Which AI models watermark text?
AI provenance is becoming more specific, not universal. Some foundation models embed subtle statistical or Unicode markers, while most generated text still carries no public, independently verifiable fingerprint.
Current watermark coverage across major models
This matrix is based on verified public documentation from each provider. Watermarking behavior varies by API configuration, consumer product tier, and post-generation edits.
| Model family | Watermark technology | Current status | How it functions |
|---|---|---|---|
| Google GeminiGoogle SynthID Docs | SynthID Text | Supported in compatible configurations | Detection depends on the watermark configuration and tokenizer used by the generating model. Google’s open-source reference implementation provides cryptographic hashing and sampling-table verification. |
| Anthropic ClaudeAnthropic Announcement | SynthID-based approach | Announced for new & future models | Anthropic announced text watermarking based on Google’s SynthID Text algorithm. Coverage varies across Claude 3.5 Sonnet / Haiku generations and public detector endpoints. |
| ChatGPT / OpenAI TextOpenAI Provenance Update | Invisible Unicode / No statistical mark | No public statistical detector | OpenAI uses cryptographic C2PA for DALL-E images and Sora video, plus subtle Unicode spacing markers in some models (e.g. o3-mini U+202F spaces). There is no official public OpenAI text detector. |
| Open-weight Models (Llama, Mistral)Why status varies | Deployment-dependent | Varies by hosting provider | Hosted API providers (e.g. Together, Groq, Fireworks) can optionally apply statistical watermarks, while self-hosted deployments control their own sampling. No universal mark exists across all instances. |
How statistical text watermarks work
During token generation, an LLM predicts probability scores across thousands of prospective vocabulary words. A watermarking system uses a cryptographic pseudo-random seed to partition words into “green” and “red” buckets, slightly boosting green token probabilities.
In human-authored text, token selections fall evenly between random splits. Across a passage of watermarked text, a detector with access to the secret key can prove that green tokens appear with statistically impossible frequency.
Text watermarks versus file provenance (C2PA)
Text watermarks exist inside the string itself and can survive copy-pasting, but can degrade with significant paraphrasing. In contrast, images and audio typically rely on C2PA Content Credentials—cryptographic manifests signed by certificate authorities that document the device or software model that created the asset.
While C2PA provides verifiable attribution, social media platforms and screenshots often strip file metadata. A thorough verification requires checking both container metadata and steganographic content signals.
Why open-weight model coverage varies
Weights for open models like Llama 3 and Mistral can be downloaded and hosted anywhere. An enterprise hosting provider can enable SynthID or custom token watermarking, whereas a developer running Ollama on their MacBook runs with completely unwatermarked generation parameters.
Frequently asked questions
Does Google SynthID detect any AI-written text?
No. SynthID Text is tied to a specific watermark configuration, pseudo-random key sequence, and tokenizer. Its detector evaluates whether token probabilities conform to the private sampling pattern. Rewriting, translation, or alternate tokenizers reduce detection confidence.
Does a missing watermark mean text is human-written?
No. The vast majority of AI text generators do not embed verifiable watermarks. Furthermore, simple paraphrasing, translation, copying, or using unsupported models will not leave a detectable watermark.
Is a text watermark the same as C2PA Content Credentials?
No. A text watermark is embedded into the statistical distribution of token choices. C2PA is a cryptographically signed provenance manifest attached to a container file (such as JPEG, PNG, or MP4). Copied plain text does not carry C2PA metadata.
How does AIWaterMarkCheck detect invisible watermarks?
Our engine performs multi-layered inspection: 26+ zero-width Unicode steganography characters, special non-breaking spacing patterns (such as ChatGPT o3-mini markers), C2PA manifests, EXIF/XMP metadata, and calibrated AI stylistic features.
Inspect your own text or media files
Scan for C2PA credentials, SynthID signals, invisible Unicode markers, and metadata clues.