OpenAI's ChatGPT Gets Watermarking in EU to Meet AI Act Rules

OpenAI is rolling out automatic watermarking for ChatGPT outputs across the European Union, marking a significant step toward regulatory compliance. The watermarks will be enabled by default for EU users but available as an optional feature elsewhere.

Regulatory Driver#

The watermarking push stems directly from the EU AI Act, which became enforceable in August and mandates that AI-generated content carry detectable markers. This requirement reflects growing regulatory pressure to help distinguish machine-generated text from human writing as AI systems become more sophisticated and widespread.

How It Works#

OpenAI's proprietary watermarking system, called textGrain, embeds subtle patterns into word choices that remain invisible to readers while preserving output quality. The approach aligns with other watermarking solutions already in use, such as Google's SynthID and the C2PA standard, but relies on a specialized detection tool to verify authenticity. The company has published a technical paper explaining the methodology.

Access and Verification#

OpenAI plans to restrict detector access to a curated set of researchers and organizations initially, with a formal approval process for expanding access over time. This gatekeeping approach contrasts with more open-source alternatives but gives the company control over how detection tools are deployed.

The Limitations Challenge#

Existing watermarking standards, while technically sound, remain vulnerable to circumvention by users with basic technical knowledge. Whether OpenAI's textGrain proves more robust against such attempts remains to be seen—a critical question as regulators expect these markers to provide meaningful assurance of AI involvement in content creation.