Google DeepMind is widening access to SynthID, a tool designed to help identify AI-generated text. The move matters because watermarking has become one of the main technical ideas for making synthetic media easier to recognize, especially as generative AI systems produce more text, images, and video.
The company has already used the same broader SynthID family for other kinds of AI output. It unveiled a watermark for images last year, later rolled one out for AI-generated video, and in May announced that SynthID was being applied in its Gemini app and online chatbots. It also made the technology freely available on Hugging Face, an open repository of AI data sets and models.
What Google DeepMind is releasing
SynthID is a watermarking system for text generated by AI models. Google DeepMind has developed it to add an invisible signal during generation, so later systems can help assess whether a passage came from a model using that watermark.
The open-source release is intended to let other generative AI developers use the technology with their own large language models. Pushmeet Kohli, the vice president of research at Google DeepMind, frames the release as a way for more developers to build AI systems responsibly by detecting whether text outputs came from their own models.
The immediate promise is not that every AI-written sentence on the internet becomes identifiable. The source makes clear that SynthID for text currently works only on content generated by Google’s models. The hope is that open-sourcing it will expand the range of tools it can work with.
That distinction is important. A watermarking tool is most useful when it is actually built into the generation process. Open sourcing can make adoption easier, but the watermark still depends on developers choosing to incorporate it into their systems.
How SynthID marks AI text
SynthID works inside the text generation process rather than after the fact. Large language models break language into “tokens,” then predict which token is most likely to come next. A token can be a single character, a word, or part of a phrase.
For each possible next token, the model assigns a percentage score that reflects how likely it is to fit. A higher percentage means the model is more likely to choose that token. SynthID adds its watermark by adjusting those probabilities as the text is generated.
In plain terms, the tool nudges the model’s token choices in a way that should not be visible to the reader. Later, to detect the watermark, SynthID compares the expected probability patterns in watermarked and unwatermarked text.
This is why the watermark is described as invisible. It is not a label, a tag, or a visible mark in the text. It is a statistical pattern introduced during generation and checked later by a detector.
What the Gemini experiment found
Google DeepMind says it tested SynthID at large scale after deploying the watermark in Gemini products. Gemini allows users to judge chatbot responses with a thumbs-up or thumbs-down, giving the company a way to compare how people rated watermarked and unwatermarked outputs.
Kohli and his team analyzed scores for around 20 million watermarked and unwatermarked chatbot responses. According to the source article, users did not notice a difference in quality and usefulness between the two groups.
Google DeepMind also found that the SynthID watermark did not compromise generated text in several areas:
- quality
- accuracy
- creativity
- speed
The results of the experiment are detailed in a paper published in Nature today. For developers and platforms considering AI text watermarking, the key point is that Google DeepMind’s test suggests this type of watermark can be added without making the model’s answers feel worse to users, at least in the Gemini setting described in the source.
That is central to adoption. A watermark that noticeably damages output quality would face resistance from users and developers. The Gemini experiment gives Google DeepMind a public argument that SynthID can be useful without creating an obvious trade-off in everyday chatbot responses.
Where watermarking still falls short
SynthID is not presented as a complete answer to AI detection. The source article identifies several limits. The watermark resisted some tampering, including cropping text and light editing or rewriting, but it was less reliable when AI-generated text had been rewritten or translated from one language into another.
It was also less reliable for prompts asking for factual information, such as the capital city of France. The reason is straightforward: when there are fewer valid ways to answer without changing the facts, there are fewer opportunities for the system to adjust the likelihood of the next word.
Soheil Feizi, an associate professor at the University of Maryland who has studied vulnerabilities in AI watermarking, says reliable and imperceptible watermarking of AI-generated text is fundamentally challenging. He points especially to cases where large language model outputs are near deterministic, including factual questions and code generation tasks.
These limits do not make watermarking useless. They do mean it should be understood as one safeguard among several. A detector may perform better in some settings than others, and transformations such as translation or heavy rewriting can weaken reliability.
Why open source changes the conversation
Feizi describes Google DeepMind’s decision to open-source the method as a positive step for the AI community. His reasoning is that the community can test the detectors, evaluate their robustness in different settings, and better understand where these techniques work or fail.
João Gante, a machine-learning engineer at Hugging Face, points to another practical benefit. Open-sourcing the tool means anyone can take the code and incorporate watermarking into a model with no strings attached. He also says this can improve privacy because only the model owner will know its cryptographic secrets.
Gante says he wants to believe that better accessibility and the ability to confirm the tool’s capabilities will help watermarking become the standard, which could support detection of malicious language model use.
Irene Solaiman, Hugging Face’s head of global policy, adds a caution. Watermarking is one part of a safer model ecosystem, not a standalone fix. She compares the broader challenge to human-generated content, where fact-checking has varying effectiveness.
That is the practical takeaway. SynthID’s open-source release could make AI text watermarking easier for developers to test and adopt. But the source makes equally clear that watermarking works best as part of a broader set of safeguards, especially in a world where AI-generated text can be edited, translated, rewritten, or used in contexts where detection is technically harder.