Can an AI Text Watermark Prove Authorship? What EU Rules Require

Can an AI text watermark prove authorship? No. Even a verified watermark can support only a limited conclusion: the passage contains a signal associated with a particular model or system. It cannot, by itself, identify who prompted, edited, copied, or published the text—or prove that a particular person authored it.

That distinction matters when a detector result informs a publication or an allegation. A watermark detector looks for a particular embedded or model-associated signal. An AI-text classifier instead estimates whether text was generated by AI from patterns in the wording. A classifier’s result is not proof that a watermark exists; and a detector’s positive result does not establish that a signal is genuine or tied to a particular model unless the method and its association are verified.

What can an AI text watermark prove—and what can it not prove?

A genuine, verified watermark may indicate that text carries a signal associated with a model. To support that conclusion, a reviewer needs to understand the detector’s method, what signal it tests for, how the signal is linked to the model, and whether the result can be reproduced on the preserved passage. A bare “AI detected” score does not establish those points.

Even a verified model-associated signal does not show who used the model or authored the final passage. Someone may have prompted the system, edited its output, copied the text, or published it. Authorship, ownership, editorial responsibility, and accuracy are separate questions. A watermark does not settle who created the wording, who holds rights in it, who stands behind publication, or whether its claims are true.

How reliable is a text-watermark detector?

Detection depends on the signal, detector, and text tested. Passage length, subject matter, rewriting, and other edits can affect whether a watermark remains detectable. A negative result therefore does not establish that a person wrote the text. A positive classifier result may indicate that text resembles AI-generated writing, but it is not confirmation of an embedded watermark. Neither kind of result identifies an individual author.

Do not generalize a vendor’s reported performance without checking its evaluation conditions. Ask what text was tested, how long the passages were, what edits were applied, which detector version was used, and how false positives and missed detections were measured. Results from one setting are not a guarantee for different material or a different detector.

Does an EU machine-readable watermark satisfy the duty to disclose AI text?

Not by itself. Article 50(2) of the EU AI Act requires providers of systems that generate synthetic audio, image, video, or text to ensure their outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. This provider duty has an exception where the system performs an assistive function for standard editing, or does not substantially alter input data or its semantics. Article 50 applies from 2 August 2026, so these duties are not yet applicable before that date. ^[1][2]

That provider marking duty is separate from the deployer disclosure rules in Article 50(4). In covered cases, deployers must disclose when an AI system generates or manipulates a deepfake. They must also disclose AI-generated or manipulated text published to inform the public about matters of public interest. For that text, the disclosure exception applies when the material has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication. The deployer disclosure duty also does not apply where the use is authorized by law to detect, prevent, investigate, or prosecute criminal offences. ^[1]

The Commission distinguishes public disclosure from machine-readable marking: an embedded mark alone does not meet a separate deployer disclosure duty. For deepfakes, disclosure must be understandable and perceivable without special technical tools or dedicated actions, and made no later than first exposure. The Act also includes context-specific rules, so publishers should check the relevant provision rather than assume every AI-assisted publication has the same duty. ^[3][1]

How should a publisher handle a watermark result?

Treat the result as a lead for review, not a ready-made attribution. Preserve the evidence, assess its limits, and decide what the combined record supports. A useful rule is to match the strength of the wording to the strength of the evidence:

• Detector result alone: If the detector reports a signal but its method, model association, or test conditions are unclear, report only that the detector returned a positive result. For example: “A detector reported a possible watermark signal in the passage; we could not independently verify its model association.” Do not state that a named person used AI or wrote the text.
• Verified signal, limited corroboration: If the method and model association are verified, but there is no reliable record of who produced or handled the passage, say that the signal is associated with the model. This still does not establish individual authorship.
• Signal plus independent records: Draft history, workflow records, or other reliable evidence may support a stronger account of how the text was produced. Explain what those records show and what they do not. Corroboration can support a conclusion about the process, but may still not prove that one person authored every part of the passage.

1. Preserve the material. Save the exact passage tested, its length, the date of testing, the detector and version, and the reported result. Note known edits and any missing context.
2. Check the method. Find out whether the tool detects a watermark or classifies text as AI-generated. For a claimed watermark, check the signal’s model association, validation, and reproducibility. Record known false-positive and missed-detection risks.
3. Seek independent records. Where lawful and appropriate, review drafts, revision history, relevant workflow or account records, and publication records. Consider who had access to the text and whether it may have been copied or edited.
4. Keep legal compliance separate. Assess whether a provider marking duty applies and whether the publication context triggers a distinct deployer disclosure duty. A detector result does not answer either legal question.

Frequently asked questions

Can ordinary editing change whether a watermark is detected?

It can. Rewriting or replacing words may alter the text a detector examines. Without testing the particular detector on comparable edits, do not assume that a signal will survive—or that a missing signal means the text was not generated or manipulated by AI.

Does the EU AI Act require disclosure for every AI-assisted publication?

No. Article 50 covers specified systems and contexts, including deepfakes and certain AI-generated or manipulated text published on matters of public interest. It also includes exceptions, including the human-review or editorial-control exception for that text. The rules apply from 2 August 2026, not before. Check Article 50 and relevant Commission guidance rather than assuming that every use of AI triggers the same disclosure rule. ^[3][1][2]

Sources

• ^[3] European Commission, guidance on Article 50 transparency obligations under the AI Act, including deployer disclosure for deepfakes and text published on matters of public interest.
• ^[1] Regulation (EU) 2024/1689, Article 50(2) and 50(4), on provider marking and deployer transparency obligations.
• ^[2] European Commission, AI Act implementation timeline, listing 2 August 2026 as the application date for Article 50.