OpenAI sets a default textGrain watermark on EU text from ChatGPT: treat it as a weak provenance signal

On 5 October 2026 OpenAI announced that it will add an invisible textGrain watermark by default to eligible ChatGPT and Codex text in the EU over the coming weeks, as its implementation of Article 50 of the AI Act. Treat a found or missing watermark as a limited provenance signal: it proves no authorship, identity, ownership or accuracy.

The announcement is relevant to anyone making decisions based on text whose provenance matters: lawyers, directors, doctors, regulators and editorial teams. The core of it is sober. A technical signal is being added that can help establish that an OpenAI system was involved in a text. That signal is not the same as a judgement about who wrote the text or whether it is correct.

What exactly has OpenAI announced about textGrain in the EU?

In its explanation of the EU provenance rules OpenAI writes that API customers worldwide can optionally activate text watermarks from 5 October 2026, and that in the weeks that follow the company will add an invisible textGrain watermark by default to eligible ChatGPT and Codex text output in the European Union. According to OpenAI, textGrain embeds a statistical signal in the word choices; it adds no hidden characters, spaces or unusual punctuation. Access to the detector will initially be limited to approved researchers and expert organisations.

In its guidelines to Article 50 of the AI Act, the European Commission states that providers of in-scope generative AI systems must add effective, reliable, robust and interoperable machine-readable markings to AI-generated or manipulated text, audio, image and video, as far as technically feasible. The European Commission also notes that the Code of Practice is voluntary while the underlying transparency obligations are legally binding, and that those transparency rules apply from 2 August 2026. The watermark itself is therefore not a new law, but a technical implementation of an existing obligation. Article 50 applies from 2 August 2026, although a limited transition period may apply to the marking and detection obligation for certain generative AI systems already on the market before that date. Anyone who wants to know more about the context will find it in our pieces on the EU AI Act and the transparency obligations and on the Article 50 transparency requirements that remain in force.

What does a textGrain watermark prove and not prove?

A found watermark may indicate that an OpenAI system generated or processed the text. According to OpenAI itself it establishes no authorship, ownership, legal responsibility, user identity or accuracy. In its updated explanation of provenance signals the company explicitly limits the conclusions organisations may draw.

That distinction is decisive. A positive signal says nothing about how much a human edited the text, who was behind the account, or whether the content is lawful and true. A negative result equally does not prove human origin: short, factual, coded, translated or heavily edited text is, according to OpenAI, harder to detect reliably, and the detector does not support all models or older output. In our estimation the biggest practical pitfall is that decision-makers read an absent watermark as confirmation that a human wrote the text. That is a fallacy: the absence of a signal is no proof of human origin.

How reliable is detection with edited or mixed text?

Reliability depends on the type of text and on how much has been edited. OpenAI itself reports that detection works better with longer, freely formulated text and more weakly with mathematical, coded or heavily edited passages. We read those figures solely as results reported by the company, not as independently verified assurance.

Independent research underlines the caution. An academic analysis on AI text watermarking after the EU AI Act states that the current debate offers insufficient verifiable certainty about both the concerns and the promises of watermarking, and that this very uncertainty constitutes a governance problem; editing and mixing watermarked and human text affect detectability. A technical study that tested ten Unicode watermarking methods against six language models shows that detectability, secrecy and knowledge of the detector setup are different properties. Our inference: a watermarking system is not a universal, forensically conclusive authorship test, and textGrain must be assessed separately for reliability before it is given weight in a decision.

What does this mean for directors, lawyers and CISOs working with sensitive text?

Our analysis: because a positive textGrain signal does not establish authorship or accuracy, it can serve as an indication but not as an accusation. In our analysis, professionals should not make a disciplinary, contractual or other consequential decision solely on a found watermark; they should consider the original text, its context and substantive review. Because a missing signal does not prove human origin with short, translated or heavily edited text, a director blocks no publication, tender or academic assessment purely on its absence. And because detector access is for now limited to approved parties, our analysis is that a CISO cannot currently deploy textGrain as their own control mechanism; organisations should therefore give primary weight to visible disclosure, access logging and process recording. Because the EU obligation is technical but not forensic, our analysis is that organisations should record the model, product, version, date and human editing for sensitive files. These records can support provenance reconstruction without broadly exposing the underlying content.

Practically, as a summary of that analysis:

• Preserve the original text and its associated context where possible.
• Record per document the model used, the product, the version, the date and the human editing.
• Combine watermarking with visible disclosure, access and process recording and an independent substantive verification.
• Do not use a positive signal as proof of misconduct and do not block an important decision solely on a missing signal.

This line aligns with broader developments around making AI transparency testable with evidence and with the accuracy principle in generative AI: provenance and accuracy are two different questions, and textGrain answers at most the first one partially.

Sources: The article draws on OpenAI's announcement and help information, the European Commission's guidelines to Article 50 of the AI Act and two independent academic studies on arXiv.