OpenAI Rolls Out Invisible Text Watermarks in EU ChatGPT and Codex Under AI Act Rules

OpenAI is introducing invisible, machine-readable watermarks for eligible text generated by ChatGPT and Codex in the European Union. The rollout begins over the coming weeks and covers all ChatGPT plans, while API customers worldwide can opt in for selected models from October 5, according to the company’s announcement and summaries of its plan.

The API setting remains off by default, and OpenAI says it won’t become a global default at launch. That split gives the company a compliance path for European consumer services while leaving developers control over whether generated text carries the signal in other markets. The approach also places OpenAI’s text provenance policy alongside broader EU transparency obligations for AI-generated content.

TextGrain embeds a statistical signal through the model’s word choices rather than inserting visible characters or metadata. A detector analyzes enough text to identify the pattern, allowing the output to be classified as AI-generated without changing how it looks when copied, pasted, or read.

Testing figures reported by AI Weekly put detection at about 80% for 200 tokens and 95% for 400 tokens under ideal conditions. The same account says replacing 25% of the words with synonyms reduced detection to 17%, while The Decoder cites tests in which performance fell as low as 1% after substantial rewriting. Those results frame watermarking as a provenance aid, not a permanent label that survives every edit.

The rollout responds to Article 50 of the EU AI Act, which requires providers and deployers of certain AI systems to disclose AI interaction and mark generated or manipulated content in a machine-readable form. European Commission guidelines published July 20, 2026, together with a transparency code of practice, complete the main framework for applying those obligations, according to Paul Weiss.

The rules cover more than text. Guidance discussed by Mindstudio and the Artificial Intelligence Act resource also addresses images, audio, video, and open-source systems, with obligations extending to organisations that deploy AI content. Because major providers operate shared services across regions, compliance features introduced for European users can shape product defaults, developer workflows, and content checks elsewhere.

OpenAI is starting access to its detection capability with researchers before making it available more broadly. That sequencing gives independent investigators a chance to study detection accuracy, robustness, and failure cases while the system enters production use.

The research-first approach matters because short passages and rewritten text create difficult testing conditions. A detector that performs well on long, untouched generations can struggle with editing, translation, paraphrasing, or mixed human and machine authorship. Public evaluation will also test whether the watermark creates useful provenance signals without encouraging false certainty about who wrote a passage. OpenAI’s announcement describes where watermarking applies and how detection works, while secondary coverage fills in the reported performance benchmarks.

Developers using OpenAI’s API worldwide can opt in to watermarked output for select models, with the feature disabled by default. That choice affects applications that publish model-generated writing, pass text between systems, or need to demonstrate provenance to customers, schools, regulators, or platforms.

The policy also creates operational questions. Teams will need to decide whether to preserve the signal through editing pipelines, how to explain detection results to users, and whether their own products must disclose AI-generated material under local rules. Anthropic’s comparable SynthID approach, cited by The Decoder, provides a competitive reference point, while OpenAI’s opt-in global API model gives developers more control over deployment behavior.

OpenAI’s immediate next steps are the European rollout for ChatGPT and Codex, continued API availability, and researcher access to detection tools. The company has not made text watermarking a global default, leaving future policy decisions tied to technical performance, user response, and regulatory expectations.

The larger test is whether statistical provenance survives real-world handling. Watermarks can help platforms and researchers identify unedited model output, but the reported collapse after synonym replacement limits their use as definitive authorship evidence. EU enforcement and the Commission’s guidance will push providers, developers, and publishers toward clearer labeling practices, while the quality of public evaluations will determine how much trust the signal earns.