Editorial
Editorial and corrections policy
This page sets out what we publish, what we refuse to publish, and what happens when we get something wrong. Corrections are logged publicly at the bottom of this page rather than edited away.
It exists because a site with a transactional domain and confident claims about other companies' technology owes readers a way to check both.
Updated 2026-09-05
The writing standard
Every substantive page is built the same way: claim, then the evidence for it, then what the evidence actually means, then what follows for the reader.
- The opening answers the question in the title within the first few sentences. Nuance comes after the answer, not before it.
- Numbers, dates, and quotes are traceable to a cited source. A figure with no source does not ship.
- Quotes are verbatim and attributed to the document they appear in, not to a summary of it.
- Where a study tested a reimplementation rather than a production system, the page says so in the same sentence as the result.
- Pages are written to be useful when skimmed and defensible when read closely.
What we cut: rhetorical filler, invented suspense, and the tic of restating the question three ways before answering it.
AI assistance, disclosed
Research and drafting on this site are AI-assisted. Publishing that fact is not a legal formality here; it is the subject of the site.
What that does and does not mean: models are used to gather sources, draft, and cross-check. They are not used to invent facts, generate sources, or fill a table with plausible-looking numbers. Every claim is checked against the source it cites before publication, and anything that cannot be checked is either labelled unverified or cut.
This is also why the evidence ladder is strict. A generative model is very good at producing a confident sentence about a watermark mechanism that no provider has ever described, which is exactly the failure mode the labelling system is built to catch.
Corrections policy
To report an error, use the contact form: give the URL, quote the sentence you are disputing, and add a source if you have one. Reports that identify a specific sentence get handled fastest. You can leave a reply address or stay anonymous. A correction gets published on its merits either way.
What happens next:
- A factual error is corrected on the page and logged below with the date, the page, what was wrong, and what it now says.
- A significant correction (one that changes the answer a reader would take away) is noted on the page itself, not only in the log.
- An update is different from a correction. When a provider changes its documentation or a new source appears, the page's updated date changes and the log does not, because nothing we published was wrong at the time.
- If we cannot verify a disputed claim either way, the claim is downgraded or removed rather than left standing on our authority.
The line between the two, stated plainly so it can be held against us:
| Change | Logged? | Example |
|---|---|---|
| Material correction | Yes, and noted on the page | A claim was wrong and a reader would have drawn a different conclusion from it. |
| Minor factual fix | Yes, log only | A date, a figure, or a name was wrong without changing the page's answer. |
| Silent update | No, updated date changes | A provider changed its documentation, or a new source appeared. Nothing published was wrong at the time. |
| Copy edit | No | Wording, structure, or layout, with no change to any claim. |
Corrections are never quietly deleted. If a page is merged or retired, its URL redirects and the log entry stays.
Corrections log
This log opens with the publication of this policy on 12 August 2026. Material corrections, the ones that changed the answer a reader would have taken away, are also stated on the page itself, as the policy above promises.
That second promise went unmet until 2026-09-03. The two corrections from August were logged here but never carried a notice on the page they fixed, because no mechanism for one existed. Both have now been annotated retroactively, and every material correction below appears on its own page as well as in this table.
| Date | Page | Kind | What changed |
|---|---|---|---|
| 2026-09-05 | /claude-watermark | Material | Until this date an FAQ on this page said Anthropic had announced a watermark detection API "but has not released it," and that no route to check text existed "for everyone, not just consumers." Anthropic put the API into private preview on 2026-09-01 and annotated its own explainer to say so, so eligible organizations under EU law could already check text while this page still said nobody could. A reader would have concluded that no one, anywhere, could check a text for Claude's watermark. The FAQ now says the API shipped on 2026-09-01, names who can and cannot use it, and separates text (no public checker) from files (a free public checker). Anthropic naming Claude Opus 5 and September 9, 2026 in its 2026-09-04 notice to Claude for Work administrators is a later development and not part of this correction. |
| 2026-09-05 | /claude-text-watermark | Material | This page said in two places that Anthropic's detection API had not shipped: the detection section called it "a first-party commitment with no date, no interface, and no stated access rules", and the lab section listed the official detector as "announced August 14, 2026, not released". Both were wrong from 2026-09-01, when Anthropic put watermark detection into private preview and updated its August 14 post to say so, stating detection is "currently in private preview, available to eligible organizations as required under EU law." A reader would have concluded that no Claude watermark detection existed anywhere, when regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, EU civil society groups and enterprises with their own verification duty could already apply for access. Both passages now say the API is in private preview, limited today to eligible organizations under EU law, with no route for the general public. Anthropic naming Claude Opus 5 and September 9, 2026 in a notice to administrators seen 2026-09-04 is a later development and not part of this correction. |
| 2026-09-05 | /claude-watermark-detector | Material | Two answers on this page still said Anthropic's detection API did not exist: one FAQ ended "What is still missing is the API itself, its access model, and a date", and another said no date had been given for a detector at all. Both were written before 2026-09-01 and left standing after the API shipped, so the page contradicted its own summary. Anthropic's help centre states that watermark detection is "currently in private preview, available to eligible organizations as required under EU law", and its 2026-08-14 explainer carries the note "Updated Sep 1, 2026: Provided up to date information on the watermarking detection API." A reader could have concluded that no detector existed for anyone. Both answers now say the API shipped 2026-09-01 into a private preview for eligible organizations under EU law, and that what is still missing is public access, published admission criteria, and a date for the expansion Anthropic says it plans. Anthropic naming Claude Opus 5 and September 9, 2026 in its 2026-09-04 notice to administrators is a later development and not part of this correction. |
| 2026-09-03 | /ai-watermark-status | Material | The OpenAI audio row read Yes for metadata, on the assumption that audio "got the same treatment" as images. It did not. OpenAI marks audio with SynthID only and attaches no C2PA Content Credentials to it, which three of its own surfaces confirm: the 2026-07-31 update, the help-centre modality table reading "Audio | SynthID watermarks", and the developer guide stating that "images include C2PA and SynthID results, and audio includes a SynthID result." A reader would have gone looking for a manifest that is not there. Metadata corrected to No, and the surface name narrowed to "ChatGPT and API audio output", since OpenAI's cited page never names ChatGPT Voice or GPT-Live. |
| 2026-09-03 | /ai-watermark-status | Material | The Microsoft 365 Copilot row read Yes for a public detector. Microsoft operates none: its page names no detection tool and routes verification through third-party C2PA readers. Corrected to No. Metadata also dropped from Yes to Partial, on Microsoft's own statement that "this additional information is added only to the metadata of images. We're working on getting the information added to video and audio content, but we don't have a specific time frame for when that will happen." A reader would have believed Copilot video and audio carried inspectable provenance today. |
| 2026-09-03 | /ai-watermark-status | Material | The OpenAI Sora video row read Yes for a public detector. No OpenAI provenance surface accepts video at all: the Verify tool and the Content Provenance API both take images and audio only. Corrected to No. This matters more than the other two, because video is the one modality where independent testing contradicted OpenAI's own claim, and it is also the one modality OpenAI gives nobody a way to check. |
| 2026-09-03 | /ai-watermark-status | Minor | The OpenAI images row implied that joining the C2PA steering committee was part of the 2026-05-19 provenance rollout. OpenAI's own page places that in 2024, alongside adding Content Credentials to DALL-E 3, and describes only "C2PA Conforming Generator Product" status as recent. Also noted: the Verify tool is still labelled a preview and has moved to openai.com/research/verify/. |
| 2026-09-03 | /ai-watermark-status | Minor | The Alibaba row read No for a detector. Alibaba Cloud sells detection through AI Guardrails, including implicit-label verification for images and video and a text_aigc_detector classifier for text, so the column is now Partial with the nature of it stated: a paid cloud service, and for text a classifier that guesses rather than a reader that decodes a watermark. The text column stays Unknown, so the row's answer is unchanged. |
| 2026-09-03 | /is-chatgpt-watermarked | Material | Graded Sora output as confirmed for both C2PA metadata and SynthID, contradicting this site's own evidence: the full OpenAI file records a November 2025 audit finding visibly watermarked Sora 2 videos with no detectable C2PA metadata, and unwatermarked Pro-tier videos that carried it, and the provider table already marked Sora video inconsistent. Now split into images and supported audio (confirmed) and Sora video (contested). The same box offered media verification as confirmed without the strippability caveat stated lower on the page, and now says verification holds only on an unmodified original. The shelved text watermark carried three different evidence labels on one page and is now Reported throughout. |
| 2026-09-03 | /deepseek-watermark | Material | Corrected on the since-merged DeepSeek question page, which redirects here as of 2026-09-10. Said DeepSeek's visible label can be removed because it is editable text, omitting the Open Platform Terms of Service clause, reported on two sibling pages here, that forbids maliciously removing, altering, forging or concealing an applied label. A reader would have concluded removal carries no restriction. The prohibition is now stated. The page also said DeepSeek's disclosure page describes two labeling places and nothing more; it describes three, and the notice at the bottom of the interface has been added. |
| 2026-09-03 | /remove-claude-watermark | Material | Called Claude's text watermark shipped rather than researched and graded it confirmed, when Anthropic's own release notes at the time dated every generally available Claude model earlier than its 2 August 2026 marking threshold. The Claude tracker said announced and in transition, and this page's own figure marked the same claim conditional, so the page contradicted both its sibling and itself. Now states that the rollout to shipped models is in progress. Anthropic named Fable 5.1 and Mythos 5.1 as marked at launch on 1 September 2026, which is a later development and not part of this correction. |
| 2026-09-03 | /claude-watermark-detector | Material | Listed files whose metadata was stripped by conversion, re-saving or a screenshot among the conditions under which Claude's text watermark may be absent. That is a C2PA file-metadata condition, and including it contradicted the page's own central point, quoted from Anthropic, that nothing is added to the text and there are no hidden characters. A reader could have concluded that re-saving a document strips the text mark. Now stated separately as the provenance layer it belongs to. A passage crediting an unnamed scanner for checking 34-plus invisible characters and disclaiming cryptographic watermarks read as this site's own count and vocabulary; it now names getgpt.app and notes that Claude's mark is statistical. |
| 2026-09-03 | /is-gemini-watermarked | Material | Called Google's consumer verification tool the confirmed, working path for images, video and audio, where an independent check in November 2025, cited on the full Gemini file, found it reliable for images but hit-or-miss for audio and video. Now says so. The page also described the dispute over Gemini text as two Google sources when there are three, omitting Google's own help page for verifying AI content, which covers images, video and audio and stays silent on text; the third source is now named. A robustness figure attributed the 98.3% signal loss to aggressive paraphrasing where the source says automated. |
| 2026-09-03 | /claude-watermark-remover | Minor | Described the invisible-character checker as scanning five kinds of character, and a next step on /remove-claude-watermark described it as finding eight. It has checked 50 since the cleaner was widened on 2026-08-17. Both places now read the count from the code rather than restating it, so they cannot drift again. The page also now says that characters doing real typographic or scripting work are flagged rather than deleted, which the tool has always done and the page did not mention. |
| 2026-09-03 | /anthropic-watermark | Minor | The provenance timeline dated the news cycle 2026-08-11 and called it nine days after the feature went live. The Claude tracker dates the first posts to the evening of 2026-08-10 and both cited sources carry that date, making the interval eight days. Timeline corrected. |
| 2026-09-03 | /claude-watermark | Minor | Stated flatly that Anthropic's Transparency Hub was last updated 23 July 2026. The date the Hub displays is a fact, but whether its wording changed then could not be verified, because a Wayback Machine check failed, as the company page already recorded. Now phrased as the last-updated date the Hub shows. |
| 2026-09-03 | /remove-claude-watermark | Minor | The GPTZero, Originality.ai and ZeroGPT homepages were typed as Community sources here and as Official on four other pages. They are first-party product documentation, so Official is correct and this page now matches. ZeroGPT was also cited over http in two places and is now https. |
| 2026-08-13 | /is-grok-watermarked | Material | Understated what xAI documents. The page treated a single Acceptable Use Policy clause as the only evidence and concluded that nothing said how Grok marks anything. It missed xAI's own Grok FAQ, which states plainly that generated images and videos carry a Grok watermark and that no setting removes it. Corrected to a per-surface answer: images and video confirmed visibly marked, text not documented. The page now also states that the AUP clause prohibits users from stripping provenance metadata and does not establish that xAI embeds any. |
| 2026-08-13 | /machine-readable-ai-marking | Material | Said there was no confirmed official EU icon and that Commission texts contain no icon specifications of any kind. Both were wrong. The Commission publishes three official icons in four variations, free to use, with placement rules in Section 2 of the Code of Practice. Corrected, with the scope made explicit: the icons serve the Article 50(4) deployer disclosure duty, not the Article 50(2) machine-readable marking duty this page covers, and their use remains optional. The separate rumour about a mandated 24x24px size, PMS 286C blue, and a required C2PA pathway is still unsupported and stays flagged. |
The check that runs before publication
Every page clears this list before it goes live, and again whenever its central claim is revised.
- Every factual claim traces to a source in the list at the bottom of the page.
- Each source has been opened, not just cited. Where a live page blocks automated reading, an archive capture is used and labelled as one.
- The page's evidence label reflects the weakest source supporting its central claim, not the strongest.
- Anything a provider has not said is described as undocumented rather than denied.
- Any number carries its sample size, its date, and what it cannot show.
- Removal and detection claims are checked against the rule that nothing unverifiable is presented as verified.
- Pages that overlap this one link to it and say how they differ.
- Dates at the top reflect what actually happened: published, revised, and where relevant, sources last reopened.
Links, citations, and reuse
External links are editorial. Nobody pays for a link on this site, and no link here is affiliate-tagged.
Sources are listed at the bottom of every page with their status label, so a reader can go to the primary document rather than take our summary of it.
If you want to cite or reuse a diagram, chart, or dataset from this site, that is welcome. A link back to the page it came from is the only thing asked for. If you are reproducing a finding, please also reproduce its sample size and limitations, since those are load-bearing.