OpenAI's text watermark in the EU: what's confirmed, what isn't, and what it means

Short answer: on October 5, 2026, OpenAI announced that it will add an invisible watermark, called textGrain, to eligible ChatGPT and Codex text output in the EU over the coming weeks. In the API the watermark is an opt-in for select models and stays off by default, and the detector is not public. OpenAI says it is responding to the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine-readable way. Detection is weaker on short or edited text.

Most of the coverage is about students and AI detectors. We read OpenAI's own post and the text of the EU AI Act, and below is what's confirmed, what isn't, and what we think anyone running a chatbot should take from it.

What OpenAI announced

In its post, Our approach to EU text provenance rules, OpenAI says its watermark adds "an invisible statistical signal to the model's word choices," so it lives in the text itself. Over the coming weeks, OpenAI will add it to eligible ChatGPT and Codex output for users on all plans in the EU only, and says it is not making text watermarking a global default at launch.

For the API the wording is different. Customers around the world can opt in for select models, and "text watermarking will remain off by default in the API." OpenAI adds that this "lets customers decide how watermarking fits their transparency obligations and the experiences they provide to users," and says it is working with cloud partners to make watermarking available for OpenAI model outputs accessed through their services in the coming weeks.

The detector is not public. Researchers and expert organizations can apply, access is granted case by case, and the tool reports only whether it detects an OpenAI watermark, without identifying the user or revealing prompts. OpenAI also says it plans to make the technology available in open source. We also checked the coverage by TechCrunch and BleepingComputer, and it matches the post.

How well it works, by OpenAI's own numbers

OpenAI is open about the limits. In its evaluations, textGrain matched or exceeded the other approaches it tested, including SynthID for text, but OpenAI adds that strong performance under ideal conditions does not guarantee reliable detection in everyday use. At a 1% false positive rate, the detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages, and detection was substantially lower for content such as math, where there is less flexibility in word choice.

Editing weakens the signal too. In a test on 400-token passages, replacing 10% of the words with synonyms reduced detection from about 92% to 66%, and replacing 25% reduced it to 17%. On quality, OpenAI says it does not see meaningful performance differences with and without watermarking across its benchmarks.

What a watermark can't tell you

OpenAI is just as direct about what a detection result does not mean. It doesn't measure how much a human contributed, it doesn't establish who owns the text or who is responsible for it, it doesn't identify the user, and it doesn't verify accuracy. It also doesn't determine "whether disclosure was required."

A missing watermark proves nothing either. According to OpenAI, the text may be too short, edited or translated for detection to work, it may come from an unsupported model or predate watermarking, or another company's tools may have generated it.

Why this is happening: Article 50 of the EU AI Act

The reason is Article 50 of the EU AI Act, which has two parts relevant to chatbots. Paragraph 1 says providers must make sure AI systems that interact directly with people are designed so that those people are informed they are dealing with an AI system, unless that is obvious from the context. Paragraph 2 says providers of systems that generate synthetic text must mark the output in a machine-readable way so it can be detected as AI-generated.

On timing, legal analyses such as this note from the Cloud Security Alliance agree on two points. Article 50 became binding on August 2, 2026, with no extension for the chatbot disclosure duty, and the Digital Omnibus delayed high-risk rules, not transparency. For systems that were already on the market before August 2, only the machine-readable marking has a grace period, until December 2, 2026. We're not lawyers, and who counts as a "provider" for a bot built on a platform is a question for a lawyer.

What it means if you run a chatbot

For most businesses, the watermark itself is a footnote. It is OpenAI's, it starts in ChatGPT and Codex in the EU, and in the API it is off unless someone opts in. If your bot runs on another provider's model, this announcement doesn't cover it, and we haven't looked at what other providers plan.

The disclosure rule in paragraph 1 is the practical one, and marking text and telling a person are separate things: as OpenAI's own post notes, a watermark says nothing about whether disclosure was required. In a messenger chat people often already know they're talking to a bot, and the "obvious from context" exception exists for that, but we wouldn't rely on it. A one-line first message such as "I'm an AI assistant, and I can pass you to a person any time" costs nothing.

What we'd do this week

• Check which model your bot runs on and whether that provider has announced anything similar.
• Add a short AI disclosure to your first message and keep a way to reach a person.
• If you use OpenAI's API, leave watermarking off unless you have a reason, and re-check OpenAI's post later, since OpenAI says it expects to revisit each part of this approach.
• Don't build flows around removing or hiding marks. The point of the rules is transparency, so we wouldn't spend time on workarounds.
• If you serve consumers in the EU, ask a lawyer which obligations apply to you.

Where this meets Mavibot

In Mavibot, the AI assistant runs on DeepSeek-V4-Flash by default, and you can connect your own model with an API key. The custom model form lets you pick a provider (OpenAI, Anthropic, Google, xAI, Mistral or DeepSeek), optionally name a model, paste your key, and optionally set an API URL. The form also says that requests go directly to the provider's API and are billed at its rates.

So if you connect OpenAI, your bot's replies are OpenAI API output, and by OpenAI's announcement the watermark there is off unless someone opts in. Nothing in the form turns it on. What OpenAI's post doesn't spell out is how the opt-in works technically, as a request setting or as something you switch on for your account, so we can't tell you what would happen if you enabled it for the same key. If you use OpenAI, check that in OpenAI's own documentation before changing anything. The first message is yours to write in the funnel builder, so the disclosure line takes a minute.

If you want an assistant that tells people what it is, answers from your own data and hands over to a person when needed, Mavibot's chatbot builder lets you set that up without code, with the default model or your own API key. For the setup itself, our AI shopping assistant guide walks through roles and instructions. Start free.

Frequently asked questions

What did OpenAI announce about text watermarks?

On October 5, 2026, OpenAI announced that it will add an invisible watermark, called textGrain, to eligible ChatGPT and Codex text in the EU over the coming weeks. API customers worldwide can opt in for select models, and the watermark stays off by default in the API.

Can anyone check whether a text carries OpenAI's watermark?

Not yet. OpenAI says detector access is initially limited to approved researchers and expert organizations, granted case by case.

Does OpenAI's text watermark affect my chatbot?

By default, no. According to OpenAI's announcement, the EU rollout applies to ChatGPT and Codex, while in the API text watermarking is an opt-in that is off by default.

Can a watermark be detected in short chatbot replies?

It is harder. OpenAI's tests found the watermark in about 80% of 200-token passages versus about 95% of 400-token passages, and replacing 25% of the words with synonyms lowered detection from about 92% to 17%.

Does the EU AI Act require my chatbot to say it is an AI?

Article 50(1) requires providers of AI systems that interact directly with people to ensure those people are informed they're interacting with an AI system, unless that is obvious from the context. This is not legal advice, so check how it applies to your case.

Can I use my own OpenAI, Claude or Gemini key in Mavibot?

Yes. In the AI assistant settings, the Custom model option lets you choose OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini), xAI (Grok), Mistral or DeepSeek API, and paste your API key. Requests go directly to the provider and are billed at its rates.

What model does Mavibot's AI assistant use by default?

DeepSeek-V4-Flash, unless you switch to a custom model with your own API key.