Article
The Complete Guide to Removing AI Watermarks from Text, Images & Files
Which AI models actually watermark text, what the watermark really is, and how to remove it.
An AI text watermark is a statistical pattern embedded in a model's word choices, and the only reliable way to remove it is a complete rewrite through a model that does not watermark. As of 2026, only Google Gemini watermarks text in production; Claude has announced watermarking for its newest models, while ChatGPT and open-weight models like DeepSeek and Llama do not watermark text.
That last point is the key to removal: because open-weight models carry no watermark of their own, you can rewrite watermarked text through one of them to break the pattern without stamping a new one. Here is the accurate picture of who watermarks, what the mark actually is, and how to clean it.
Who watermarks text in 2026
What an AI text watermark actually is
A text watermark is a statistical bias in word choice, most commonly the SynthID-Text method from Google DeepMind. When the model picks among plausible next words, a secret key nudges the choice, and across a long passage the bias becomes detectable to whoever holds the key. It is not a hidden Unicode character and not file metadata. Two related things often get confused with it: stray invisible characters that come from copying out of a chat interface, and C2PA provenance metadata on image files. Both are worth cleaning, but neither is the text watermark.
Which providers watermark text
Google Gemini — live
Gemini has applied SynthID-Text watermarking to its output since 2024, and Google offers SynthID detection. This is the one production text watermark in wide use today.
Anthropic Claude — announced
Anthropic announced text watermarking in August 2026 using a SynthID-Text approach, applied to models launched after its cutoff. Current models launched before it and are not yet watermarked, with retrofits planned. No public detector has shipped.
OpenAI ChatGPT — not yet
OpenAI has discussed provenance signals for text but has not shipped a text watermark. ChatGPT output today carries no statistical text watermark.
Open-weight models (DeepSeek, Llama, Qwen, Mistral) — none
Self-hosted and open-weight models do not apply a provider watermark to their text. This is precisely why they are useful as a rewriting engine: they can break another model's watermark without adding one of their own.
How to remove an AI text watermark
Step 1: Rewrite through a non-watermarking model
This is the step that removes the watermark. The pattern is in the words, so you have to change enough of them. ClaudeRewrite rewrites the wording through DeepSeek V4 Flash, preserving every fact, number, name, quote, and code block, and verifies meaning with an automatic check. Do not rewrite with the same model that wrote the text (it can re-apply its mark), and do not rewrite with Gemini (it watermarks its own output).
Step 2: Strip stray invisible characters
Deterministic and lossless: remove zero-width spaces, joiners, and smart-quote artifacts left over from a chat interface. These are not the watermark, but they are common AI tells, so ClaudeRewrite clears them in the same pass.
Step 3: Strip file metadata
For images and documents, remove C2PA, EXIF, XMP, and IPTC metadata. ClaudeRewrite handles this for JPG, PNG, WebP, GIF, and TIFF.
Why manual editing often fails
Hand-editing rarely removes a text watermark, because each word's bias is seeded by the words before it, so light synonym-swapping leaves most of the pattern intact. The watermark is designed to survive light editing. A full rewrite is what changes enough of the words to break it.
An honest limit
No tool can guarantee your text passes every AI detector: Anthropic and Google keep their detectors private, and style-based detectors are inconsistent and prone to false positives. A rewrite through a non-watermarking model reliably breaks the statistical watermark itself, which is the strongest available removal, with that residual uncertainty stated plainly.
References
Get started
Ready to remove Claude's watermark?
Try ClaudeRewrite free — 5 rewrites per month, no card required.
Try it free