AI Watermarks Face a Compliance Problem Europe Cannot Ignore

The watermark may not survive the text.

The European Union now requires AI-generated text to carry a watermark so people and institutions can identify it as machine-produced. OpenAI’s textGrain, released in October and developed to comply with the new EU law, is meant to provide that signal—but tests show the watermark often disappears after additional editing or when the text contains fewer than 150 words.

That creates a regulatory problem with an irritatingly simple shape: Europe requires watermarking, but the law does not define a clear legal benchmark for how well the system must work. Without that benchmark, it remains uncertain whether textGrain meets the EU’s transparency rules, even as compliance could cost OpenAI significantly and become harder as models grow more sophisticated.

How textGrain hides a signal in ordinary writing

AI models calculate the probabilities of different possible next tokens from the text already written. They select one token, repeat the calculation, and continue until the response is complete—a statistical process that leaves room for a hidden intervention.

A watermark uses a secret key to generate pseudorandom values that influence which token the model suggests. That creates a statistical pattern in the finished text, which a detector holding the same key can search for without showing anything to ordinary readers.

textGrain uses entropy—the measure of how unpredictable a model’s word choices are—as a resource that can be budgeted during watermarking. Developers specify how much unpredictability they want to allow, and textGrain adjusts to that target, giving them more control over the trade-off between a detectable signal and variation in the model’s output.

OpenAI says its detector has a false-positive rate of 1 percent. In practical terms, it is designed to wrongly flag about one in 100 unwatermarked passages. That is a measurable limit, not a magic truth machine.

A positive result does not show how much a person contributed to a passage. A negative result does not prove that a human wrote it. The system can identify a statistical pattern—or fail to find one—but it cannot reconstruct authorship from the text alone.

Anthropic takes a broader route

Anthropic announced in August that it would integrate a text watermark into its models, responding to the EU regulation that requires the technology across generative AI tools. The company is now rolling out the change more broadly across Claude, including Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6.

The rollout covers all users globally, despite its European regulatory trigger, and users cannot opt out. The system is described as leaving an invisible trail that shows whether text was written with Claude, while the markings are also described as entirely visible. Either way, normal users cannot check whether their writing would be flagged.

At the moment, only eligible organisations defined under EU law can detect the watermark, including regulators, law enforcement, journalists, and researchers. Companies that must verify watermarking under the EU’s AI Act can also use the detection system.

Anthropic has accepted that shorter text may not contain enough of the pattern to be detected. Users have also criticized the system because proofreading or fact-checking text with Claude could trigger detection, while bad actors could rephrase the material with another AI system to hide the watermark.

The company has pushed back against claims that watermarking will degrade output quality. Yet on Opus 4.6 and Claude Sonnet 4.6, changing the settings that control word selection through the API will remove the watermark. Users can also select older models that have not received the changes, including Opus 3, first released in 2023.

Other companies have indicated that they will comply with the EU rules. OpenAI has added textGrain to ChatGPT, though it will be added automatically only in Europe.

The larger issue is not whether companies can embed a signal. They can. The issue is whether that signal survives the messy conditions of real writing—editing, short passages, model changes, and deliberate attempts to remove it.

Europe’s rules are trying to create trust through detection. But if a watermark vanishes below 150 words or after routine editing, compliance risks becoming a box-ticking exercise with a secret key attached. Technology has many talents. Being reliable under pressure remains an optional upgrade.

Stay connected via Google News

Follow us for the latest AI News