ChatGPT Text Watermarking Is Coming to Europe Under the EU AI Act

OpenAI is preparing to add invisible watermarks to text generated by ChatGPT and Codex in the European Union, giving machines a new way to identify content that may have originated from its AI systems.

The change will roll out over the coming weeks to eligible ChatGPT and Codex users across all plans in the EU. OpenAI is calling the underlying technology textGrain, and unlike a visible AI label or metadata tag, it works by subtly changing the statistical pattern of the words a model chooses while generating a response.

A person reading the text should not notice anything unusual, while a compatible detector can look for the statistical pattern the model leaves behind.

OpenAI's text provenance announcement explains how the system is being introduced in response to European transparency requirements.

How ChatGPT Text Watermarking Actually Works

The watermark is not a hidden sentence, special character or invisible space inserted into a ChatGPT response.

Instead, textGrain changes the probabilities when the model selects between possible words or word parts. Across a sufficiently long passage, those choices can create a statistical signature that OpenAI’s detector can recognise.

That makes the watermark part of the wording itself.

Copying and pasting an unchanged response does not automatically remove the signal because there is no separate metadata file that has to travel with the text. At the same time, the approach is not permanent. Heavy editing, paraphrasing or translation can weaken the pattern enough to make detection less reliable.

OpenAI specifically warns that text watermarking should not be treated as definitive proof of who wrote something.

A detected watermark may indicate that an OpenAI model generated or processed the text. Still, it cannot reveal who used the model, how much human editing took place, who owns the writing, or whether the information is accurate. OpenAI’s explanation of provenance signals provides more detail on those limitations.

The opposite is also true. Failing to detect a watermark does not prove that a human wrote a piece of text.

Europe Is Getting the Watermark First

OpenAI is not making text watermarking a worldwide ChatGPT default, at least for now.

Eligible ChatGPT and Codex outputs in the European Union will receive the watermark as the system rolls out over the coming weeks. API customers outside Europe have a different option: developers can choose to enable watermarking for supported models, but the feature is not necessarily enabled by default.

OpenAI says it is also working with cloud partners to make the technology available when its models are accessed through their services.

The regional rollout reflects a regulatory requirement rather than a decision to watermark every OpenAI text response globally.

Article 50 of the EU AI Act requires providers of systems that generate synthetic text and other media to use machine-readable marking that enables detection of AI-generated or manipulated content where technically feasible. European Commission guidance on Article 50 explains how those transparency obligations apply.

That helps explain why text provenance has moved from a research problem to a product requirement for major AI companies operating in Europe.

A Watermark Is Different From an AI Detector

The distinction between watermarking and conventional AI detection is important.

Most third-party AI detectors examine writing after it has been produced and estimate whether its vocabulary, sentence structure, or other characteristics resemble machine-generated text. Those systems do not necessarily have access to a signal the model intentionally embeds.

textGrain takes a different approach. OpenAI deliberately creates a statistical pattern during generation and then looks for that pattern later.

In theory, that provides a stronger provenance signal than simply guessing whether writing sounds like AI.

It still does not make detection certain.

OpenAI acknowledges that its detector can miss watermarked text and can sometimes produce false positives. That is one reason the company is not releasing the text watermark detector publicly at launch.

Instead, approved researchers and expert organisations can apply for access so OpenAI can evaluate how the technology performs outside controlled testing.

The detector will report whether an OpenAI watermark appears to be present. OpenAI says it will not identify the user or expose the prompts or conversations behind the text.

Short Text Remains Difficult to Identify

Text watermarking becomes harder when there are not enough words for a statistical pattern to emerge.

OpenAI notes that very short outputs may not provide enough text for reliable watermark detection, while certain kinds of content such as code can present additional challenges.

This limitation is practical. A long passage gives the model many opportunities to make word choices that collectively form a detectable signal. A short answer may not contain enough variation to establish the same pattern with confidence.

Language can also affect detection.

OpenAI has tested textGrain across multiple European languages and found that detection performance can vary. The company can adjust the watermark strength to improve detectability while trying to preserve response quality.

OpenAI says its testing has not found a meaningful reduction in model quality from enabling the watermark, and the effect on generation speed is negligible. OpenAI’s technical explanation of textGrain provides more information about how the system works.

Editing Can Weaken the Signal

Perhaps the biggest limitation is that text is easy to change.

An image watermark can be designed to survive some resizing, compression or editing. Text can be rewritten sentence by sentence, translated into another language, summarised, or passed through another model.

Each of those transformations may alter the statistical pattern on which textGrain depends.

OpenAI openly acknowledges this problem. Substantial rewriting or translation can reduce the likelihood that its detector will find the original watermark.

That means the technology is better understood as a provenance signal than as a universal AI-writing test.

This difference matters in schools, workplaces and publishing. A watermark result alone cannot establish plagiarism, misconduct or authorship.OpenAI’s own description of the system makes clear that detecting a watermark does not tell an investigator how much of the final document a person wrote or edited.

The EU AI Act Is Changing How AI-Generated Content Is Identified

OpenAI’s decision reflects a broader shift in the European AI market.

The EU’s transparency rules require providers to make AI-generated or manipulated content machine-detectable in covered circumstances. The law also imposes separate disclosure requirements on certain AI users, including publishers of AI-generated text on matters of public interest when there has been no human review or editorial control.

There is an important exception: public-interest text that has undergone human review or editorial control and for which a person or legal entity accepts editorial responsibility does not face the same disclosure requirement under that part of Article 50. EU AI Act Article 50 sets out those transparency obligations.

The rules are therefore more nuanced than simply requiring every piece of AI-assisted writing in Europe to display an “AI-generated” label.

For OpenAI, however, the provider-side requirement creates a technical problem it must solve at the model level. textGrain is its answer.

The broader shift also places OpenAI alongside other AI companies experimenting with provenance systems. TechAmerica.ai has previously examined invisible watermarking in Claude-generated text, showing how provenance is becoming a wider industry concern rather than an OpenAI-only issue.

OpenAI Is Taking a Cautious Approach to Detection

The company appears deliberately reluctant to present watermarking as a perfect solution.

OpenAI could have launched a public website where anyone could upload text and receive an AI-generated or human-written verdict. It has chosen not to do that with textGrain.

The restricted detector rollout reflects the potential consequences of a false result. A system that incorrectly labels human writing as AI-generated could create problems in education, employment, journalism or other settings where authorship matters.

OpenAI instead plans to give approved researchers and expert organisations initial access while gathering evidence about how the watermark performs in real-world conditions.

That approach also distinguishes text provenance from some image and audio verification systems, where embedded metadata or other provenance signals can sometimes be inspected more directly.

What the Change Means for ChatGPT Users

Most people using ChatGPT in Europe should not notice a visible change in the text they receive.

There will be no watermark logo at the end of a paragraph, no hidden characters attached when text is copied, and no unusual punctuation inserted for detection. The signal comes from themodel’ss word-selection pattern.

The bigger difference exists behind the scenes.

For the first time, ordinary ChatGPT text generated in the EU will increasingly carry an intentional OpenAI provenance signal that an authorised detector may be able to identify later.

It will not solve the wider debate over AI-generated writing. Watermarks can be weakened, detection can fail, and the signal says little about how much human work went into a final document.

But the rollout establishes something that did not previously exist at this scale: a built-in mechanism designed to leave evidence of AI generation inside the text itself. Europe’ss AI rules are pushing that technology out of the laboratory and into everyday ChatGPT responses. What happens next will depend on whether invisible watermarking proves durable enough to be useful without being treated as more conclusive than it really is.