OpenAI says watermarks can't prove *who* wrote the text

OpenAI has published its approach to the EU AI Act's requirement for machine-readable watermarking of generated text ^[2]. The company plans to add metadata to outputs from ChatGPT and its API, and is working on watermarks that survive editing. It also says something every organisation using generative AI should notice: watermarks can show where text came from, but they cannot measure how much a person contributed to it, and the absence of a watermark does not prove a person wrote it.

That is the provenance problem in a sentence. A watermark is like a postmark: it tells you when and where a letter left, but it doesn't tell you who wrote it or what they changed before they sent it.

What Article 50 requires

Article 50.2 of the AI Act says providers of general-purpose AI models with systemic risk must ensure outputs are marked in a machine-readable format ^[1]. For models already on the market before 2 August 2026, that requirement applies from 2 December 2026 ^[3]. The marking must allow the content to be detected as artificially generated or manipulated.

Article 50.4 says deployers—organisations and people using the AI—must disclose deepfakes and AI-generated text published to inform the public on matters of public interest, unless a human has reviewed it and holds editorial responsibility ^[1].

OpenAI is a provider, so Article 50.2 is its duty. The company says it will add metadata indicating AI generation to text outputs from ChatGPT and the API ^[2]. It is also researching watermarks that can survive common edits like paraphrasing and translation, though it notes that all current watermarking methods can be removed or spoofed ^[2].

What watermarks cannot do

OpenAI's statement is blunt about the limits. It says watermarks "are not a reliable way of determining whether content is human or AI-generated, since they do not account for the extent of human contribution in creating a piece of content" ^[2]. Someone could generate a draft with AI, rewrite it heavily, and the result would still carry a watermark even though the person did most of the work. Conversely, someone could strip the watermark from generated text and present it as their own.

The company adds that the absence of a watermark does not prove human authorship ^[2]. Text can lose its mark through editing, or never have one if it came from a system that does not watermark.

That matters for anyone who will have to show later what was AI-generated and what was not. The marking duty is on the provider, but the disclosure duty under Article 50.4 is on the deployer, and proving contribution is the deployer's problem.

Why a record of generation matters

This is where a digital witness fits. The watermark is on the output, readable by whoever receives the content. A witness is for the organisation that generated it, and it seals a record of what was generated, when and by which system.

Say a publisher uses a generative model to draft an article. The model's output carries a watermark, as required. A journalist then rewrites parts of the draft, checks the facts and adds new material. The published piece is mostly the journalist's work, but it started with AI. If someone later questions whether the article should have been disclosed under Article 50.4, the watermark in the final text might be gone, faint or misleading, because the text has changed.

A witness seals the original generation event: the model version, the prompt, the output and the time. The content itself stays with the publisher. When the question comes, the publisher can show what the AI generated, what the journalist changed and when the final version was published. That record, sealed at each step, hasn't been quietly rewritten since.

The witness is WitniumChain docs. It takes a fingerprint of the generation record, not the content, and seals that fingerprint on a chain where any later change would show. To check, you calculate the fingerprint again and compare it with the sealed one. If they match, the record is unchanged. The first 2,000 witnesses a month are free.

A witness does not measure human contribution any more than a watermark does. It does not authorise a decision to publish or to disclose. What it does is give you a record of what the AI generated, provably unchanged, so the question of contribution starts with facts both sides can check.

What this means for you

If you use generative AI and publish the results, ask yourself these questions now:

• When you edit AI-generated text, how will you show later what the AI wrote and what you changed?
• If the watermark is stripped or altered in editing, what record do you have of the original generation?
• When Article 50.4 asks whether content was AI-generated, who keeps the record: you, or the provider whose watermark may no longer be there?
• If someone questions a piece of content months after publication, can you show which model version generated it, what the prompt was and when it happened?

OpenAI's approach shows the watermark doing what it is designed for: telling the reader, or a machine, that content came from an AI. The organisation's own record of what was generated is a separate matter, and one a watermark does not solve.

What to watch

Article 50.2 applies to models already on the market from 2 December 2026 ^[3]. Providers have until then to implement machine-readable marking. Organisations using those models should decide now how they will keep their own records of generation events, because the disclosure duty under Article 50.4 already applies, and questions about contribution will come later.

The AI Office's guidance on Article 50 is expected in 2026. Until then, the text of the regulation is the rule.