OpenAI is putting a hidden signature into the words its chatbot writes, but only for some of the people using it. On 5 October the company said it would add textGrain, an invisible statistical watermark, to eligible text from ChatGPT and Codex in the European Union over the following weeks. From the same day, API customers anywhere could switch the watermark on for selected models. It stays off unless they do.

The split matters because one marketing team can now produce marked and unmarked copy without changing how it works. A blog draft written in ChatGPT on an EU account will carry the mark once the rollout reaches it. The same draft generated through a content tool built on OpenAI’s API will not, unless the tool’s developer has opted in. OpenAI ties the change to the EU AI Act’s transparency rules, which apply from 2 August 2026. It follows Anthropic, which said in August that future Claude models would watermark their text worldwide.

What does OpenAI’s textGrain watermark actually do?

OpenAI’s textGrain nudges which words a model picks, leaving a statistical pattern that a detector holding a secret key can look for. Nothing is added to the page: no label, no symbol, no hidden characters. A reader sees ordinary prose. Because the signal lives in the wording itself, copying and pasting the text does not strip it, according to TechCrunch’s account of the technical report OpenAI wrote with researchers from the University of Pennsylvania and Yale.

The scope is narrow. In ChatGPT and Codex the mark applies to eligible users on all plans, in the EU only, and OpenAI said it was not making watermarking a global default at launch. On the API it is opt-in, off by default and limited to select models. OpenAI did not name those models, explain how to switch the feature on, or mention a price. It said it was also working with unnamed cloud partners to offer watermarking on OpenAI models accessed through their services in the coming weeks.

OpenAI says the mark does not identify a user, account, prompt or conversation. It also says a watermark cannot show who owns a piece of text, whether it is accurate, or how much a person contributed to it.

How reliable is the detection, and who can run it?

OpenAI’s own evaluation shows detection works best on longer, unedited text. At a target false-positive rate of 1%, its detector found the watermark in about 80% of 200-token passages on psychology topics, rising to about 95% at 400 tokens. A token is a fragment of a word, so 200 tokens is roughly 150 words, about the length of a product description or a short LinkedIn post. Detection was substantially lower on mathematics, where there are fewer equally good ways to phrase an answer.

Light editing does real damage. In OpenAI’s tests on 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing a quarter of them cut it to 17%. OpenAI says short, edited or translated text may go undetected, and that a missing watermark “does not prove human authorship”.

No independent evaluation of textGrain has been published. Research on comparable systems points the same way. A preprint posted to arXiv in July tested three watermarking methods, including an open-source implementation of Google DeepMind’s SynthID-Text, and found that paraphrasing removed the mark from almost every text where it had first been detected. The paper has not been peer reviewed.

The detector is not public. OpenAI opened applications on 5 October for approved researchers and expert organisations, with access granted case by case, and cited the risk of false positives and missed watermarks as its reason for holding back. It plans to release the technology as open source but gave no date.

What changes for marketing teams and the agencies they use?

For marketing teams, whether content carries a mark is now set by product, location and settings rather than by who wrote it. Anthropic’s approach covers more ground than OpenAI’s. Its watermark, a version of SynthID-Text, applies globally because the company says it does not yet have a reliable way to limit it by region. It covers future Claude models; models launched before 2 August fall under an EU transition period, and Anthropic says it is adding watermarking to them over the coming months. Anthropic said on 1 September that its detection API was in private preview for eligible organisations, including regulators, media, researchers and enterprises with their own verification obligations.

That gives a single marketing department several routes into its content, each behaving differently. Freelancers and agencies may use ChatGPT, a writing tool built on the API, Claude, or a mix, and the client rarely knows which. None of it is visible to a reader. What changes is what a regulator, platform or researcher with detector access could later check, and how much weight that check deserves given OpenAI’s own figures on editing.

The legal requirement is less specific than “watermarking”. Article 50(2) of the AI Act requires providers of generative AI systems to mark output in a machine-readable format and make it detectable as artificially generated. It does not prescribe a method, and it does not apply to systems performing an assistive function for standard editing. The European Commission gives systems placed on the market before 2 August until December 2026 to meet the marking obligation.

OpenAI had built a text watermark before. A 2024 Wall Street Journal report said the company held it back partly over concern that users would move to rivals that did not mark their text. In the EU, the AI Act now puts the same marking obligation on its rivals. OpenAI has not said when, or whether, textGrain will become the default for ChatGPT users outside the EU.