Can You Tell If Text Was Written by ChatGPT? An Invisible Watermark Starts in Europe

Not yet, for most people. On 5 October 2026 OpenAI announced that over the coming weeks it will add an invisible watermark to text written by ChatGPT and Codex, but only for users in the European Union, to meet the EU AI Act. The detection tool itself is not public: it will go first to researchers and expert organizations the company approves. Even in ideal conditions, editing weakens the mark: replacing a quarter of the words with synonyms cut detection to 17%. For ChatGPT users in Saudi Arabia, nothing changes for now.

What is a text watermark?

Watermarks on images are familiar: a faint logo over the picture. Text is different, because you cannot put a logo over words. According to OpenAI's official announcement, its technology, called textGrain, adds an invisible statistical signal to the word choices themselves. When a model writes a sentence it usually has several suitable words to choose from, and the watermark nudges it slightly and consistently toward certain choices. A reader will notice no difference, but a detector looks for that pattern and judges whether the text carries an OpenAI watermark.

OpenAI says the watermark does not hurt answer quality. In its tests on Astra, its newest model, results with and without the watermark were close. On GPQA Diamond, for example, the model scored 94.44% without the watermark and 93.94% with it.

Who is affected, and when?

Does the watermark catch every text? OpenAI's own numbers say no

The welcome part of OpenAI's announcement is how frank it is about the limits. These are its figures, all at a target false positive rate of 1%, meaning the tool wrongly flags about one in every 100 human texts:

• Short text is harder. The detector found the watermark in about 80% of passages of roughly 200 tokens (a token is roughly part of a word), and in about 95% of 400-token passages, on subjects such as psychology.
• Constrained text is harder. On subjects such as mathematics, where word choice is limited, detection rates were substantially lower.
• Editing weakens it fast. In 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.

That is why the company decided not to open the detector to everyone at launch, starting instead with researchers who can help it assess reliability.

What a watermark does not tell you

This part matters for every teacher, HR manager and business owner thinking about detection tools. According to OpenAI itself:

• A watermark does not measure human effort. It may show that an OpenAI model wrote or processed part of a passage, but not how much thinking or editing a person added.
• A watermark does not establish ownership or responsibility, or whether using the text was lawful.
• A watermark does not identify the user or link the text to an account or conversation.
• A watermark does not verify accuracy.
• No watermark does not prove a human wrote it. The text may be short, edited or translated, written with another company's tool, or older than the watermark itself.

What does this mean for you in Saudi Arabia?

For students and teachers: text that ChatGPT writes for users in Saudi Arabia does not carry this watermark by default today, and the official detector is not public. Do not rely on any single tool to accuse someone of copying from AI; the company behind the technology says the presence or absence of a watermark does not settle who wrote a text.

For businesses: if you use the OpenAI API in your product, nothing changes unless you switch watermarking on yourself. If your company serves customers in the EU, the transparency obligations in Article 50 of the EU AI Act apply from 2 August 2026, according to the European Commission's page, including that AI outputs in text, image, audio and video be marked in a machine-readable format. For more on how European rules affect tools like ChatGPT, read our guide to ChatGPT's EU classification.

How should you handle text you suspect was AI-written?

As long as detection tools are not conclusive, the safest course is practical steps that are fair to everyone:

1. Talk to the author. Ask them to explain the main idea in their own words, or to clarify a specific paragraph. People who wrote a text themselves can explain it easily.
2. Ask for drafts. Version history in Google Docs or Word shows how the text developed step by step.
3. Set a clear policy in advance. Decide in the classroom or the company when AI use is allowed and how it should be disclosed, instead of discovering the problem after submission.
4. Judge the content, not the tool. An accurate, useful text its author reviewed carefully may be acceptable even with AI help, and the reverse is also true.

The bottom line

Detecting text written by ChatGPT is moving from guesswork to technology, but it is still early: the watermark starts in Europe only, the detector is limited to researchers, and light editing weakens it a lot. Until it spreads, the best test is still to ask the author about what they wrote and discuss it with them.

Sources: OpenAI's announcement of its approach to EU text provenance rules (5 October 2026): https://openai.com/index/eu-text-provenance/ , the European Commission's page on the Code of Practice on transparency of AI-generated content, the text of Regulation (EU) 2024/1689 on EUR-Lex, and coverage by TechCrunch.