ChatGPT text in the EU will carry an invisible watermark: what Pakistani freelancers should know
OpenAI is watermarking ChatGPT and Codex text in the EU under the EU AI Act. How textGrain works, what it can't prove, and what to tell EU clients.
ChatGPT text in the EU will soon carry an invisible watermark. If your clients are in Europe, this matters to you.
OpenAI said on 5 October 2026 that, over the coming weeks, it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union, to meet the EU AI Act. It is not a global default at launch. Here is what is actually changing, what the watermark can and can’t tell anyone, and what it means if you freelance for European clients from Pakistan.
What is changing, and where
According to OpenAI’s announcement and its Help Center:
- ChatGPT and Codex in the EU: an invisible watermark is being added to eligible text output, for users across all plans, in the EU only, over the coming weeks.
- The API, worldwide: API customers anywhere can opt in to watermarked text for select models, from their project or organization settings. It is off by default.
- Cloud partners: OpenAI says it is working with cloud partners to make watermarking available for OpenAI models accessed through their services in the coming weeks.
- Detector: access is limited, for now, to approved researchers and expert organizations, granted case by case. It is not publicly available at launch.
Why now: OpenAI says the EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way. TechCrunch reports that the Act’s transparency rules took effect on 2 August.
What this means for you in Pakistan: the ChatGPT rollout as announced is for users in the EU. It still matters to you: your European clients now work under these rules, and if you build products on the OpenAI API, watermarking is now a setting you can turn on.
How the watermark works
OpenAI’s method is called textGrain. It adds an invisible statistical signal to the model’s word choices. There are no hidden characters to strip out. A detector with the secret key looks for that pattern.
Because the signal lives in the words themselves, TechCrunch notes that it travels with the text when it is copied and pasted.
OpenAI says it saw no meaningful change in its models’ performance on its benchmarks with watermarking switched on.
How reliable is detection?
OpenAI is unusually direct about the limits. From its own evaluations:
| Situation | What OpenAI reported |
|---|---|
| 200-token passages (psychology-type content), 1% false-positive target | Watermark found in about 80% |
| 400-token passages, same content and target | About 95% |
| Mathematics-type content | Substantially lower, because there is less freedom in word choice |
| 400-token passages with 10% of words swapped for synonyms | Detection fell from about 92% to 66% |
| Same, with 25% of words swapped | Detection fell to 17% |
TechCrunch adds that OpenAI said short passages, maths answers and translated text are harder to detect. Those limits are part of why OpenAI is giving the detector only to approved researchers at first.
What a watermark does not tell anyone
This is the most useful part of OpenAI’s announcement for freelancers. In OpenAI’s own words, a text watermark:
- does not measure human contribution. It can indicate that an OpenAI system generated or processed part of a passage, “but not how much human judgment, editing, or creativity went into it”;
- does not establish ownership or responsibility: who owns the text, whether its use was lawful, or whether disclosure was required;
- does not identify the user: no person, account, prompt or conversation is attached to it;
- does not verify accuracy.
And the other way round: a missing watermark does not prove human authorship. The text could be too short, edited or translated, from an unsupported model, created before watermarking started, or made with another company’s tools.
The wider picture
TechCrunch reports that the announcement came two months after Anthropic said it would watermark text generated by Claude, applying it worldwide, and that Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that have committed to follow the EU’s code of practice on AI-generated content.
Kaam ki baat for freelancers
A watermark can show that an OpenAI system generated or processed text. It cannot show how much of the work was yours. So the safest position is not to hide AI use, but to make your own editing, judgment and expertise the real value you deliver.
What to do this week:
- Check your contracts. Does your agreement with EU clients say anything about AI use? If not, consider agreeing how you use it before a dispute, not after.
- Be clear about where you use AI. For example: “I use AI tools for first drafts and research; all final copy is edited and fact-checked by me.”
- Add value AI can’t. Your own examples, local knowledge, interviews and edits are what clients pay for.
- If you build on the API, decide whether to turn watermarking on for your users, and tell them.
FAQ
Will my ChatGPT text be watermarked if I’m in Pakistan?
The ChatGPT and Codex rollout announced on 5 October is for users in the EU only. API customers anywhere can choose to switch watermarking on.
Can a client run my work through OpenAI’s detector?
Not at launch, unless they are an approved researcher or expert organization. OpenAI says it isn’t making the detector public yet because of the risk of missed watermarks and false positives.
Can editing remove the watermark?
OpenAI’s own tests show heavy editing weakens it a lot (25% of words swapped cut detection to 17%). That isn’t a reason to disguise AI text; it is a reason why a watermark result alone shouldn’t be treated as proof either way.
Sources
- OpenAI, Our approach to EU text provenance rules . Checked 6 Oct 2026 . Published 5 Oct 2026; archived copy saved
- OpenAI Help Center, Provenance signals in OpenAI-generated content . Checked 6 Oct 2026 . Archived copy saved
- TechCrunch, OpenAI will start watermarking ChatGPT's text in the EU . Checked 6 Oct 2026 . Published 5 Oct 2026, 1:36 PM PDT (6 Oct, 1:36 AM PKT); archived copy saved
Drafted with AI assistance from the sources listed above; every figure and link is checked against those sources before publishing. Spotted an error? Tell us. Discuss it on LinkedIn.