Which AI models watermark text?
AI text provenance is becoming more specific, not universal. Some systems can add an invisible statistical watermark; most text still has no public, independently verifiable mark.
Current watermark coverage
This is a research snapshot based on public provider documentation. Coverage changes by model, product, rollout, watermark configuration, and whether the content was rewritten.
| Model family | Watermark | Current status | What that means |
|---|---|---|---|
| Google GeminiGoogle guidance | SynthID Text | Supported for compatible watermarked generation | Detection depends on the watermark configuration and tokenizer used by the generating model. Google’s public reference implementation notes that its hashing does not automatically transfer to Gemini App output. |
| Anthropic ClaudeAnthropic explanation | SynthID-Text-based approach | Announced for new and future models | Anthropic says its approach is based on SynthID Text. Public detection details and coverage may vary by model and rollout. |
| ChatGPT / OpenAI textOpenAI provenance update | No public text detector documented | Do not assume a hidden text watermark | OpenAI’s current provenance work emphasizes image and audio signals. A missing signal still cannot establish that text was written by a person. |
| Open-weight modelsWhy the status varies | Deployment-dependent | Varies by model and operator | A hosted provider can add a watermark, while a local deployment controls its own generation settings. No single watermark can be assumed across every deployment. |
How statistical text watermarks work
A language model usually has several plausible next tokens. A watermarking system uses a private configuration to slightly favor a pseudorandom subset of those choices. Across a long enough passage, a compatible detector can test whether the pattern is more consistent with the configured watermark.
The configuration matters. Google documents keys, n-gram length, sampling-table settings, and tokenizer compatibility as part of the method. That is why an arbitrary AI-text classifier cannot be relabeled as a SynthID detector.
Text watermarks versus image provenance
Text watermarks live in token choices and can survive copy-paste, but they can weaken after heavy editing or translation. Image provenance usually combines pixel-level watermarking with file-level credentials. C2PA credentials can show signed creation and editing history, while metadata may disappear after a screenshot or re-export.
AIWaterMarkCheck checks the evidence that is available in the submitted copy. It does not claim that a missing credential, missing watermark, or low heuristic score proves human authorship.
Why open models vary
Open-weight models can be run by different providers with different samplers, tokenizers, and watermark settings. A model name alone therefore cannot guarantee that every output carries the same watermark. Always ask which exact model, revision, provider, and watermark configuration produced the text.
Frequently asked questions
Does Google SynthID detect any AI-written text?
No. SynthID Text is tied to a compatible watermark configuration and tokenizer. Its detector is probabilistic and can become less confident after rewriting or translation.
Does a missing watermark mean text is human-written?
No. Most generation systems do not expose a public watermark detector, and editing, translation, copying, or unsupported models can remove or avoid the signal.
Is a text watermark the same as C2PA?
No. A text watermark is a statistical pattern in token choices. C2PA is signed provenance attached to a file or asset. C2PA can describe creation and editing history, while copied text usually loses file metadata.
Can AIWaterMarkCheck detect SynthID today?
Not directly. AIWaterMarkCheck currently checks C2PA, readable generator metadata, and replaceable heuristic adapters. It does not claim a generic SynthID detector.
Check the evidence in your own file
Use our checker for C2PA credentials, readable generator metadata, and model-based signals. Results are deliberately labelled as verified provenance, AI-like signals, no supported signals found, or inconclusive.
Open the checker