The Mark Tells You Just One Thing
OpenAI has explicitly set out what watermarking can and cannot do. The mark tells you only that an OpenAI system generated or processed part of the text. It does not tell you who used the system, how much a person contributed, who owns the text, or whether its content is correct[1]. Anthropic has also said that its watermarks contain no information about individual users[2].
| Question | What the watermark tells you |
|---|---|
| Did an AI system generate or process part of the text? | A detected mark indicates yes |
| Who used the system? | Does not tell you |
| How much did a person contribute? | Does not tell you |
| Who owns the text? | Does not tell you |
| Is the content correct? | Does not tell you |
| Does no detected mark mean a person wrote it? | Not proof that a person wrote it |
Which Text Gets Marked Now?
On October 5, 2026, OpenAI announced that it would add invisible watermarks to eligible text generated by ChatGPT and Codex in the EU over the following several weeks. From that same day, API customers worldwide can also enable the feature themselves on selected models. However, it is off by default in the API, and it is not being enabled by default worldwide at launch[1].
Earlier, on August 14, Anthropic announced that it would add watermarks to text generated by upcoming Claude models. Because there is no reliable way to apply them by region, they will be used worldwide from the outset[2]. This means that people in Korea who use Claude to draft work documents will also be covered starting with the new models.
OpenAI uses a method called textGrain. It adds a statistical signal to word choices, leaving a pattern that is invisible to people but can be recognized statistically by a detector[1]. Anthropic has based its approach on SynthID-Text, a method Google DeepMind published in the journal Nature in 2024[2].
Detection Results Alone Cannot Measure How Much You Revised
The figures reported by AI Matters show that watermark detection rates vary with text length and the extent of editing. With the false positive rate set at 1%, about 95% of 400-token texts and about 80% of 200-token texts were detected. Replacing 10% of the words in 400-token texts with synonyms reduced the detection rate from about 92% to 66%; replacing 25% brought it down to 17%. Detection rates for mathematical texts were much lower[3][4].
Anthropic's explanation points in the same direction. It says light editing will not completely remove the mark, but rewriting the text with every word changed will[2].
You therefore cannot treat a higher or lower detection result as a measure of your contribution. Merely swapping some words for synonyms lowers the detection rate, but that does not mean more of your judgment went into the text. OpenAI also states that the mark alone cannot tell you how much a person contributed[1].
The detectors are not available to everyone, either. OpenAI initially grants access only to approved researchers and specialist organizations[1]. Anthropic's detection API is in private preview. It is available to the groups specified by EU law (regulators, law enforcement, the press, fact-checkers, independent researchers, educational institutions, and EU civil society organizations) and businesses required to carry out checks for compliance[2]. For now, a team lead or company reviewer cannot simply submit a document and get an immediate verdict.
Record Your Work in Version Histories and Notes
If you write a document based on an AI draft and want to show your contribution, you need records other than the mark. A version history showing your revisions, a list of the materials you consulted, and notes on the stages at which you used the tool can serve that purpose.
The items to include in your notes are not very different from those for a book manuscript. The four columns suggested in the guide "If You Used AI in Your Manuscript, Here Is What to Tell Your Publisher" (tool, stage of use, scope of use, and what a person verified) can be used as they are for work documents, too. When a team lead or client asks instead of a publisher, the same records can provide your answer.
For work documents, keep track of three things.
- Save the AI draft separately, exactly as you received it. Without the original, you cannot compare versions to see what you changed.
- Save a separate version after each major revision. Name it to show what you did, such as restructuring the document or verifying the figures.
- List the materials you consulted and the stages at which you used the tool at the end of the document or in a separate note.
Who the Law Makes Responsible for Marking
The transparency obligations under Article 50 of the EU AI Act took effect on August 2, 2026. Providers of systems that generate text with AI must add machine-readable marks to their output so it can be detected. Exceptions apply to assistive functions for standard editing or cases where the input is not substantially changed. Generative AI systems placed on the market before August 2 have a grace period until December 2026[5].
Those publishing text also have obligations. If they publish AI-generated text to inform the public about matters of public interest without human review or editorial responsibility, they must add a label that users can recognize. Text that a person reviews and takes responsibility for, by contrast, is exempt from this labeling requirement[5].
The mark does not show that you took on the review and editing, either. A person needs to keep that record as well.
In Korea, the AI Basic Act took effect on January 22, 2026. For AI-generated content that is not a deepfake, an invisible watermark may be used instead of a visible label. The government decided to allow a guidance period of at least one year from the date the law took effect, deferring fact-finding investigations and administrative fines during that time[6].
References
- [1] OpenAI Developer Community, OpenAI's approach to EU text provenance rules, October 5, 2026.
- [2] Anthropic, How Claude's text watermarking works, August 14, 2026.
- [3] AI Matters, report dated October 6, 2026.
- [4] TechCrunch, report dated October 5, 2026.
- [5] European Commission, Quick Facts: Transparency rules for AI systems.
- [6] Ministry of Science and ICT, press release on the AI Basic Act.
Frequently asked questions
If no watermark is detected, is that proof that I wrote the text?
No. OpenAI states that the absence of a detected watermark is not proof that a person wrote the text. A model from another company may have been used, or the signal may have weakened during revision. Keep versions of your draft and notes as evidence of what you wrote.
Does the watermark contain my name or account information?
Anthropic states that its watermarks contain no information that can identify users. OpenAI also explains that the mark alone cannot tell you who used the system. All the mark can tell you is whether an AI system generated or processed part of the text.