ChatGPT starts watermarking text in the EU as the AI Act's marking deadline approaches

One part of the EU AI Act is now reaching ordinary chatbot users. On October 5, 2026, OpenAI said it is rolling out an invisible watermark on eligible ChatGPT and Codex text for users in the EU "over the coming weeks", across all plans. API customers anywhere can opt in. The move responds to the Act's rule that generative AI output must be marked in a machine-readable way. That rule has applied since August 2, 2026, and providers whose systems were already on the market have until December 2, 2026 to comply.

For background on which AI Act obligations are already in force, see our explainer on what applies now.

What the law requires

Article 50 of the AI Act sets transparency duties. The one at issue here is Article 50(2). Providers of generative AI systems must ensure their output is marked in a machine-readable format and detectable as AI-generated. According to the Commission's AI Act page, Article 50 obligations became applicable on August 2, 2026.

The Digital Omnibus on AI, Regulation (EU) 2026/1744, was adopted on July 8, 2026, published in the Official Journal on July 24 and in force from July 27. It delayed the high-risk rules: Annex III systems now apply from December 2, 2027, and product-embedded Annex I systems from August 2, 2028. It did not delay Article 50. It added a transitional period for providers whose systems were already on the market: under new Article 111(4) of the AI Act, inserted by the Omnibus, providers of generative AI systems placed on the market before August 2, 2026 must comply with Article 50(2) by December 2, 2026 (recital 38 describes it as four months). ChatGPT and Codex fall into that group.

On June 10, 2026, the Commission published a voluntary Code of Practice on marking and labelling AI-generated content. It is meant to give providers a practical way to show compliance. The Commission's signatory list, last updated September 24, 2026, names about 190 organisations. OpenAI, Anthropic, Google, Microsoft, Meta and Mistral are listed as signatories of Section 1 of the code.

What OpenAI is doing

OpenAI's post, "Our approach to EU text provenance rules", sets out its approach:

• ChatGPT and Codex (EU only): an invisible watermark is added to eligible text output, for all plans, rolling out over several weeks. OpenAI says it is not making this a global default at launch.
• API (worldwide, opt-in): customers can enable the watermark for select models. It is off by default. OpenAI says cloud partners will offer it "in the coming weeks".
• The technique: called textGrain, it adds a statistical signal to the model's word choices. OpenAI says it matched or exceeded alternatives it tested, including Google's SynthID for text. That is OpenAI's own evaluation. It plans to open-source the technology and has published a technical report.
• Detection: access is not public. Researchers and expert organizations can apply, and access is granted case by case. The detector says only whether an OpenAI watermark is present. OpenAI says it does not identify users or reveal prompts.
• Images and audio: OpenAI says its image and audio tools are unchanged.

OpenAI is unusually frank about the limits. At a 1% false-positive rate, it reports detection of about 80% for 200-token passages and about 95% for 400-token passages. Math content performs "substantially worse". Editing weakens the signal: swapping 10% of words for synonyms cut detection from about 92% to 66%, and swapping 25% cut it to 17%. OpenAI also stresses that a missing watermark "does not prove human authorship".

How Anthropic's approach differs

Anthropic took a broader approach. Its help center says supported Claude models embed an imperceptible text watermark wherever Claude is offered, not just in the EU, and files such as PNG and JPEG carry C2PA Content Credentials. Anthropic says it signed the Article 50(2) Code of Practice. Text-watermark detection is in private preview for eligible groups, including regulators, researchers, media and fact-checkers. TechCrunch first reported the plan on August 11, 2026.

The two companies agree on the main point: a mark shows a model was involved, not who is responsible or whether the content is accurate. They differ on scope. OpenAI is EU-only for its consumer apps and opt-in on its API. Anthropic applies marking globally.

Who is affected

• EU ChatGPT and Codex users: nothing visible changes. Text you copy out may carry a statistical mark that an approved detector can find, especially in longer passages you haven't edited much.
• Developers on the OpenAI API: this is the open question. The watermark is off by default. The AI Act puts the marking duty on providers of generative AI systems. If you build a product on a model API and serve EU users, you may count as a provider for your system. Leaving the setting off could then mean you need another way to mark output. We could not find Commission guidance that addresses an off-by-default upstream setting directly. Check with counsel, and see our explainer on the AI Act if you only use model APIs.
• Schools, publishers and platforms: detection isn't publicly available from either company. Watermarks won't be a plagiarism checker any time soon.

What's confirmed and what's uncertain

Confirmed: Article 50 has applied since August 2, 2026. The transitional period for existing systems ends December 2, 2026 (Article 111(4)). OpenAI's EU rollout and opt-in API setting. Anthropic's global marking. Both companies' signatures on Section 1 of the code, per the Commission's list.

Uncertain: how national authorities will judge opt-in API marking. Whether other major providers will meet the December 2 date.

Meanwhile in California

The US has no federal equivalent, but California is building its own provenance rules. On September 30, 2026, Governor Gavin Newsom signed 13 more bills, almost all of them on AI. Two concern provenance: AB 2713, on provenance data under the California AI Transparency Act, and SB 1000, which strengthens that Act. Others include SB 947, which bars employers from relying solely on AI for discipline or firing decisions, and SB 951, which requires notice when AI leads to layoffs. Earlier, on September 9, he signed SB 813, creating a framework for independent organizations that verify AI systems, and AB 1405, an AI auditor registry. Newsom called on the federal government to "step forward with robust, national regulations".

For how OpenAI and Anthropic fit these rules into their wider businesses, see our explainers on OpenAI and Anthropic.