SynthID Text
Google DeepMind reference code for watermarking text from language models and detecting that watermark.
Reference code from a Nature paper, with frequentist and Bayesian watermark detectors.
SynthID Text is Google DeepMind's reference implementation of its text watermarking and detection method. It suits researchers who generate text with their own models and want to verify it later, not people checking unknown text.
What it is
A Jupyter notebook extends Hugging Face Gemma 3 and GPT-2 models so they watermark text as they generate, then detects the watermark with a frequentist detector (an exact binomial p-value) or a trainable Bayesian detector.
Who it's for
It suits researchers and developers who control the generating model. It cannot tell you whether arbitrary text came from ChatGPT or Claude, because it detects only text carrying its own watermark, and the repo says it is not for production.
Key features
Watermark generation mix-in
Adds watermarking to Gemma 3 and GPT-2 models in PyTorch.
Frequentist detector
Needs no training and returns an exact binomial p-value for a given key.
Bayesian detector
A trainable detector the README calls more powerful, scored with JAX on CPU.
Self-contained notebook
Runs in Colab or locally with Gemma 3 270M IT or GPT-2.
PyPI package
The core library is distributed on PyPI for the notebook.
Nature paper reference
Implements the method from the published Nature paper.
Up and running in 3 steps
Our honest verdict
What's great
- Detection comes with a statistical p-value rather than a vague percentage.
- Runs on a T4 GPU for Gemma 3 270M, or on CPU with GPT-2.
- From the group that published the method, with the paper to cite.
- Apache-2.0 licence.
Where it falls short
- Detects only watermarked text from a model you configured; it cannot flag ordinary chatbot output.
- The repo says it is for reference and research reproducibility, not production.
- Its Threefry-based hashing is not a vetted cryptographic primitive and keys are limited to 63 bits.
- You need Python, a virtual environment and a Jupyter notebook.
Pricing & license
SynthID Text is free under the Apache-2.0 licence. The repo shows no paid plan; the only cost is running the notebook yourself, for example on a free Colab runtime.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can SynthID Text detect ChatGPT writing?
No. It detects only text generated with its own watermark under a key you hold.
Is it ready for production?
The README says the implementation is for reference and research reproducibility and is not designed for production systems.
What hardware does it need?
The README says any GPU such as a T4 works for Gemma 3 270M IT, and GPT-2 runs on any runtime, faster on a high-RAM CPU or GPU.
