OpenAI Text Watermark Goes Live in the EU as the AI Act's De

OpenAI has begun the phased rollout of an OpenAI text watermark for eligible ChatGPT and Codex output inside the European Union. The identifier is invisible to readers and machine-readable by design: no label, banner or disclaimer attaches to the text. The company said on October 5 that the mark will reach eligible accounts over the coming weeks, tied to the transparency obligations of the EU AI Act.

Two defaults define the rollout, and they point in opposite directions. Watermarking switches on automatically for eligible ChatGPT and Codex users in the EU across all subscription tiers. On the API, the same capability sits behind an opt-in that stays off unless a developer enables it. Coverage is geographic, so accounts outside the EU are untouched in this phase. The split is the most consequential part of the announcement: the API is where enterprise text production runs at volume, and a default that stays off leaves that output unmarked unless a customer chooses otherwise.

Surface	Default state	Coverage
ChatGPT, eligible plans	Watermark on	EU users
Codex	Watermark on	EU users
API, eligible models	Watermark off, opt-in required	EU users
Detection tool	Closed	Approved researchers

What the Watermark Actually Does

OpenAI describes the mechanism as a statistical signal applied to generated tokens rather than an added string of characters. It calls the system textGrain and says it performed at or above the level of competing approaches, including the watermarking work Google has published. Nothing about the output changes for the person reading it.

The design choice shapes how the mark travels. Metadata-based provenance, of the kind attached to some image formats, disappears the moment a file is re-encoded or pasted into another application. A statistical pattern woven into the wording travels with the words themselves, so a copied paragraph still carries it. That is the argument for the token-level route, and it is also why the signal holds up against copying but not against rewriting.

Detection is not open to everyone. OpenAI is restricting access to its detection tool to approved researchers. Publishers, fact-checkers and the general public cannot run the check themselves, which keeps the verification capability inside a vetted group. For a newsroom trying to establish whether a document is machine-written, the mark exists but the means to check it does not.

OpenAI also sets two boundaries on what a result means. A positive detection does not establish that a whole document was machine-written, since a single machine-generated passage inside a longer human-authored file can trip the signal. A negative result does not establish human authorship either, because the watermark survives only so much editing. Read together, the two limits mean the mark cannot carry the weight of a binary AI-or-human label.

The Regulatory Clock Behind the Rollout

Article 50(2) of the EU AI Act requires providers to ensure that AI outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Those transparency rules began applying on August 2, 2026. Systems already on the market before that date received a transition window, and the European Commission has set December 2, 2026 as the point at which the marking and detection obligation bites for them.

OpenAI sits in that second group. So do Anthropic, Microsoft, Google and Meta, all of which face the same December deadline for products placed on the market earlier. Newer entrants were covered from the start. That grouping explains the timing: the deadline lands on vendors already shipping, and OpenAI is one of them, so the announcement reads as a compliance step rather than a product launch.

The Commission has also produced a voluntary Code of Practice on Transparency of AI-generated Content, which lays out practical measures for meeting the Article 50 requirements. Around 190 companies have signed on. Signing is not the same as shipping: the code describes methods, while the December date sets the enforcement horizon.

Why the API Default Matters More Than the Consumer One

For a ChatGPT subscriber in the EU, the change is invisible in the literal sense. The text looks the same, reads the same and is copied and pasted the same way. The watermark rides along inside the token sequence.

The API decision carries more commercial weight. Enterprise pipelines that call OpenAI models to draft documents, generate product copy or populate internal tools will produce unmarked output unless someone deliberately turns the feature on. That creates a split outcome: a consumer chat in Munich is watermarked by default, while a corporate application running the same underlying models in the same jurisdiction is not, unless its engineering team opts in. Put plainly, the regime covers the least consequential output automatically and the most consequential output on request.

For organisations that need to demonstrate provenance downstream, opt-in becomes a configuration decision with compliance consequences. It also raises the practical question of who owns the obligation once text leaves the model and enters a customer's own product. OpenAI's default places that choice with the API customer rather than with OpenAI itself.

Procurement teams should expect the question to move into vendor reviews. If an organisation relies on generated text for regulated disclosures, marketing claims or customer communications, the ability to produce a provenance signal on demand turns into a checklist item in the way data-processing terms did. Providers that gate detection behind researcher access leave their customers without the means to run that check independently.

Where the OpenAI Text Watermark Can Fail

Watermarking is strongest on text that leaves the model untouched. Paraphrasing, translation, heavy editing or mixing machine passages into a larger human draft all reduce the signal's reliability, and OpenAI acknowledges the mark degrades under editing. Anyone motivated to strip it has an obvious route: rewrite the output, and the statistical pattern weakens.

The restricted detector compounds the problem from the other direction. A provenance system that only approved researchers can query is difficult to audit in public. The trade-off is deliberate, since a widely available detector is also a widely available target for reverse engineering, but it leaves publishers and platform trust teams relying on a signal they cannot independently verify.

There is a wider asymmetry in the regime. The obligation falls on providers placing systems on the market, so text produced through a hosted service can be marked at the source. Text generated by open-weight or self-hosted models that a company runs on its own infrastructure sits outside that arrangement, and nothing in OpenAI's rollout changes it. A compliance officer who assumes that all AI text circulating inside the EU carries a mark is working from a false premise. The failure modes stack: the mark weakens under editing, the check stays closed to the public, and a large share of EU-facing AI text never enters a regime that would apply it at all.

What Competitors Face

Google, Microsoft, Meta and Anthropic are working against the same December date, and the voluntary code gives them a shared template for how to comply. Any of them could match or undercut OpenAI's approach before then. OpenAI has established a public position first, with a consumer-default-on model and a researcher-gated detector, which sets a reference point rivals will be measured against. The meaningful comparison is not which vendor ships first but which vendors let anyone outside a vetted group verify the result, since a closed check produces the same evidentiary gap regardless of whose model wrote the text.

The scoping also leaves a global question open. OpenAI has limited this phase to the EU rather than switching watermarking on worldwide. Other jurisdictions are moving on content provenance at different speeds, and a provider that maintains separate regimes per region carries more engineering overhead than one that applies a single global standard. Extending the watermark beyond the EU would simplify that, but it would also export a detection limit to markets with no legal requirement to accept it.

The Practical Read

For EU consumers, nothing changes in daily use. For developers and platform teams building on OpenAI's API, the decision point is now: if downstream provenance matters, opt in before the December deadline rather than after a compliance review raises it.

For anyone who evaluates text for a living, the watermark is a signal with a known error rate and a closed detector. Treating it as proof in either direction would be a mistake. It is one input among several.

Two milestones are worth tracking from here: whether detection access widens beyond approved researchers, and whether OpenAI carries the watermark outside the European Union once the December obligation is met.

Why this matters

The EU AI Act has turned content provenance from a research problem into a product requirement with a date attached. For providers, the December deadline sets a floor that every major model vendor must clear in the same window, which makes watermarking a baseline feature rather than a differentiator. For buyers, the lesson is narrower and more useful: provenance guarantees follow configuration, and configuration is a choice someone in the organisation has to make. Clearing the compliance box and earning trust in a given piece of text are two different tests, and this rollout settles only the first.

Sources

Our approach to EU text provenance rules

Photo by Brecht Corbeel on Unsplash

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.