Simple Unmark guide
How to Remove AI Watermarks From Text
- Published
- Last updated
- Sources checked
- Primary references: 4
Short answer
First work out which mechanism applies. Hidden characters can be found and deleted exactly with a scanner. A statistical watermark lives in word choices, so the wording has to be rewritten. Google and Anthropic say light edits are not enough; OpenAI's textGrain test shows synonym replacement cuts detection sharply. Then check the facts. We cannot independently confirm the remaining watermark signal.
Hidden characters and statistical watermarks need different treatment, so a complete pass handles them one after the other. You can do every step below by hand; a tool only makes the first one exact and the second one faster.
Choose the privacy boundary before every rewrite.
Private is the lowest-cost path and stores no text in our application database. Confidential verifies Google Confidential Space in your browser and encrypts to a key the workload created for that one request, so our web backend receives no text — though DeepInfra still performs the rewrite. Confidential AI, with the model inside a confidential GPU, is coming soon.
Compare the three modes and their limits1. Identify the mechanism
Start here, because it determines whether the job is exact or probabilistic. Text from a provider that documents a token-choice watermark — Gemini, Claude since August 2026, and eligible ChatGPT and Codex text in the EU under OpenAI's textGrain — carries a statistical signal and no added characters. Text from a provider with no documented text watermark may still carry stray invisible characters from the copy path.
Our provider status tracker records what each provider documents, with sources and a verification date, so you are not guessing.
2. Preserve the original
Keep an untouched copy before you change anything. You will need it to compare meaning, numbers, names, formatting and citations afterwards, because a rewrite is exactly the kind of edit that can alter a figure or drop a hedge without anyone noticing.
3. Run deterministic cleanup
Scan for zero-width spaces, direction controls, soft hyphens, tag characters and unusual Unicode spaces, and remove them. This part is exact: a good tool tells you which code points it found and how many.
Be conservative. Preserve zero-width joiners and non-joiners in valid emoji and same-script contexts, because they are essential to scripts including Arabic, Persian, Hindi and Malayalam. Do not "normalise" every non-ASCII character, which mangles legitimate text.
4. Rewrite the statistical layer properly
Change syntax, sentence boundaries, clause order, transitions and vocabulary across the whole passage. Google describes SynthID as robust to modifying a few words and mild paraphrasing, and Anthropic contrasts light editing, which probably will not remove Claude's mark, with a complete rewrite, which will. OpenAI's textGrain numbers show detection falling under lighter edits: replacing 25% of words with synonyms cut it from about 92% to 17% on 400-token passages. Aim for the stricter standard, a complete rewrite, because you cannot test which case applies.
Leave direct quotations, code and fixed terms as they are. They have to stay exact, and they carry less watermark anyway. Short passages give a rewrite little room, which cuts both ways, because short passages are also harder to detect.
Example: a light edit versus a rewrite
An illustrative editing example written by us. It shows the scale of change, and it has not been run through any watermark detector.
- Original: "Our team reviewed the proposal carefully and identified three main risks: budget overruns, unclear ownership and a tight timeline."
- Light edit: "Our team reviewed the proposal closely and found three main risks: budget overruns, unclear ownership and a tight timeline." Two words changed. Most of the original word choices, and so most of any watermark evidence, are still there.
- Rewrite: "After a close read of the proposal, we see three risks. Costs could run over, nobody clearly owns the work, and the schedule leaves little slack." Same three risks, new structure, almost no original wording left.
5. Verify, and calibrate your claims
Compare every factual claim, number, name and quotation against the original. For high-stakes content, have a qualified human review it.
Then be realistic about what you have done. We do not have access to the providers’ text watermark detectors, so we cannot measure what signal is left. You have changed the wording the signal depends on; you have not proven the signal is gone. An AI detector score is not proof of authorship in either direction.
Sources
- SynthID text watermarking and detection
Google AI for Developers · First-party source
- How Claude's text watermarking works
Anthropic · First-party source
- Our approach to EU text provenance rules
OpenAI (5 October 2026) · First-party source
- UAX #15: Unicode Normalization Forms
Unicode Consortium · First-party source
Published by Simple Unmark. Our Methodology explains how we source claims, what we test, and what we do not claim.
Apply this to your own text.
Guests get three free cleans without an account. Choose Private or browser-verified Confidential processing before each rewrite. Deterministic character inspection is free and runs only in your browser.
Keep reading
- How AI Text Watermarks WorkHow hidden-character marks and statistical token watermarks differ, how detection works, and why removal claims should always be bounded.
- What Is SynthID Text?SynthID Text explained: how Google's text watermark is applied during generation, how detection works, what weakens it, and how it differs from image watermarking.
- Hidden Unicode vs Statistical Text WatermarksThe two mechanisms confused most often: invisible Unicode characters versus keyed token-choice watermarks. What each one is, what removes it, and what neither proves.
- AI Watermarks vs AI Detectors vs Provenance MetadataThree different things routinely confused: keyed watermarks added at generation, statistical AI detectors that guess, and C2PA metadata attached to files.
