OpenAI Announces Provenance Markers for Some Text and a Restricted Detector

On October 5, 2026, OpenAI announced optional text watermarks for certain API models and a subsequent rollout of invisible markers on some ChatGPT and Codex text in the European Union. OpenAI’s post also explains that access to the associated detector will be restricted. For content teams, the news opens a conversation about provenance and transparency, but does not offer a definitive way to prove who wrote a text or how much a person contributed.

What provenance markers were announced?

Starting October 5, OpenAI API customers worldwide can opt in to text watermarks on certain models; the option is off by default, according to the company’s explanation. Separately, OpenAI said it will add invisible markers to some ChatGPT and Codex text in the European Union over the following weeks. That wording describes a rollout announced for the future, not availability that has already been completed in all cases.

The distinction between the two approaches matters for editorial workflows. One applies to API customers who can choose to enable the marker on certain models. The other refers to some ChatGPT and Codex text in the EU. They should not be treated as a single, universal feature, nor should anyone assume they cover all text generated or edited with AI.

What can a watermark prove?

OpenAI clarifies that the marker does not determine how much a person contributed to the text, nor does it prove authorship, accuracy, or responsibility, according to the announcement. A provenance marker therefore does not replace editorial review, fact-checking, or internal policies on when to disclose the use of AI tools.

The company also warns that editing the text can weaken detection. This limits any conclusions someone might try to draw from a positive or negative result. In practice, a team should not use detection as an automatic verdict on a person’s conduct, the truthfulness of a post, or the quality of a piece. The source describes provenance mechanisms, not a complete certification of the creative process.

Who can use the detector?

Access to the detector is being opened to approved researchers and organizations, not the general public, OpenAI reports. The available research provides no basis for presenting the tool as something any brand can activate in its everyday workflow. Those who believe they need access should consult the criteria and follow the process OpenAI publishes, rather than basing an editorial procedure on assumed availability.

It is also worth distinguishing the detector from the API watermarking options. The fact that a customer can opt in to a marker on certain models does not mean everyone gets access to the detector or that its result will resolve authorship disputes. They are separate parts of the announced provenance approach.

What should a content team review?

A team can agree on how it internally identifies AI-generated or AI-assisted drafts, what human review each post requires, and what disclosures its brand needs. The decision should be based on clear policies, not on the expectation that an invisible marker will determine authorship, accuracy, or responsibility. OpenAI also does not present its markers as a measure of how much each participant contributed.

HeyMark has not yet tested these markers or the detector. Its relationship to the editorial workflow is more limited: it lets teams plan and publish content on supported social networks, collaborate, and review messages and performance; it does not verify provenance or authorship. To organize posts created with the help of AI tools using a clear workflow, see the guide to connecting Claude with HeyMark.