My original X Article about Claude's invisible watermark reached 64.7K views before I prepared this update.5 That reach gave the argument time to breathe. Anthropic published its FAQ on Friday; it added a corporate account of the planned system, not a final answer to the argument.
Anthropic describes the planned system in its FAQ. Independent testing still needs to establish how it performs. The company says a signal of Claude involvement cannot prove that Claude authored the work. Any institution using a detector would need to keep that distinction intact.
The first argument was about authorship and control
My concern was never that AI use should remain hidden. McQueen Analytics asked 2,400 adults whether people should be told when AI played a major role in creating something. In that sample, 83.1 percent agreed. The same survey found 87.1 percent agreement that people should have a human choice for important decisions, and 77.1 percent selected real choice about when to use AI as something they wanted.1
Those findings belong together. People want disclosure, but they also want a clear account of what the person and the model each did.
Our trust work has never treated an official statement as proof. A company's explanation is useful because it records what the company is promising. Trust comes from whether people can verify the promise, see the same rule applied in practice and correct a failure when the system gets it wrong.
Before the FAQ, a detector result could easily be overstated as an authorship verdict. Consider an essay a student drafted before asking Claude for a grammar pass, a researcher's translation or a programmer's request to reformat code and comments. Claude can process each one while the ideas, judgment and responsibility remain human.
I wrote the X Article to keep that line visible. A machine-readable mark can serve transparency only if institutions keep it within the evidence it provides.
What Anthropic said on Friday
Anthropic published its FAQ on August 14. The company says future Claude models will use a version of Google DeepMind's SynthID-Text method to comply with the EU AI Act. It describes a keyed pattern used when the model chooses among words that would all fit, rather than hidden characters added to the text. Anthropic says the pattern does not add tokens, raise the price or carry information about a person, organization or chat.2
Anthropic represents that ownership, authorship, legal responsibility and a user's rights under its terms remain unchanged. Independent end-to-end results are unavailable because the public does not yet have the future models or detection API.
The FAQ's stated limits give the public a list of claims to test later.
Anthropic says a watermark can answer only how likely it is that Claude was involved in writing part of a passage. Short and factual text may not carry enough choices to detect. Light proofreading may leave little or no detectable signal because Claude changes so few words. Translation is different: Claude chooses every word in the translated text, so the translation carries the mark.
The company plans to offer a detection API, but it is still working out the implementation. Even then, Anthropic says the result will not distinguish "Claude wrote this" from "Claude heavily edited this." The language is narrower than the shorthand likely to appear in a school, workplace or client dispute.
The conversation after the FAQ
Anthropic's X post linking the FAQ had more than 11 million views when I checked Tuesday morning, along with about 4,700 likes, 1,800 reposts and 1,900 replies.3 The post put more precise claims into the public conversation. Its reach is evidence that Anthropic entered the conversation at scale, not that the audience trusted the explanation.
A post criticizing the policy's treatment of translations and the planned detector showed 2.3 million views. An interactive explanation of the token-selection method showed 1.7 million. A teacher welcomed a way to identify generated student work, while other people asked what happens to a resume, paper or client document that a person wrote and Claude only helped edit.
Other posts corrected claims that Anthropic was inserting hidden characters or a personal tracking code. At the same time, people promoted supposed watermark removers even though Anthropic has not released the detection API needed to test those promises.
By Tuesday morning, an expanded signed-in X review covered 264 unique posts from 227 accounts. Every status URL resolved when checked independently. The review surfaced support for provenance, objections to invisible marking, technical explainers, classroom arguments and remover pitches. That inventory shows a continuing conversation, not a percentage or an X-wide measure of opinion.4
The response did not end the argument. People kept returning to the difference between model involvement and human authorship. Nothing in that inventory shows that the FAQ restored trust.
The FAQ did not settle the trust question
Anthropic's FAQ makes several claims:
- The text mark is a statistical pattern, not hidden characters.
- It cannot be traced to a person, organization or conversation.
- It does not add tokens or change the price.
- Light proofreading may not leave enough signal to detect.
- Translation is marked because Claude chooses the translated words.
- A public detection API is still coming.
- A detected mark does not settle authorship, ownership or responsibility.
That gives the public clearer claims to test. Until the models and detector are available, these are still Anthropic's claims about a system the public cannot inspect end to end.
What another institution does with the result remains a risk. A detector can report likely Claude involvement. A school, employer, platform or client can still turn that narrow finding into an accusation that Claude wrote the work. Anthropic says its signal does not support that conclusion.
Any policy using the detector needs the tested passage, the confidence level, independent error testing, the limits for short or edited text and a way for the affected person to explain how Claude was used. A positive result cannot be an automatic verdict.
The FAQ may be a necessary response. Repairing trust will require independent performance tests, limits on how a detector is used and recourse for anyone wrongly accused.
Processing is still not authorship
Giving the X Article time to breathe brought a detailed response from Anthropic and a fuller view of the reaction. Some early claims concerned hidden characters or personal tracking. Anthropic says it is not using either; that clarification is useful but remains a company claim until the product can be tested.
My position remains that useful disclosure has to preserve the fact that processing is not authorship.
Anthropic draws that boundary in the FAQ: the watermark estimates Claude involvement but cannot decide authorship. Trust will depend on whether the product, detector and institutions using it keep that limit in practice.
Other reads on Claude's watermark
- Read the original X Article for the argument as it began and the McQueen Analytics survey results behind it.
- Anthropic's August 14 FAQ explains the planned watermark, its limits, proofreading, translation, detection and ownership.
- Anthropic's X response shows how the company summarized the FAQ and how people answered it.
Source notes
- McQueen Analytics AI Trust Survey 1 included 2,400 completed responses among adults ages 18 to 64. The results were 83.1 percent agreement that people should be told when AI played a major role, 87.1 percent agreement that people should have a human choice for important decisions and 77.1 percent selecting real choice about when to use AI. These sample estimates were not weighted to represent the U.S. population.
- Anthropic, How Claude's text watermark works, August 14, 2026, checked August 18. The note records Anthropic's description of its own planned system, not an independent performance test. Anthropic says it uses a version of SynthID-Text; the pattern adds no hidden characters, tokens or identifying information. The FAQ says light proofreading may leave too little signal to detect, translations are marked, a detection API is forthcoming, and a mark cannot distinguish Claude writing from heavy Claude editing. It also says the mark does not decide ownership, authorship or legal responsibility. The future models and detection API were not publicly available for an end-to-end check at capture.
- Anthropic's August 14 X post linking the FAQ, checked in a signed-in X session on August 18 at about 9:30 a.m. EDT. The visible counts were 11,018,463 views, 4,732 likes, 1,832 reposts, 1,938 replies and 3,289 bookmarks. Counts change, and X views are not unique people.
- Expanded signed-in X reaction review, August 18, 2026. Multiple search families covered direct Claude and Anthropic watermark terms, SynthID and detector terms, proofreading and translation, remover claims, the Anthropic FAQ URL, and replies and quote-posts around Anthropic's August 14 response across Top and Latest surfaces. Deduplication produced 264 direct status URLs from 227 accounts, dated from August 14 at 7:16 p.m. UTC through August 18 at 2:18 p.m. UTC. All 264 resolved through X's public embed endpoint; 196 used “watermark” or “SynthID” in their own visible text, while the remainder included short or context-dependent replies returned from the same search families. Named examples were also checked directly from Arturo Villa, Thariq, ProfessorF, Alex Prompter and Antonio Ripa. The inventory is search-ranked, incomplete and not a probability sample; it supports reaction examples, not a distribution of opinion on X.
- Carl McQueen's original X Article post passed 64.7K views before this update was prepared. That count describes displayed reach, not unique readers, agreement or persuasion.
McQueen Analytics research note
Our trust framework follows responsibility, independent verification, meaningful choice and the route for correcting a failure. Here, Anthropic has published claims about a future watermark and detector. Trust will depend on whether people can verify those claims and challenge the authorship conclusion another institution draws from the signal.