Editorial standards
Methodology
This page exists so the claims elsewhere on the site can be checked. It covers who publishes Simple Unmark, how provider claims are sourced and dated, what we test, and what we cannot measure and therefore do not claim.
- Last updated
- Provider status verified
- Published pages
- Tools: 6 · Guides: 9
Who publishes this
Simple Unmark is an independent tool for removing invisible Unicode characters and rewriting text that may carry a statistical watermark. It currently handles pasted text only. It is operated as a small commercial product, funded by credit purchases rather than by advertising or affiliate links to the detectors and providers it writes about.
We have no commercial relationship with Google, OpenAI or Anthropic, and no access to any provider's private watermark keys or detectors. Where that limits what we can prove, the limitation is stated on the page rather than papered over.
How provider claims are sourced
Every provider status claim is held in one registry that feeds the provider pages, the status tracker and this page. A claim carries a confidence level and the date a person last re-read its sources.
- "Documented" requires a first-party page from the provider itself. Press coverage never promotes a claim to documented, however extensive.
- "Reported" means credible independent coverage exists without first-party confirmation, and the page says so in those words.
- "Unverified" means neither, and we would rather publish that than imply certainty.
- Provider status is reviewed monthly, and within 72 hours of a major announcement.
- When a provider's documentation contradicts something we published, we change the page and the date rather than quietly leaving both up.
What we test on the deterministic layer
The character layer is exactly testable, so it is where we make exact claims. The full set of code points we remove or normalise is published as a coverage table on the invisible character remover. The browser-side TypeScript scanner and isolated Python workload implement the same published classes, with matching fixtures in both test suites so cross-language drift is caught before release.
That layer is covered by unit tests, including cases that must not change: zero-width joiners and non-joiners are preserved in valid emoji and same-script contexts, because deleting them there corrupts text in scripts such as Arabic, Persian, Hindi and Malayalam. Floating or malformed joiners are removed.
How worked examples are made
Each tool page shows one worked example, and its note says which kind it is. A hidden-character example shows what the deterministic cleaner produces for that input. A rewrite example is an illustrative edit written by us to show how much a thorough rewrite changes. It is not saved model output, and it has not been run through any watermark detector or AI detector.
What we cannot measure on the statistical layer
We cannot measure watermark removal, and we will not publish a number that implies we can. Detecting a statistical text watermark requires the provider's key. Google publishes its approach and a reference implementation but no general detector for arbitrary text. Anthropic offers restricted detection API access in private preview for eligible organizations; Simple Unmark does not have access. OpenAI's textGrain detector, documented on 5 October 2026, is limited to approved researchers and expert organizations. None of those is a public check we can run.
So our claim is deliberately narrow. A clean rewrites the word choices that carry the signal, which is the change the providers themselves describe as weakening detection. We do not claim a score, a percentage, or that any text is undetectable.
- We do not claim guaranteed removal, or that output is undetectable.
- We do not present light paraphrasing as proven watermark removal.
- We do not claim to defeat named detectors or institutional policies.
- We do not describe rewritten output as human-written.
- We do not present invisible characters as proof that text is AI-generated.
- We do not publish benchmark numbers we cannot reproduce from a stated method.
What we would publish with detector access
If we receive authorized access to a provider's text watermark detector, we intend to test the rewrite against it. We would publish the method and results, including failures. The proposed test below is specified in advance so the result cannot be framed after the fact.
- A fixture set of provider-generated passages across lengths and genres, released publicly so anyone can rerun the test.
- Detector score before and after cleaning, reported as a distribution rather than a single figure.
- Semantic similarity and factual-consistency checks on the same pairs, because a rewrite that loses meaning is a failure even if the score drops.
- Failure cases and the conditions that produce them, including short and highly factual text where we expect the weakest result.
- Latency, token use and cost per request, so the operating claims are as checkable as the quality ones.
Privacy in the measurement itself
Analytics never receives submitted content, cleaned output, filenames, or any fragment of them. Product events record a landing route, tool or page category, coarse word-count band, authentication and privacy mode, and action or result category. Purchase and signup identifiers are used only for local verification or deduplication, never as analytics properties.
Request records in our own database store the content category, selected privacy mode, counts or byte size, status, timings, token use, cost and credit activity rather than content. After a clean, an optional reply can store a rating, selected issues, and a short comment linked to that request. The comment is only what the person writes in the feedback box. The cleaned output is stored with that reply only when they choose to share it. The submitted text is not stored. Private text mode sends plaintext through the application to the Python workload and DeepInfra. Confidential text makes the browser verify the separately deployed Confidential Space workload, encrypt to its attested request-scoped key, and decrypt the response locally; the application receives a signed metadata-only receipt. DeepInfra processes the text rewrite.
Confidential AI is the planned third mode, with the model inside an attested confidential GPU and no external model API receiving plaintext. Selecting it while it is unavailable stores a pseudonymous content-free interest record so we can compare demand with actual Private and Confidential usage. The deterministic inspector and free text converters process your content entirely in your browser. Content-free analytics can record tool use and copying; the entered text is never uploaded. Full trust boundaries are in the privacy policy.
Corrections
If something here is wrong, we would rather fix it than defend it. Corrections to technical or provider claims go to privacy [at] simpleunmark.com, and a material correction updates the page's visible date. We do not bump dates for cosmetic edits, because a date that moves for no reason is worth nothing to a reader.
Current coverage at a glance
- Reference rows
- 61 named, plus documented Unicode ranges
- Providers tracked
- 3, each with a verification date
- Measured removal claims
- None — no public detector exists to measure against
Every primary reference used on this site
- GPT detectors are biased against non-native English writers
Liang et al. (2023)
- SynthID overview
Google DeepMind · First-party source
- SynthID text watermarking and detection
Google AI for Developers · First-party source
- Scalable watermarking for identifying large language model outputs
Nature (2024)
- How Claude's text watermarking works
Anthropic · First-party source
- Transparency Hub
Anthropic · First-party source
- Advancing content provenance
OpenAI · First-party source
- Content provenance guide
OpenAI Platform docs · First-party source
- Our approach to EU text provenance rules
OpenAI (5 October 2026) · First-party source
- textGrain: Entropy-Calibrated Watermarking for Language Model Text
OpenAI technical report (5 October 2026) · First-party source
- How Claude marks AI-generated content
Anthropic · First-party source
- Code of Practice on Transparency of AI-Generated Content
European Commission · First-party source
- Anthropic posts 'How Claude Marks AI-Generated Content' without explaining how
John Gruber, Daring Fireball (11 August 2026)
- Anthropic's 'watermark' text adulteration in Claude is a perversion of writing
John Gruber, Daring Fireball (16 August 2026)
- Follow-up thoughts on watermarking schemes for AI-generated text
John Gruber, Daring Fireball (17 August 2026)
- Anthropic's watermarking, how it (probably) works, worse than it seems
Ben Thompson, Stratechery (12 August 2026, subscribers)
- Tech analyst Ben Thompson dismisses the 'clearly absurd' concept embedded in AI watermarking
Business Insider (14 August 2026)
- Anthropic's weak watermarks appease a weak law
James Padolsey
- LLMs aren't writing
Dan Moren, Six Colors
- UAX #9: Unicode Bidirectional Algorithm
Unicode Consortium · First-party source
- UAX #15: Unicode Normalization Forms
Unicode Consortium · First-party source
- UTR #36: Unicode Security Considerations
Unicode Consortium · First-party source
- watermarks-remover file cleaner documentation
watermarks-remover · First-party source
Competitor pages are not cited as technical sources. They are evidence of how a market positions itself, not of how watermarking works. See the provider status tracker for the current dated claims.
