Regulation ·
OpenAI to watermark ChatGPT in the EU but leaves the API optional
Original title, in French: “OpenAI filigranera ChatGPT dans l'UE mais laisse l'API en option”
- Publication
- ActuIA (France)
- Author
- Stephane Nachez
- Published on
- 6 October 2026 · read in the page structured data · updated on 6 October 2026
Summary
On 5 October 2026, OpenAI published a technical report on textGrain, the statistical watermark it will apply to text generated by ChatGPT and Codex in the European Union, in order to comply with Article 50 of the AI Act. The marking obligation has applied since 2 August 2026, with a deadline of 2 December 2026 for systems already on the market. In the API, the watermark remains disabled by default worldwide, which places the compliance responsibility on companies that integrate an OpenAI model under their own brand, according to the European Commission's guidelines. The detection rates measured by OpenAI vary widely depending on text length, topic and language, from 42.2% in Romanian to 69.0% in Spanish. Access to the detector remains restricted to approved researchers and organisations. Anthropic, Google and the vLLM project have made different scope choices for their own watermarks.
Summary based on the full text of the article.
Read the article on ActuIA →Figures cited
-
69 %
textGrain watermark detection rate by language (observed range)
ChatGPT/Codex, EU, OpenAI evaluation on 500 prompts translated into the 24 official languages of the Union, 1% false-positive threshold
« Au même seuil de 1 % de faux positifs, la détection va de 42,2 % en roumain à 69,0 % en espagnol. »
Source: OpenAI, rapport technique textGrain / centre d'aide OpenAI
-
190 organisations
Number of organisations that have signed the European code of practice on transparency of generated content
List published by the European Commission, end of July 2026
« la liste publiée par la Commission qui recensait environ 190 organisations fin juillet. »
Source: Commission européenne
-
200 tokens
Text length threshold above which the watermark remains required
European code of practice on transparency of generated content
« Ce code exempte de filigrane les « textes très courts », définis comme ceux de moins de 200 jetons. »
Source: Code de bonnes pratiques européen sur la transparence des contenus générés
-
1 %
Target false-positive rate for the textGrain detector in published evaluations
OpenAI evaluations on the ELI5 dataset and on multilingual translations
« Ils dépendent de la longueur du texte et du sujet, avec un détecteur réglé pour viser 1 % de faux positifs. »
Source: OpenAI, rapport technique textGrain
Classification
- Country
- France, Spain
- Company size
- All sizes
- Industries
- Software vendors, IT services companies, telecoms · All industries, or industry not specified
- Companies
- vLLM (cited), ETH Zurich (cited), Red Hat (cited), KInIT (cited), Cornell (cited), Mistral (cited), Microsoft Azure (supplier), OpenAI (supplier), Anthropic (supplier), Google (supplier)
- Keywords
- ai act, watermark, generative ai, regulatory compliance, chatgpt, content transparency
Entry written by Marsanne Studio from the cited article, with the help of an AI model (claude-sonnet-5). The rights to the article and its illustrations belong to ActuIA and its author. Entry translated from French by Claude Sonnet 5.5.