Technology

ChatGPT in Europe Is Getting a Hidden Label. Here's What That Means

Martin HollowayPublished 2month ago4 min readBased on 15 sources
Reading level
ChatGPT in Europe Is Getting a Hidden Label. Here's What That Means
Photo by Jernej Furman from Slovenia / CC BY 2.0

OpenAI said on October 5, 2026, it will add a hidden label to text made by ChatGPT and Codex in the European Union to follow the EU AI Act. OpenAI

The change will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans, but only in the EU. Developers anywhere in the world who use the API can turn on text watermarking for select models starting October 5, 2026. It is off by default. OpenAI is not turning it on for everyone worldwide at launch. TechCrunch

That means two tracks. ChatGPT and Codex have it on by default in one place. API use is opt-in everywhere else.

The reason is Article 50 of the EU AI Act. Clear labelling rules in key cases apply from August 2, 2026. Providers must mark AI-made content in a way machines can spot. Groups that deploy generative systems must clearly label AI-made or changed text under Article 50(4).

Brussels has published help for that rule. The Commission released a Code of Practice on Transparency of AI-generated Content, guidelines for providers and deployers of generative and interactive systems, and a set of icons for labelling AI-made content. Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that agreed to follow the code.

The label is hidden, like ink you cannot see. OpenAI slightly shapes the choice of words to leave a pattern readers miss but a detector can find. The pattern stays when text is copied and pasted. OpenAI said it does not name the user and caused no meaningful change in how well the model works when turned on.

OpenAI described the method in a technical report for a system called textGrain, written with researchers from the University of Pennsylvania and Yale. It also published a Safety news post titled Our approach to EU text provenance rules on October 5, 2026.

Testing shows limits. Changing 10% of marked words to synonyms cut detection from about 92% to 66%. Short passages, math answers, and translated text are harder to detect. For now only approved researchers and expert groups can use the detector.

Text is only one track. Images made with ChatGPT, Codex, and the OpenAI API include both C2PA metadata and SynthID watermarks. OpenAI describes a layered approach that includes SynthID watermarking for images and audio.

Anthropic plans a broader rollout for text. It said it will watermark text made by Claude worldwide. One lab uses EU defaults plus global API opt-in. The other plans global defaults.

The broader context here is how long this kind of labelling has been promised and how slowly it has shipped. OpenAI, Google and others pledged to watermark AI-made content in a 2023 White House process. EU lawmakers then spent three years writing what disclosure means in law, including a provisional political agreement on May 7, 2026, on revised AI rules.

In my view, people should read this as compliance plumbing, not as a fix for spotting AI text. A mark that survives copy and paste helps filters and audit tools. It does not survive careful rewriting. It does not solve short or tightly limited answers. It does not say who wrote what.

Looking at what this means for builders, three details stand out. First, off-by-default API watermarking means mixed results. The same models will make marked and unmarked text based on settings and place, so logging that setting becomes part of the record. Second, closed detectors limit checks. Researchers can test the claims, but product teams, teachers, and newsroom technologists cannot yet run routine checks. Third, text now joins images and audio. C2PA and SynthID already cover pictures and sound in OpenAI's system. TextGrain fills a gap, but joining signals across types is still work for deployers.

Worth flagging for deployers, the law covers both providers and deployers. A provider-side mark covers only part of the chain. Product teams still need clear notices for users, icons where fitting, and data handling that survives sharing. The method can get better in sampling, entropy calibration and detector design. If detectors become open, work together, and earn trust, these quiet marks could make AI content easier to trace and manage.