.
The method just uses normal cryptography, and a deterministic algorithm, running on the image-to-check and the model-to-check-against, and see if they "match".
Reliable?
The details are down below, but basically synthid sits in all randomness added to the decisions of generative AI.
If there are enough decisions matching the secret "password" the companies have, then you get a practical certainty the decisions in that original image were in fact made by that "model" with that "password".
Now, just because there is a threshold of astronomical probableness, that anyone would call "certain" in layman terms, doesn't necessarily mean the companies are using such a responsible threshold.
Unfortunately, we can't know how they are doing it, they won't give us the password for reasons that are also down below. So we can't check for ourselves.
The threshold is very high though. Since it gives every clean image an equal probability floor of false detection, the fact we don't have any known false positives, means it should be rarer than one in a million.
Annoyingly, every corporation could set their own threshold differently.
Why can only Big Slop itself check SynthIDs?
Probably a few reasons.
Mainly, you need both the "model" and password to check. So by releasing the means to check, they would also release the model to anyone and let them generate without adding a synthid, or with simply changing the password and not telling anyone.
Even if they somehow manage to only release part of the "model", maybe a section responsible for some of the finer detail or later steps, then releasing the password would let people more easily reverse engineer those stages, potentially undoing them and changing the decisions to change the synthid.
Knowing the noise would also help reverse engineers figure out the true "model" outputs better and let them more efficiently distillation train their own ai. And as we all know, training your ai on Big Slops ai is theft and very very bad and please respect intellectual property. IP rights are very important to Big Slop!
Lastly, naturally, Big Slop don't want to make it any easier than legally required. If more people get away with slop because it's less accessible to prove and fewer people do it, then sloppery gets more widespread.
This is also the reason these are inconvenient to use, and noone has the api access to make free 3rd party sites that check multiple synthids at once and are nicer to use.
How?
The process uses the way the "model" (image sloppery) works, and a secret password to generate predictable noise, to prove that that noise and that "model" were used to make that image.
The companies of course promise (are legally required to) use such passwords to slop all their images.
To simplify, generative AI currently works via prediction. You make it predict existing shit, and eventually, it, as a byproduct, also becomes good at predicting non-existent shit that's kinda similar to all that.
For that, it currently ends up with a preference rating for a set list of choices. So whatever task you give it, you break it into limited choices, and make the neural net rate them, by assigning each choice a preference value. For text that could be the next character, word, or in practice "token" (for efficiency reasons).
Usually, for a few reasons, you don't want to pick the top ranked choice every time, but somewhat at random from among the highest rated choices.
Images are made small to large, you start with a low res smudge, and "upscale" it with a description in mind. The ai will hallucinate interpretations into the noise step by step until you get an image. That's diffusion. The randomness among the choices here is usually an image of noise blended over the last step's output-image.
SynthID basically replaces this randomness with cryptographic noise, which comes out to the same result, picking randomly from the top rankings based on their preference values.
Then, if you have access to the model, you can reverse the process of creating the image, and see what sort of decisions were picked at each step. If you then also have your cryptographic noise, you can compare them. And if there is an extremely improbable match, then only someone with access to that noises seed value ("password") could have made those decisions (since to those without knowledge of the password everything was just random).
The noise is random for each image, itself "salted" by that image, simply so that the "randomness" remains "random", making the slop part of the operation work better, and also to prevent anyone from figuring out the "password" (seed values) and messing with the process.
I found this video, which goes even more in-depth. It's about text, but still applies.
https://www.youtube.com/watch?v=FnAqruxx-QE
Here is the paper for synthid: https://arxiv.org/abs/2510.09263
