The news
OpenAI announced on October 5 that it will bring text watermarking to ChatGPT and Codex in the European Union, adding an invisible mark to AI-generated text so detection systems can identify it. The move is designed to comply with the EU's AI Act, PYMNTS reported.
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. Developers using OpenAI's API anywhere in the world can turn it on for select models starting October 5; it is off by default. OpenAI said it is not making text watermarking a global default at launch.
The system, called textGrain, nudges which words the model picks, creating a statistical pattern that is invisible to readers but that a detector can find, according to TechCrunch's account of OpenAI's description. OpenAI published a technical report on the method co-written with researchers from the University of Pennsylvania and Yale. The company said the watermark does not identify the user and that it saw no meaningful change in its models' performance with the feature switched on.
Detection will be limited at first. OpenAI said it will give initial detector access only to approved researchers and expert organizations, according to both TechCrunch and PYMNTS. PYMNTS reported that OpenAI cautioned that a watermark cannot show how much human editing went into a passage, does not establish ownership, does not link users to text and does not verify accuracy, and that the absence of a watermark does not prove a human wrote a passage.
OpenAI's own figures show the mark can be weakened. Replacing 10% of words with synonyms dropped detection from about 92% to 66%, and short passages, math answers and translated text are harder to detect, TechCrunch reported.
Article 50 of the AI Act, which applies from August 2, 2026, requires providers of AI systems that generate synthetic text, images, audio or video to mark those outputs in a machine-readable format so they are detectable as artificially generated. The law exempts systems that only perform an assistive function for standard editing or that do not substantially alter the input.
The numbers
- Detection rate, unedited text (OpenAI, via TechCrunch)
- About 92%
- Detection after 10% of words swapped for synonyms
- 66%
- AI Act Article 50 applies from
- August 2, 2026
- API watermark default
- Off (opt-in, select models)
Why CEOs should care
For companies that publish in Europe, the practical effect is that text EU-based staff draft in ChatGPT will soon carry a hidden marker. Marketing, communications and legal teams should write a provenance policy that sets which content may be AI-drafted, who reviews it and when the company discloses AI use. Article 50 also puts a disclosure duty on deployers that publish AI-generated text to inform the public on matters of public interest, though that duty does not apply when the text has gone through human review or editorial control and someone holds editorial responsibility for it.
For chief information officers and procurement teams, the watermark is a vendor question. Companies building on OpenAI's API must decide whether to switch textGrain on, since it is off by default. Teams using other model providers should ask how those vendors meet Article 50's marking duty and whether their marks survive the company's own editing and translation steps. OpenAI's figures show that editing and translation weaken detection, so a watermark alone is not a reliable audit trail.
Boards and general counsel should avoid two mistakes. A missing watermark is not proof of human authorship, a point OpenAI itself made, and a detected mark is not proof of misconduct. Employers and schools that discipline people based on detector results would be leaning on a tool that OpenAI describes as limited and that, for now, only approved researchers can use.
The bigger picture
OpenAI is following Anthropic, which announced watermarking for its Claude models in August, TechCrunch reported. That decision drew backlash from some Claude users, who argued that they had supplied the instructions and decisions while Claude was only the tool. OpenAI could hear similar objections from professional ChatGPT users who see AI-assisted drafts as their own work.
Because textGrain lives in word choice rather than in file metadata, the mark weakens as people edit, shorten or translate a passage. Keeping detection restricted and the feature off by default outside the EU suggests OpenAI sees text watermarking as a compliance tool for now, not a general-purpose AI detector.
What’s next
Watch for the EU rollout to reach all eligible ChatGPT and Codex users, whether OpenAI widens detector access beyond approved researchers, and whether it makes watermarking a default outside the EU. Companies should also track how EU regulators enforce Article 50's marking duty and whether other model providers publish their own detection results.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
