OpenAI Is Putting an Invisible Watermark on ChatGPT Text in the EU

OpenAI has begun quietly tagging the words ChatGPT writes in Europe. The company announced on Monday that eligible text generated by ChatGPT and Codex in the European Union will soon carry an invisible, machine-readable watermark — a statistical fingerprint baked into the model's word choices that readers cannot see but software can detect. It is the most aggressive step yet by a major AI lab to make AI-generated text traceable, and it exists because Brussels — not Silicon Valley — demanded it.

OpenAI's logo displayed at its headquarters. The company announced on October 5, 2026 that ChatGPT and Codex text generated in the EU will soon carry an invisible watermark.

The Announcement: textGrain Rolls Out in the EU

OpenAI said on Monday, October 5, 2026, that it will embed a hidden signal in text generated by ChatGPT and Codex for users in the European Union, according to reports from Gizmodo and TechCrunch. The system is called textGrain, and it will become mandatory for eligible ChatGPT and Codex text output across the EU over the coming weeks, applying to every subscription plan.

The move is a direct response to the EU AI Act's transparency requirements, which took effect on August 2, 2026 and require AI companies to mark generated content so it can be identified in a machine-readable way. OpenAI's rollout answers that rule with a system designed to survive the most common way a watermark dies: being copied and pasted.

Developers worldwide get a choice. From October 5, API customers anywhere can switch the watermark on for selected models — but it stays off by default. The feature will also reach users through cloud partners such as Microsoft Azure in the coming weeks, extending the signal beyond OpenAI's own apps.

How an Invisible Watermark Actually Works

textGrain does not stamp a visible logo or hide special characters in the text. Instead, it nudges the model's word choices against a secret key, embedding a statistical pattern in the selection of words that is invisible to the reader but detectable by software built to look for it. Because the signal lives in the word choices themselves, it travels with the text through copy and paste.

OpenAI says the watermark does not identify the person who generated the text. It does not reveal prompts or conversation history, does not establish who owns the output, and does not say how much of a document was written by AI versus a human. It answers one question only: does this text carry OpenAI's watermark?

The company published a technical report on the method, co-authored with researchers from the University of Pennsylvania and Yale, and has said it plans to open-source the watermarking implementation. For now, however, access to the detector itself will be limited to approved researchers and expert organizations through an application process.

What It Can — and Can't — Catch

The system is far from foolproof, and OpenAI has been unusually candid about its limits. According to the company's own assessments, the watermark was identified in roughly 80% of texts of 200 tokens and about 95% of texts of 400 tokens. Performance drops sharply once a human starts editing.

Replacing just 10% of the words with synonyms dropped detection from about 92% to 66% in OpenAI's tests, and replacing a quarter of the words cut it to roughly 17%. Short passages, heavily constrained text, math answers, and translated text are also harder to verify. OpenAI warns that the detector can produce false positives and false negatives — and, crucially, that a missing watermark "does not prove human authorship."

The limits matter because they define what this technology actually is: a tool for researchers and platform operators, not a lie detector for the public. A teacher who runs a student essay through the detector and finds nothing has learned nothing. A newsroom that finds a watermark knows only that OpenAI's software touched the text — not whether the information in it is true.

European Union member-state flags. The watermark is OpenAI's response to the EU AI Act's transparency rules, which took effect on August 2, 2026.

Why Brussels Is Driving This

The real story of textGrain is not the technology but the geography. OpenAI built the watermark because European regulators made machine-readable identification of AI-generated content a legal obligation — and outside the EU, where no such rule exists, the company leaves it switched off by default.

OpenAI is not alone in signing onto the European approach. Anthropic, Google, Meta, Microsoft, and OpenAI are all signatories to the EU's code of practice on AI-generated content. Anthropic went further than anyone: in August 2026 it rolled out text watermarking for Claude worldwide, mandatory for all users with no opt-out — a contrast OpenAI pointedly refuses to follow outside Europe.

Google's DeepMind has shipped watermarking of its own, notably the SynthID system, years earlier. The pattern across the industry is consistent: the default setting for provenance tracking is being set by regulation, not by corporate conscience.

An illustration of invisible digital watermarking. OpenAI's textGrain embeds a statistical signal in the model's word choices that specialized software can detect.

What This Means

For years, the AI industry's answer to the deepfake-text problem was essentially "trust us." textGrain is the first time a dominant lab has shipped a machine-verifiable answer — but only where a law forced its hand, which tells you how much faith to place in voluntary approaches elsewhere.

The practical effects will fall unevenly. Teachers, professors, and employers will be tempted to treat the detector as a cheating detector; it is not, and OpenAI's own warnings make that explicit. Newsrooms and platforms, on the other hand, gain a real forensic tool for tracing coordinated campaigns — if they can get detector access. Researchers studying how AI text spreads online finally get a way to measure what was previously invisible.

The deliberate choice to restrict detector access while promising to open-source the implementation reveals the tension OpenAI is managing: widespread detection is the point of a watermark, but a public detector also becomes a training target for anyone building watermark-removal tools. That arms race has already begun. Swapping a few words demonstrably degrades the signal, and purpose-built evasion tools will follow.

What Happens Next

The EU rollout begins over the coming weeks, and OpenAI says eligible ChatGPT and Codex users in Europe will see it automatically, on every plan. API adoption will be the broader test: with the switch off by default and available through partners like Azure, most of the world's AI-generated text will still go untagged.

Watch three things. First, whether detector access expands beyond approved researchers — without wider access, the watermark is a promise, not a practice. Second, whether the open-sourced implementation survives contact with evasion tools, or degrades into the same cat-and-mouse game that consumed every previous attempt at content provenance. Third, whether the United States, which has no equivalent federal rule, ever imposes one — because OpenAI has just shown that the technology ships exactly as far as the law requires, and not an inch further.