OpenAI Plans AI Text Watermarking Under EU Rules

What happened

OpenAI plans to add a digital watermark to text generated by ChatGPT and Codex within a few weeks. The change is a response to EU artificial intelligence rules, introduced two months earlier, which require AI-generated content to be identifiable.

The watermark will be available worldwide, but it will be enabled by default only within the EU. Users in other regions can therefore ignore it. Competitor Anthropic introduced a similar feature for its AI tools in August.

According to TechCrunch, the watermark appears to be possible to edit out, so its practical value remains uncertain. The story was carried by Computerworld and originally appeared in Computer Sweden.

Why it matters

This is a provenance story as much as a regulatory one. As AI tools become standard in business writing, customer communications and internal documents, organisations will need a clear answer to whether a piece of text was machine-generated. A watermark built into the model's output is one attempt to provide that answer at the source, rather than relying on downstream detection.

For leaders in digital transformation, the detail that matters most is the default setting. A capability that exists everywhere but is switched on only in one jurisdiction creates two different experiences for users, and a different compliance posture for the organisations that deploy these tools across regions.

By the numbers

• Within a few weeks: the window OpenAI gave for introducing the watermark.
• Two months: the time since the EU's new AI rules were introduced, which the company cites as the trigger.
• August: the month Anthropic introduced a comparable feature for its AI tools.

The Renascence take

Most readers will assume a watermark is a trust signal. In practice, when it is off by default outside the EU and can reportedly be removed, it works more like a compliance artefact than a reassurance for the people reading the text.

The underlying principle is that defaults shape behaviour far more than availability does, so a disclosure that is optional by default is rarely seen as a meaningful signal. Customer-obsessed operators should not outsource transparency to a model vendor's watermark. They should design their own disclosure into the experience, make it visible and understandable at the point of reading, and test whether it still holds up when content is edited or passed between channels.