OpenAI to Add Invisible textGrain Watermark to ChatGPT and Codex Text in the EU

Why this matters: Europe is turning AI transparency into a technical default, and the signal now travels inside the words themselves.

October 6, 2026 — OpenAI says it will add an invisible watermark called textGrain to eligible ChatGPT and Codex text for users in the European Union, in a rollout announced on October 5, 2026 and due over the coming weeks. The company frames the move as compliance with the EU AI Act, while developers elsewhere can choose whether to switch the same feature on.

Image: Layered digital text pages over a map of Europe, examined by a detector lens. Credit: Red Eye Eagle News / original AI illustration.

Editor's Highlight: This is not a visible label. According to the company, the mark lives in the pattern of word choices, survives copying and pasting, and can only be checked with a detector that is not public at launch.

Table of Contents

• Key Facts
• What OpenAI Is Switching On
• How textGrain Works
• How Reliable the Signal Is
• Why the EU AI Act Is Driving This
• How It Compares With Rivals
• What Readers and Builders Should Do
• Bottom Line
• FAQ
• Sources

Key Facts

Detail	Information
Announcement	October 5, 2026, according to the company
Technology	textGrain, an invisible statistical text watermark
EU rollout	Eligible ChatGPT and Codex text, all plans, over the coming weeks
API access	Opt-in worldwide for select models from October 5; off by default
Detector access	Limited at launch to approved researchers and expert organizations
Legal driver	EU AI Act Article 50 transparency duties, applicable from August 2, 2026
Detection example	About 95% for 400-token passages in one reported evaluation, at a 1% false-positive target

What OpenAI Is Switching On

The rollout has three separate parts, not one global switch.

First, eligible EU text in ChatGPT and Codex will be watermarked automatically, across plans, over the coming weeks. Second, API customers anywhere can opt in for select models; that setting stays off by default. Third, the detector starts gated to approved researchers and expert organizations.

Newsroom Focus: The split matters. In Europe the mark is a compliance default. Outside Europe it is, for now, a choice each developer makes.

How textGrain Works

According to the company, textGrain adds no visible badge or hidden character. It subtly shapes which words the model picks when several choices would fit, and across a longer passage those choices form a statistical pattern a detector holding a secret key can test for.

Because the signal sits in the wording, the company says it survives copying and pasting, and does not identify the user, account, or prompt. The technical work was co-authored with researchers from the University of Pennsylvania and Yale, and the company says it plans to open-source the technology.

How Reliable the Signal Is

OpenAI presents this as a probability signal, not a verdict. Its reported figures show where it works and where it breaks down.

Test	Reported detection
200-token passage	About 80%
400-token passage	About 95%
400 tokens, 10% of words swapped	Falls from about 92% to 66%
400 tokens, 25% of words swapped	Falls to about 17%

Those figures come from company evaluations reported by multiple outlets, at a 1 percent false-positive target. Short answers, maths, constrained text, and translations are harder, because there are fewer natural word choices to shape.

Analysis Note: Light editing is the weak point. Swapping one word in ten is ordinary revision, yet it cuts detection sharply — making this supporting evidence, not a stand-alone accusation.

The company cautions that a missing mark does not prove human authorship: a passage could be short, edited, translated, older, or from another system.

Why the EU AI Act Is Driving This

The legal backdrop is Article 50 of the EU AI Act. Reporting on the rollout says those transparency duties began to apply on August 2, 2026.

Providers must make generated content detectable in a machine-readable way where the law applies, with separate disclosure duties for some public-interest text. For systems already on the market before August 2, reporting describes a grace period for marking until December 2, 2026 — a date that helps explain the October timing.

Practical Takeaway: A watermark is not a visible disclosure. Publishers still need their own labelling and editorial standards.

How It Compares With Rivals

OpenAI is not the first mover here. It is, on this evidence, the more regional one.

Provider	Approach reported
OpenAI	textGrain mandatory for eligible EU ChatGPT and Codex text; API opt-in elsewhere
Anthropic	Comparable text watermarking rolled out in August 2026, applied worldwide for Claude
Google DeepMind	SynthID watermarking shipped years earlier for its systems

That contrast matters for businesses: a regional default creates two workflows, and teams must know which produced a given document. On quality, the company says it saw no meaningful performance change — a company claim to treat as untested until independent results appear.

What Readers and Builders Should Do

For ordinary EU readers, little should look different: the text reads normally, with no new badge. For schools, employers, and publishers, a positive result may support a provenance check, but it does not identify a person, measure human contribution, or prove a claim is true.

For API builders, turn the setting on if provenance helps governance or compliance, test it on your own outputs, and do not promise more certainty than a probabilistic signal can deliver.

Newsroom Focus: The detector gate is the bottleneck to watch. If only a vetted group can check the mark, most newsrooms and moderators cannot verify it themselves at launch.

Bottom Line

textGrain moves provenance from guessing after publication to marking at generation, for eligible EU text. It is useful on longer, lightly edited passages, and weak after rewriting or translation. Watch whether the detector opens up, API opt-in rises, and regulators accept a fading signal.

FAQ

Will every ChatGPT answer in Europe carry the watermark?

No. The company says eligible ChatGPT and Codex text in the EU will be covered over the coming weeks. Coverage can vary by product, model, length, and type of output.

Can anyone check a paragraph for the mark?

Not with a public tool at launch. According to the company, detector access starts with approved researchers and expert organizations.

Does copying and pasting remove the watermark?

Copying alone should not, because the signal is in the word choices. Rewriting, paraphrasing, or translating the text can weaken it substantially, according to the reported tests.

Does a missing watermark prove a human wrote the text?

No. The company cautions against that reading. The text may be short, edited, translated, older, unsupported, or made by another system.

Can API customers outside the EU use it?

Yes, by opting in. The company opened that option worldwide for select models on October 5, 2026, with the setting off by default.

Sources

• AI Weekly — EU rollout summary. https://aiweekly.co/alerts/openai-to-watermark-chatgpt-codex-text-in-eu-under-ai-act
• DEV Community — Article 50 and detection explainer. https://dev.to/yanming_liu_e0e17dd6bf1be/openai-rolls-out-invisible-textgrain-watermarks-for-chatgpt-and-codex-under-eu-ai-act-4301
• AI Daily Post — EU mandate and API opt-in report. https://aidailypost.com/news/openai-adds-optional-ai-watermark-api
• Digital Pulse Brief — textGrain explainer, verified October 6, 2026. https://digitalpulsebrief.com/openai-textgrain-watermark-chatgpt-codex-eu/
• Hero image — original AI illustration created for Red Eye Eagle News. hero-openai-to-add-invisible-textgrain-watermark-to-chatgpt-and-codex-text-in-the-eu-2026-10-06-evening.jpg