If a Claude trace is found, who should be considered the writer?
Say a Claude watermark is detected in a freelance manuscript or hiring assignment. Can you immediately conclude that “AI wrote all of it”? Conversely, if no watermark appears, can you trust that a human wrote it?
Neither conclusion is valid. Claude’s text watermark can indicate only that Claude may have generated or processed a sentence. It cannot determine the original author, ownership, or how much a person revised it.
That said, the old explanation that “no one can check watermarks yet” now needs updating. As of September 2026, the text-detection API is in private preview for qualified institutions and companies, while a tool to check credentials in files created by Claude is publicly available.
Text checks and file checks are different functions
What ordinary users can check right now is a file’s Content Credential. The public Claude Content Checker verifies C2PA-based signed metadata embedded in a file; it does not inspect watermarks in copied-and-pasted text.
| Category | Text watermark | File Content Credential |
|---|---|---|
| How it is marked | Statistical patterns are reflected in word selection | C2PA-based signed metadata is attached to the file |
| Current way to check | Private-preview Detection API for qualified institutions and companies | Public Claude Content Checker |
| What it can show | Whether Claude may have been involved in generating or processing text | Whether a Claude Content Credential is found in the file |
| What it cannot show | Original author, ownership, or extent of human editing | Original author, ownership, or the file’s complete editing history |
The text Detection API preview is available to regulators, law-enforcement agencies, journalists, fact-checkers, independent researchers, educational institutions, EU civil-society organizations, and companies with relevant verification obligations. An interest-registration path is open, but approval criteria, API specifications, and a general-release date have not been confirmed.
Anyone can use the file checker, but it is not a complete certificate of provenance either. Metadata can disappear through resaving, format conversion, or screenshots, and even a discovered credential does not identify who authored the original.
A watermark is a pattern of word choices, not hidden characters
It does not secretly insert special characters or separate metadata into the text. Anthropic uses an adaptation of SynthID-Text, published by Google DeepMind in Nature. When the model selects the next word, it reflects a statistical pattern tied to a secret key, and the detector analyzes whether enough of that pattern remains.
Because of this structure, signals can accumulate in long drafts or extensive rewrites, while short phrases and light proofreading may leave too little evidence to detect. Factual sentences or code whose wording is difficult to change may also have weak signals because there are fewer words to choose from; heavy editing, translation, or mixing with other writing can make detection fail.
It is also not accurate to say that watermarking always remains when Claude proofreads human-written copy. The watermark applies only to words selected by Claude, so a few grammatical or punctuation fixes may leave almost no detectable signal. If Claude substantially rewrites the sentences, however, detection becomes more likely.
A positive result is not proof that Claude alone authored it, and a negative result is not proof of human authorship. A detection result should not be treated as an automatic finding of copyright or contract violation. It is one signal that calls for a closer look at the generation time and model record, manuscript versions, and human editing history.
The original SynthID-Text study compared about 20 million Gemini responses and reported no quality decline in user and standard evaluations. But that does not establish Claude’s real-world false-positive or false-negative rates, or Korean detection performance. Anthropic has likewise not published language-specific minimum lengths, decision thresholds, or accuracy figures for the Claude implementation.
Today, reconcile generation records before relying on a detector
The first action is to place the original and final versions of a piece of content side by side. Even without access to the Detection API, documenting Claude’s scope of involvement and the human editing process creates evidence you can use later for publication or dispute response.
- Confirm the model and generation time. Find the model and date in API logs or work records, then compare them with official support documentation to see whether watermarking applied at that time. Supported models released after August 2, 2026 have marking enabled from launch, but that does not mean every older model is already supported.
- Distinguish the work Claude performed. Record it as one of: “full draft,” “major rewrite,” “translation,” or “light proofreading.” Also note the limitation that short text and light proofreading can be missed even when Claude was involved.
- Check the type of inspection method. Qualified organizations should review the application path for the text Detection API in official support documentation. Without access, do not present third-party AI-detection results as if they were Claude watermark results.
- Inspect files separately. If there is a supported file created by Claude, check its Content Credential in the public Content Checker. Do not combine that result with the body-text inspection result; record it in a separate field.
- Link human edits and approvals. Preserve the initial draft, final version, revisers, and approvers together. In the final assessment, separately record “possible Claude involvement,” “applicable model and time,” “human editing history,” and “disclosure-obligation review.”
Content ID: ______
Generation date · model: ______
Claude task: full draft / rewrite / translation / proofreading
Original · final version preserved: yes / no
Human reviewer · editor responsible: ______
Text-detection access: yes / no
File credential: positive / negative / not applicable
Disclosure-obligation review: complete / incomplete
The success criterion for this review is not separating people and AI with a single watermark. The first deliverable is a state in which you can explain how far Claude was involved, who changed what, and who assumed final responsibility.
EU rules also require separating “marking” from “disclosure”
Machine-readable marking applied by an AI provider and disclosure to readers by a content publisher are separate obligations. Article 50(2) of the EU AI Act requires providers of AI systems that generate synthetic content to ensure the outputs are marked in a machine-readable format. However, this obligation is exempted where an AI system assists with standard editing or does not substantially alter the input data provided or its meaning. Whether a specific proofreading task qualifies for the exception cannot be decided by the label “proofreading” alone; its actual scope of change must be examined.
Article 50(4) imposes a disclosure obligation on deployers publishing AI-generated or manipulated text to inform the public on matters of public interest, while providing an exception where there has been human review, editorial control, and editorial responsibility. The related transparency obligations apply from August 2, 2026.
Which obligations apply to a particular company or post, however, depends on jurisdiction, the content’s purpose, the organization’s role, and its editorial process. Watermark detection alone must not be used to determine a disclosure obligation or legal liability.
If you want to dig deeper
How Claude’s text watermark works Anthropic explains the watermark’s principles, the current access scope for the Detection API, and the limits of detection results. anthropic.com
How Claude marks AI-generated content This official support document lets you check applicable models and product paths, as well as the difference between text and file marking. support.claude.com
Code of Practice on Transparency of AI-generated Content See how the European Commission distinguishes transparency obligations for providers and deployers. digital-strategy.ec.europa.eu
Scalable watermarking for identifying large language model outputs The original study covers SynthID-Text’s design and its quality evaluation using roughly 20 million Gemini responses. nature.com
