OpenAI to watermark ChatGPT text by default in the EU

OpenAI will automatically add a text watermark to ChatGPT outputs for users in the European Union, according to Ars Technica. The company is also making the feature available elsewhere, but outside the EU it will not be turned on by default.

For creators, marketers and small businesses, the change matters less as a visible label and more as a compliance and trust signal. The watermark is designed to be detected by a separate tool, not noticed by a reader, and OpenAI’s own testing shows there are limits to how reliably it can survive edits.

What changes for ChatGPT users in Europe

OpenAI plans to enable watermarking by default for EU users of ChatGPT, Ars reports. The rollout is expected “in the coming weeks,” according to the company.

The watermarking system is called textGrain. OpenAI has published a technical paper about the method, and Ars describes it as broadly similar to earlier large language model watermarking approaches: the model makes subtle patterns in word choice that are not obvious to humans and are not intended to change the overall quality of the output.

Detection is not public for everyone. OpenAI says it will give access to the detector to a limited group of researchers and organizations, with a request-and-approval process for others over time. That means most users should not expect to paste text into a widely available OpenAI checker and get a result, at least based on what has been announced.

Outside the EU, the same watermarking capability will be offered but left off by default. Ars also reports that, in the API, it appears the feature will remain optional and disabled unless users choose to enable it.

Why Europe is getting the default setting

The reason is regulatory. Ars reports that the move is tied to the EU AI Act, which took effect in August and requires AI-generated content to be marked in a way that another tool can detect.

That distinction is important. This is not the same as a visible caption saying a passage came from AI. Instead, OpenAI’s approach is a hidden statistical signal embedded in generated text.

Ars notes that AI watermarking remains an imperfect area. Existing standards and projects, including SynthID and C2PA, already exist, but the report says these systems can be relatively easy to work around for people with basic knowledge. For more context on one of those efforts, see our coverage of Google’s SynthID website for checking AI-generated media.

How reliable is textGrain likely to be?

OpenAI’s own tests suggest textGrain can work, but not with certainty. Ars reports that the company’s testing showed a 92 percent successful detection rate.

The weakness is editing. According to the same report, changing just 10 percent of the generated text reduced detection success by almost 30 percent. Changing 20 percent of the text could reduce successful detection by almost 75 percent. The report also says detection tends to be weaker for shorter text and translated text than for longer text.

For everyday teams, that creates a practical gray area:

• A long, lightly edited ChatGPT draft may be more detectable.
• A short caption, subject line or translated passage may be harder to identify.
• Substantial editing can weaken the watermark.
• The detector will not be broadly available to everyone at launch, based on OpenAI’s stated access plan.

This means the watermark should not be treated as a perfect authorship test. It is better understood as one compliance mechanism that may help institutions identify some AI-generated text under controlled conditions.

What creators and businesses should watch

If your team uses ChatGPT in the EU, the main near-term impact is process-related. AI-assisted copy, support drafts, internal documentation and marketing text may carry a detectable signal by default. That does not necessarily change the usefulness of the output, but it could matter for clients, schools, publishers or platforms that set their own disclosure rules.

For users outside Europe, OpenAI’s approach is more cautious. The company is making watermarking available, but not forcing it on by default outside the region where regulators require it. That differs from Anthropic, which Ars says introduced text watermarking for its models in August and enabled it globally.

The bigger takeaway is that text watermarking is moving from research topic to product setting, but with clear limitations. Businesses that need reliable AI disclosure should not rely only on hidden watermarks. Policies, human review and explicit labeling may still be necessary, especially when content is edited, translated or repurposed across channels.

For tool buyers comparing assistants, watermarking may become one more governance feature to consider alongside quality, privacy controls and workflow fit. If you are weighing major chatbot platforms, our ChatGPT vs Claude vs Google Gemini buying guide may help frame those broader trade-offs.