OpenAI said on October 5 that it will embed an invisible watermark in text produced by ChatGPT and Codex for users in the European Union, a move tied to new EU rules that require content generated by artificial intelligence (AI) to carry machine-detectable marks.

The watermark will “roll out over the coming weeks to eligible ChatGPT and Codex users on all plans, but only in the EU,” TechCrunch reported, citing the company’s blog post. Developers using OpenAI’s application programming interface (API) anywhere in the world can switch it on for select models starting the same day, but it is off by default. OpenAI is not making text watermarking a global default at launch.

How the watermark works

The mark is not a visible symbol. According to TechCrunch, it subtly shapes the model’s word choices, leaving a statistical pattern that a detector can recognize. Because the signal lives in the words themselves, it travels with text that is copied and pasted elsewhere. OpenAI said the watermark does not identify the user and that it saw no meaningful change in its models’ performance with it switched on.

The method, called textGrain, is set out in a technical report by researchers from OpenAI, the University of Pennsylvania and Yale University. The report explains that a language model writes one token, a word or part of a word, at a time, and that textGrain links each choice to pseudorandom values derived from a secret key and the preceding text. The authors wrote that the detector “requires only the generated text and the secret key and does not need to know the budget used during generation.”

That budget is the central trade-off in the design. The report describes it as a limit on how much of the model’s sampling randomness the watermark may remove in exchange for a stronger signal. The authors note that some earlier watermarking approaches could lead the same prompt to produce identical responses, which matters when an application needs several distinct answers.

Known limits

OpenAI acknowledged the watermark can be weakened. “In one test, replacing 10% of words with synonyms dropped detection from about 92% to 66%,” TechCrunch reported. Short passages, math answers and translated text are also harder to detect, the company said.

Citing those limitations, OpenAI said it would “provide initial detector access only to approved researchers and expert organizations.” The company also cautioned that a missing watermark does not prove a human wrote a piece of text, and that a watermark shows only that an OpenAI system generated or processed part of a passage, not how much human judgment or editing went into it.

The EU rules behind it

Article 50 of the EU’s AI Act, which sets transparency obligations for certain AI systems, applies from August 2, 2026, according to the European Commission. Among its requirements, providers of generative AI must add machine-readable marks so that AI-generated or manipulated content can be detected. National market surveillance authorities and the EU’s AI Office are responsible for enforcement.

The Commission has also backed a voluntary Code of Practice on Transparency of AI-generated Content, which about 190 organizations had signed by the end of July. Signatories to the section for AI providers include Anthropic, Google, Meta, Microsoft, Mistral and OpenAI. Following the code is one way to demonstrate compliance; companies that do not adopt it must show they meet the marking obligations by other adequate means.

Why it matters

For students, workers and publishers in the EU, text drafted with ChatGPT will now carry a hidden signal that approved organizations can test for, even after it is pasted into another document. The move follows a similar step by Anthropic, which said two months earlier that it would watermark text generated by its Claude models worldwide; TechCrunch reported that some Claude users objected, arguing that they supplied the instructions and decisions behind the text.

TechCrunch also noted that OpenAI had built a text watermark before but held it back, partly over concern that users would switch to rival services that did not watermark, according to a 2024 Wall Street Journal report. Limiting the rollout to the EU means users elsewhere will not have watermarked ChatGPT text by default for now.

What to watch

The consumer rollout is due over the coming weeks. Key questions include whether OpenAI widens access to its detector beyond approved researchers, and whether it extends the watermark to users outside the EU.

The caveats

The detection and performance figures come from OpenAI’s own testing and have not been independently verified. The company itself says the system can be weakened by editing and does not work equally well on all kinds of text.

Corrections and tips: email the editor using the address on the Contact page.