textGrain watermark on EU text from ChatGPT

You must now treat a textGrain watermark as a limited provenance signal only: it indicates that an OpenAI system may have generated or processed text, but it proves nothing about authorship, accuracy, legal responsibility or the identity of the account holder. Absence of a watermark proves nothing about human origin.

An analysis of 8 October 2026 of the invisible textGrain watermark being added by default to EU text from ChatGPT and Codex argues that the signal is a technical implementation of Article 50 of the AI Act, not a forensic authorship test. The example is OpenAI's rollout of the watermark to eligible EU users in the weeks following 5 October 2026, with detector access initially limited to approved researchers and expert organisations. In our assessment, the development matters because professionals in law, medicine, regulation and editorial work now face a new signal they may be tempted to treat as conclusive proof of origin or accuracy. It is neither. The distinction between what a watermark can and cannot establish will determine whether it becomes a useful tool or a source of false confidence in your decision-making.

What does a found watermark actually tell you?

A positive textGrain signal indicates only that an OpenAI system generated or processed the text. It establishes no authorship, ownership, legal responsibility, user identity or accuracy. The watermark embeds a statistical signal in word choices; it adds no hidden characters or unusual punctuation. OpenAI's own guidelines are explicit: organisations may not draw conclusions about who wrote the text, whether a human edited it, or whether the content is lawful and true. The presence of a watermark is therefore a technical indicator of involvement by a specific vendor's system, nothing more.

What does an absent watermark tell you?

An absent watermark proves nothing about human origin. Short, factual, coded, translated or heavily edited text is harder to detect reliably. The detector does not support all models or older output. The biggest practical pitfall is reading an absent signal as confirmation that a human wrote the text. That is a fallacy. You cannot use a missing watermark as evidence that content is human-generated, and you must not make disciplinary, contractual or other consequential decisions solely on its absence.

How reliable is the detection?

Reliability depends on text type and editing depth. OpenAI reports that detection works better with longer, freely formulated text and more weakly with mathematical, coded or heavily edited passages. Independent academic research on AI text watermarking after the EU AI Act notes that the current debate offers insufficient verifiable certainty about both the concerns and the promises of watermarking. A technical study testing Unicode watermarking methods against language models shows that detectability, secrecy and knowledge of the detector setup are separate properties. Before you give weight to a watermark in any decision, you should assess textGrain separately for reliability in your specific context. Treat reported detection rates solely as results from the vendor, not as independently verified assurance.

What controls should you put in place?

1. Record the model, product, version and date — document which OpenAI system generated or processed sensitive text, when, and what version of the detector was used.
2. Document human editing and review — maintain records of who edited the text and when, to support provenance reconstruction without exposing underlying content.
3. Require visible disclosure — do not rely on invisible watermarks; require visible labelling of AI involvement in sensitive outputs.
4. Log access to detector results — if you gain access to the detector, record who ran it, when, and what decision it informed.
5. Conduct substantive review independently — treat a watermark as an indication only; make no decision based on presence or absence of a signal alone.

What should you not do?

• Treat a found watermark as proof of accuracy — provenance and accuracy are two different questions, and textGrain answers at most the first one partially.
• Treat an absent watermark as proof of human origin — short, edited or translated text may not be detected reliably.
• Make consequential decisions on watermark status alone — consider the original text, its context and substantive review before acting.
• Deploy textGrain as your own control mechanism — detector access is currently limited to approved parties, so you cannot yet use it as a primary control.
• Assume the watermark survives editing — mixing watermarked and human text affects detectability.

Tooling can flag provenance and record it; it cannot replace your professional judgement about whether text is fit for purpose. A watermark is a technical signal, not a substitute for the substantive review, editing oversight and access controls that remain your responsibility. The EU obligation is technical but not forensic. Your records—of which model was used, when, by whom, and how the output was edited—are what will support provenance reconstruction if you need it later. The watermark itself is a useful addition to that record, but only if you understand its limits.