OpenAI says it will start adding invisible watermarks to eligible ChatGPT and Codex text in the EU over the coming weeks.
TechCrunch reported that OpenAI plans to apply the watermark only in the European Union for ChatGPT and Codex at launch, while developers using the company's API can enable it worldwide for select models from now. The Register said the move is intended to comply with the EU AI Act's requirement that generative AI output be identifiable in a machine-readable way.
In plain English, this is not a visible label added to text. The watermark works by nudging the model's word choices so the finished passage carries a statistical pattern that a detector can look for later. That matters because copied and pasted text would keep the pattern, but it also means the system depends on detection rather than on readers seeing a clear mark.
What OpenAI is changing
According to TechCrunch's account of OpenAI's announcement, the company will roll out watermarking over the coming weeks to eligible ChatGPT and Codex users on all plans in the EU. For API customers, the option is available starting now for select models anywhere in the world, and it is off by default.
The Register reported that OpenAI does not plan to make text watermarking the default outside Europe for ChatGPT and Codex, unlike Anthropic's worldwide approach for Claude described in both source articles. That makes this both a product change and a regional policy response rather than a global switch across all OpenAI text output.
The same reporting also notes that OpenAI already applies other provenance signals to AI-generated images, audio and video. The new step is specifically about generated text, which has been harder to mark in a way machines can later identify.
How the text watermark works
OpenAI calls its method textGrain. As TechCrunch explained, large language models generate text by choosing the next word from a range of likely options. OpenAI's system subtly steers some of those choices, creating a hidden pattern in the wording. A detector can then test whether that pattern is present.
The Register described the same idea as an invisible statistical signal added to the model's word choices. In practice, that means the text itself carries the marker rather than an attached metadata tag. If the text is copied into another document or system, the signal may travel with it.
That design also sets a clear limit. OpenAI is not saying the watermark identifies a specific user. TechCrunch reported that OpenAI said the watermark does not identify the user, and that the company saw no meaningful change in model performance with watermarking enabled.
Why detection is imperfect
The most important caveat is that OpenAI's own testing says the watermark is easier to detect in longer passages and easier to defeat when text is edited. TechCrunch reported that replacing 10 percent of words with synonyms cut detection from about 92 percent to 66 percent.
The Register added further detail from the technical paper: in 400-token passages, replacing 25 percent of words reduced detection to 17 percent. It also reported that short passages of around 200 words may be flagged only 80 percent of the time, compared with 95 percent for a 400-word passage, and that results are worse for functional text such as maths or code.
OpenAI's reported detection rate falls as more words are replaced with synonyms.
OpenAI has also cautioned, according to TechCrunch, that a missing watermark does not prove a human wrote the text. The passage may be too short, too heavily edited, translated, or produced by another provider's AI system.
The Register further noted that watermarking based on word substitution may bring another risk: changing a word could alter meaning, especially in passages where wording carries a precise fact or figure. The article says the textGrain paper does not address that likelihood directly.
Why this matters for organisations
The immediate driver is regulation. TechCrunch said the EU AI Act's transparency rules took effect on 2 August and require AI companies to mark AI-generated content in a way other systems can identify. For organisations using ChatGPT, Codex or OpenAI's API in Europe, that creates a more concrete compliance and governance question: whether to rely on watermarking as one signal of provenance, not as a definitive test.
The sources also show why some teams may be cautious. OpenAI is limiting detector access initially to approved researchers and expert organisations, according to TechCrunch's report, because of reliability and responsible-use concerns. That suggests many businesses will not immediately have broad direct access to the checking tool itself.
For developers, the change is more immediate and optional. API customers can choose to turn watermarking on for select models worldwide, as both TechCrunch and The Register reported. That may appeal to organisations that want a machine-detectable signal attached to generated text for provenance checks or compliance-related evaluation, though the sources do not say how widely customers plan to adopt it.
The reporting also points to competitive differences. Anthropic is applying text watermarking worldwide for Claude, while OpenAI is starting with the EU for ChatGPT and Codex and leaving the API option off by default elsewhere. This suggests providers are making different choices about how far to extend compliance-style controls beyond Europe's legal requirements.
Why watermarking is more of a signal than a verdict
The strongest conclusion supported by these reports is that OpenAI is adding a machine-detectable text watermark to meet a specific regulatory requirement, not introducing a definitive test for authorship. In the EU, eligible ChatGPT and Codex output will begin carrying that signal over the coming weeks, and API customers can opt in for select models worldwide.
What matters most is the trade-off. The watermark may help systems check whether OpenAI generated or processed part of a passage, but OpenAI's own results show that editing, short passages, translation, maths and code can all make detection less reliable. Interpreting the evidence conservatively, organisations assessing provenance or compliance should treat the watermark as supporting evidence rather than a final answer.