AI Watermark Remover — Claude, ChatGPT and Gemini Compared
A text editor can remove selected Unicode characters and help you review wording. That does not make it a universal provenance checker. This tool handles pasted text only, not images, audio, video or file metadata. It does not calculate classifier scores.
Use the Claude text editing tool for local character operations, or read what can and cannot be verified before interpreting a report.
Three providers, different deployment boundaries
| Provider | Text mechanism | Practical boundary |
|---|---|---|
| Claude | A version of SynthID-Text: a statistical pattern in generated word choices. | Supported models carry the mark. Official detection access is in private preview for eligible organizations. |
| ChatGPT / OpenAI | textGrain adds a statistical signal to model word choices. | OpenAI's October 5 announcement describes opt-in watermarking for select API models and a phased EU ChatGPT rollout. It is not a global text-marking default. |
| Gemini | SynthID modifies token-selection probabilities to carry a signal in generated text. | Google documents text marking in the Gemini app and web experience. Its media verification instructions do not establish a public pasted-text verification service. |
Claude: a pattern in the words
Anthropic's technical explanation ties the signal to the randomness used when choosing words. This is different from counting invisible Unicode characters. A character operation report is reproducible, but it does not read Claude's statistical signal. The provider's key and a suitable detector are needed for that separate task. Anthropic’s technical explanation ↗
ChatGPT: do not generalize one rollout to every user
OpenAI's October 5, 2026 announcement names textGrain and separates API opt-in from eligible ChatGPT output in the EU. Its text detector initially requires approval. A statement about that regional rollout should not be turned into a claim that every American user's ChatGPT passage is marked. OpenAI's public verification page currently accepts image and audio files; that is a separate capability. OpenAI’s text provenance announcement ↗
Gemini: the text and media versions are separate
Google describes SynthID for text as adjusting the choices made during generation. It also documents watermarking and verification for images, audio and video. Sharing the SynthID name does not make those workflows interchangeable. This text-only editor neither uploads a picture to Gemini nor tests a Gemini text watermark. Google DeepMind’s SynthID documentation ↗
Choose the task you can actually check
For pasted text, compare the selected character edits with your original. If you need new wording, review the revised passage for changed facts, qualifications and emphasis. A readable revision is an editing outcome, not proof about which model contributed to it. File provenance belongs in a supported file-checking workflow, outside this editor.
Disclosure rules still apply after editing. Keep your original and a record of the assistance you used when your school, employer or publication requires that history.