SynthID Checker: How to Verify AI Images and Text

I've been digging into invisible watermarks, SynthID, StableSignature, TreeRing — the stuff baked into pixels by Gemini, DALL-E, etc. Can't see them, can't Photoshop them out, they survive screenshots. Got curious how robust they actually are, so I threw together noai-watermark over a weekend. It runs a watermarked image through a diffusion model and the output looks the same but the watermark is gone. A single pass at low strength fools SynthID. There's also a CtrlRegen mode for higher quality. Strips all AI metadata too.

Mostly built this for research and education, wanted to understand how these systems work under the hood. Open source if anyone wants to poke around.

github:

I thought it was about time to make a new PSA about SynthID. English is not my first language so there's probably some mistakes.

As you may be aware, this subreddit does not allow AI detectors. This is because they are extremely unreliable. Try going to deviantart or any other art site and test them with a few images, it should be easy to find something they say is AI despite it being uploaded over five years ago.

This is my go to example see how it was uploaded in 2018, there is no way this could be generated but nearly all the detectors I have tried say that it is.

Then why is SynthID allowed, you might say?

SynthID is different. It's a thing developed by Google, when you use their generators they add a hidden pattern to the image which is difficult to remove without lowering the quality of the image or changing it significantly. This is what it is looking for, not if it looks like it's generated like the other detectors. Therefore you can be sure if it says that SynthID is detected, an AI has done something to the image in some way. It could be something as simple as upscaling, changing something or it could be completely generated. The point is if it has SynthID you cannot trust the image.

Recently OpenAI(ChatGPT) have implemented the same technology in their generator. There is an important thing though Google and OpenAI can only detect their own images. Not each other's so you have to run the image through both of them.

So how do I use them?

For OpenAI it's very simple they have a page for it.

Google of course has to be more complicated. It used to be able to do a "search by image" and then go to about the image but that no longer works. Now you have to use Gemini to get an answer. Add the image you suspect and type "@Verify AI". Annoying but I don't know any better way.

Edit: When using Gemini it is very important that you type "@Verify AI" and not "SynthID" or any other variation of it. Otherwise there is a high chance that it will hallucinate and give you the wrong answer. It is also important to use a clean new session for each image.

Another edit: To confirm that Gemini actually use the verification tool you can press the three little dots underneath its response and then select "see response details".

Now I can't stress this enough, these two methods will only detect OpenAI's and Google's images. Not any other generator. So when an image comes up negative for SynthID that doesn't mean it's not AI. It can only prove that an image is fake not disprove it.

So before you post an image here please check for yourself you will save yourself some time and not waste others.

Final edit: as pointed out by a mod, this issue was caused by verifying an AI image first then asking it to verify other images in the same conversation. Yes, it showed it was using verify ai in the answer details but it’s still remembering the Synthid from the first check. Each image needs to be opened in another conversation.
————————————————————————

To be clear, this was taken unedited by my camera phone.

Flagged yes then after multiple photos I had recently taken flagged with a Synthid, I tried again and it said no.

ETA taken on an iPhone.

Also to add: While there is a Google portal for journalists and “verification professionals”, the rest of us are stuck with Gemini to verify Google SynthID. This route is not as perfect a tell of AI as I (and I suspect others as well) thought.

ETA many Synthid checks were done in one conversation. This seems to have confused the checker

I'm kind of OK with watermarks. Actually think its a smart idea. But as OP has shown easy to remove. Wish there was someway to enforce it.

Right now the internet and world are all upset about AI pulling out the pitchforks with AI posts. That won't last. Diffusion models and AI is here to stay. Once it becomes more widely accepted we'll all wish there was a way to sign AI stuff so that it is known and obvious.

'It's worth noting that in spite of being the same technology, Gemini's detector wont detect Open AI's SynthID signature and OpenAI's verifier tool won't detect Google's SynthID signature.

I tested this myself by generating an image of a beach in Gemini, downloading it from the chat conversation, and then uploading it to OpenAI's verifier tool, where it returned that the Image was not generated by OpenAI's tools.

It's insane to me that they're not interoperable. Seems like they could each just check for both signatures (and in the future, anyone else that adopts it could just all share their signatures around between verifiers), but here we are. You literally have to check one image in both places to be sure.

The only way to guarantee it actually uses the SynthID verifier is to type @Verify AI in the box. That forces it to use the tool and you should get consistently correct results. If you type anything other than that (“Synthid?”), Gemini might choose to use the verifier tool or it might decide to come up with an answer on its own.

After a decent amount of research and sanity loss, I have developed a way to reliably remove the SynthID watermark from generated images.

the guy who tricks old people on facebook a few weeks ago: 'darn it, people know im using AI!'

the same guy today: 'check out this statue of jesus made by frogs'

With the last version of synthid, all you had to do was add a layer with gaussian noise and lower the opacity to 3% and take a screenshot of the canvas -- it's not quite as simple anymore.

On the new version, I found adding noise to the top of the image required much higher levels of noise that really degraded the image quality before it would pass, and then Gemini would flag it as uncertain because of the noise.

So that got me thinking, what if I put the noise behind the image and lowered the opacity and then brought the levels back up to match the original? I ended up using two layers of noise, one monochrome/one not to create more variation in the tones.

That worked pretty well. I took a screenshot of final product and voila, synthid not detected. Took about 10 minutes to figure this out.

I only put this out there because people are going to be leaning on these tools pretty heavily as a source of truth for what is or is not an AI image. The fact is that they're still very easily fooled and completely fallible.

Original image generation:

No synthid detected:

That's not a positive for synth ID it's just Google claiming it's AI.

Hey guys,

With the Gemini 3 rollout, the question of watermarking keeps coming up. We did some digging into the tech docs and the new app features (plus what the SynthID-Text open-source repo tells us) to give a definitive answer.

Here the link to the Repo:

TL;DR: YES, Gemini text is watermarked, but only Google can easily prove it.

Google has explicitly confirmed that the text generated by their user-facing chatbots (Gemini on the web and in the app) is watermarked using SynthID-Text.

• How it works: It's not visible text. It's a subtle statistical signature embedded during the generation process by slightly changing the word probability scores (token selection). This pattern is robust against simple edits, copy-pasting, and paraphrasing.

This is the part that matters most:

|| || |Detection Method|Result|Reason| |Google's Detector (in the Gemini App)|Yes (High Confidence)|Gemini has the private, secret keys needed to decode the watermark pattern.| |Open-Source SynthID-Text (from the GitHub repo)|No (Reports "Uncertain")|The open-source code gives you the method but not the keys used by Google's production models.| |Third-Party Detectors (GPTZero, WasItAIGenerated, etc.)|Unreliable / Varies|These detectors look for stylistic patterns or public watermarks (like C2PA metadata), not Google's secret statistical signature. They're often wrong.|

Bottom line: If a professor or client wanted to definitively prove your text was from Gemini, they would need an authorized version of Google's detector. Your generic checker won't work.

Since the core problem is a statistical signature, the solution is deep, structural rewriting. You can't just change a few words; you have to break the underlying pattern of token probabilities.

• Rephrasy AI Humanizer: This is one of the tools we've personally developed and that seems to work against the statistical pattern checks. Unlike simple paraphrasers that just swap synonyms (which SynthID is designed to resist), humanizers like Rephrasy aim to restructure the text's rhythm, sentence complexity, and flow—which is exactly what's needed to disrupt the watermarking signal.
• The Claim: Tools like Rephrasy often claim a high success rate against detectors like Turnitin and GPTZero (which look for AI-like style), but the evidence suggests they also significantly reduce the confidence score of the statistical watermarks by altering the probability distribution of the original tokens.
• Use Case: If you need to polish Gemini-generated text for high-stakes submissions (where you can't risk even a low "AI score"), running it through a robust humanizer is currently the best adversarial approach.

• Gemini App (Consumer): Watermarking is enforced by default.
• Gemini API (Developers): For text endpoints, watermarking is highly likely to be implemented, though it may not be a simple on/off switch like it is for Images (where the addWatermark parameter is clear). Google's public statements confirm it's a core safety feature for the text model.

• Transparency: It's a step toward verifiable content, which is good for fighting misinformation.
• Plausible Deniability: Minor editing and paraphrasing won't completely remove the statistical signal, but tools like Rephrasy are designed to lower the confidence score to "Uncertain" or "Not Watermarked" by applying adversarial rewriting.
• No Universal Detector: Just because a third-party tool says your text is "human" doesn't mean Google's internal tool would agree.

What do you all think? Is the text watermark a necessary step for trust, or just another headache for high-stakes writing?

I work a lot with AI, and after seeing several Gemini-generated images, I started seeing something which I have reason to believe is a SynthID artifact.

I can only visually see it on anime/illustration-style images. I don't see it in realistic images.

I thought one of the reasons why I was clocking it only by seeing the image was because I was generating said images with Gemini, thus I knew what to look for.

But then recently I randomly saw an article on an IT news website which had an illustration as cover image, and I immediately noticed the same artifacts. I copy-pasted it onto Gemini and it did indeed contain the SynthID watermark.

For the image I've included as an example in this post, it's on the girl's hair/face. It looks like very subtle diagonal white/black banding. I can see it naked eye, but for your reference, I've ampliated it and messed with the curves (second gallery image) to try and make it more visible.

I've asked a friend who's also working with AI image generation and he claims that he, too, can see the same thing, and that it often "destroys" the image visually for him.

I've seen Google's papers. They claim SynthID to be invisible and not alter image quality, but it seems that these claims are grounded on tests where people were shown random images, some with the watermark and others not. It seems they never did a test where people could compare between the same image with and without SynthID.

Curious, what's baked in? aside from the meta data (which can be removed) what's visually in the pixel? and how can it not be taken out? that's interesting.

Open ai verification doesnt work for me for a reason. I think it's because of my country

Can you link me the chat that gave you these results? There are possible explanations for this discrepancy. The scroll bar on the side of the false positive result indicates to me that you did not start a fresh session when running SynthID on the image.

, there weren't enough votes to determine the quality of your post...

Hi,

the SynthID technology has been spreading quickly and I've been wondering if someone tried the watermark remover already?

I don't wanna get detected basically on

Thanks for your help