AI is Hacked - SynthID flags AI Content During Generation

[theeyeinside]

https://www.cyberseo.net/blog/google-geminis-synthid-a-hidden-trap-for-autoblogging/

How SynthID works​

In text, this is not done through metadata or hidden symbols, but rather, at the level of token probabilities. When a language model selects the next word, it uses a probability distribution. For example, the probabilities for “cat,” “dog,” and “bird” are 0.31, 0.28, and 0.14, respectively. SynthID shifts these probabilities slightly according to a pseudorandom rule.

The result is text that remains readable and natural to humans but that, when analyzed for word sequences (e.g., n-grams of five tokens), exhibits a distinctive statistical pattern. This pattern is invisible to the reader, but Google’s detector can identify it: “This text was generated and watermarked by SynthID.”

The principle is similar for images. During generation, tiny distortions are embedded at the pixel and frequency levels. These distortions are imperceptible to the human eye, yet they survive compression, resizing, cropping, and filtering. With its detector, Google can easily identify images created by its model.

If you want to dive deeper into the technical details, check out the Introducing SynthID Text.

---------

Eventually all the models will do this and AI content will be easily uncovered

Writers will rejoice and not lose their jobs after all ...

 

[moonshine7000]

Won't that hurt Google bottom line to make money.Who uses an AI API,a spammer.

 

[RoboTech]

Google be like: Subscribe to our cutting-edge AI. Your life will be so much easier!

Also Google: Wait… you actually used AI? You absolute idiot.

 

[Builder Dave]

And then everybody will shift to self-hosted AI without SynthID or to AI providers without SynthID.

 

[moonshine7000]

And then everybody will shift to self-hosted AI without SynthID or to AI providers without SynthID.

Google thinks the spammer market is stupid

 

[Vikjohn]

Open source is the future of ai

 

[flackz]

Won't that hurt Google bottom line to make money.Who uses an AI API,a spammer.

Just as if they use any other blackhat tool. Isn't that not something we are up to at this forum?

 

[4440]

is it a problem?

Use Ai from other suppliers to post-edit your AI content multiple times here and there.

I don't think AI watermark is a real problem, as long as it is token-probability based.

 

[flackz]

is it a problem?

Use Ai from other suppliers to post-edit your AI content multiple times here and there.

I don't think AI watermark is a real problem, as long as it is token-probability based.

Exactly. Post-editing or rephrasing can easily break token-probability watermarks. However, images are a different story. Google’s SynthID in Imagen and Nano Banana embed pixel-level patterns that remain detectable even after compression, scaling, cropping, or applying filters. News outlets have confirmed that the image watermark is nearly invisible yet robust, whereas text marks can be eroded by extensive rewriting.

 

[mr Lotus]

SynthID slightly adjusts the probability of choosing words or pixels to create a “hidden watermark” that humans can’t notice, but Google’s detector can identify as AI‑generated content.

 

[flackz]

The key finding here is that Google's AI models should be avoided at any means.

 

[elBlackhatero]

The key finding here is that Google's AI models should be avoided at any means.

I usually use it in moderation, and when I do, I completely rewrite everything with other APIs, and finally run an embedding to compare semantic similarity (for example, cosine similarity). It’s not something that will work in the long run, but for now, it does.