What happened

OpenAI, the company behind ChatGPT, will add invisible, machine-readable watermarks to eligible text from ChatGPT and Codex, its coding tool. The rollout will reach eligible users in the European Union over the coming weeks. It will cover all plans in the region. Readers will not see a symbol, hidden character, or special label.

OpenAI calls the system textGrain. It changes the statistical pattern behind word choices as the model writes. A detector can search for that pattern later. The system does not insert invisible spaces, extra characters, or unusual punctuation into the text.

Why the EU comes first

OpenAI says the rollout responds to transparency rules in the EU AI Act. Those rules require generated text to be identifiable in a machine-readable way. Starting in the EU also lets OpenAI study real-world use before making a global decision.

The company is taking a different approach with its API. Customers worldwide can opt into watermarking for selected models. That option is off by default. OpenAI is not making text watermarking a global default at launch.

Why this matters

The watermark could give readers, websites, and organizations a clue about where text came from. That clue may be useful when people want to disclose or investigate AI assistance. It could also create a shared signal across many kinds of writing.

But a watermark answers only one question. It may suggest that an OpenAI system helped produce text. It cannot show that the text is accurate, fair, or safe. It also cannot identify the person who used the system.

A detector result should therefore be treated as evidence, not a final verdict. This matters wherever a wrong label could affect a student, worker, writer, or applicant.

The important limits

OpenAI says textGrain can remain detectable after some edits. Reports also note that editing can make the mark harder to find. Rewriting sentences, changing the style, or translating the passage may weaken the original pattern. The more the text changes, the less certain the result may become.

OpenAI will not make its detector public at launch. Approved researchers and expert organizations can apply for access first. The company cites the risk of missed watermarks and false positives. It wants outside groups to help test and improve the system.

What is confirmed

The reported rollout covers eligible ChatGPT and Codex text for EU users. It will arrive over several weeks and will not become a global default immediately. OpenAI has also offered an opt-in path for API customers worldwide.

OpenAI describes textGrain as an early technology with important limitations. The reports do not establish that the mark will survive every edit. They also do not establish that a detector can reliably classify every passage.

What to watch next

The next test is ordinary human editing. Can the signal survive a careful rewrite? How often does the detector miss it? How often does it flag text incorrectly? The answers will shape whether organizations use the mark as one clue or as a decision rule.

The broader issue is trust. AI companies can make text easier to identify, but people still need to judge the words themselves. Provenance can show where text may have come from. It cannot replace fact-checking or human responsibility.