Independent newsroom

Tuesday, October 6, 2026

The File

Every claim traces to a primary source. Uncertainty is labeled. Corrections are public.

OpenAI · Tuesday, October 6, 2026

OpenAI will put a hidden watermark in ChatGPT text in Europe to meet new EU rules

The File desk · Oct 6, 2026, 10:42 AM UTC

Status

Confirmed as OpenAI’s own announcement, dated Oct. 5, 2026. The catch rates are OpenAI’s tests, not an outside lab’s. The

description is from the European Commission’s page. The official EUR-Lex text was not verified; that page returned a robot check.

  • Confirmed

    OpenAI said it will add an invisible watermark to eligible ChatGPT and Codex text in the European Union over the coming weeks, and that API customers anywhere can turn watermarking on starting Monday. It stays off by default.

  • Confirmed

    The detector is not public. OpenAI says only approved researchers and expert groups can apply, because detection still misses too often.

OpenAI said Monday it will add an invisible watermark to eligible ChatGPT and Codex text in the European Union over the coming weeks, to meet the

. That law requires companies that make AI which writes text, images, or sound to mark the output so a machine can tell it was made by AI. Starting Monday, customers of OpenAI’s API anywhere can choose to turn the watermark on for some models. It stays off unless they switch it on. OpenAI is not giving the public a detector.

The system is called textGrain. It nudges the model’s word choices so a statistical signal is left behind. OpenAI’s own tests, aimed at a 1 percent false alarm, caught about 80 percent of 200-token passages and about 95 percent of 400-token passages on psychology-type writing. It did substantially worse on math. Swapping 10 percent of the words for synonyms dropped detection from about 92 percent to 66 percent. Swapping 25 percent dropped it to 17 percent. Those figures are OpenAI’s.

OpenAI says the watermark made no meaningful difference on its benchmarks for GPT-6 Astra. It plans a fuller technical report in the coming weeks and an open-source release. The post lists what a watermark cannot tell you: how much a person contributed, who owns the text, who the user is, or whether the text is true. A missing watermark does not prove a human wrote it. Checks for images and audio stay public at openai.com/verify.

The company is not releasing the text detector to the public at launch. It says missed watermarks and false positives are still too common. Approved researchers and expert groups can apply to use it.

The European Commission’s page on the

says the transparency duties apply from Aug. 2, 2026. Signing the code is voluntary. The marking rules in Article 50 of the are legal duties anyway. The Commission’s AI Act page says providers of generative AI have to ensure that AI-generated content is identifiable. The official text on EUR-Lex was not verified. That page returned a robot check.

What is still unknown or disputed

Primary sources

Every claim in this story is drawn from the documents below. If a fetch failed, that is recorded on the card.

  1. Source 1

    Our approach to EU text provenance rules

    OpenAI · October 5, 2026

    Over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union.

    https://openai.com/index/eu-text-provenance

  2. Source 2

    Code of Practice on Transparency of AI-generated Content

    European Commission · October 5, 2026

    These transparency obligations, applicable from 2 August 2026. Even though adherence to the code is voluntary, the transparency requirements under article 50 of the AI Act are legal obligations.

    https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content

  3. Source 3

    AI Act

    European Commission · October 5, 2026

    providers of generative AI have to ensure that AI-generated content is identifiable.

    https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai