OpenAI to use hidden digital watermark in ChatGPT outputs for the EU

Summary

OpenAI is beginning to use 'textGrain,' an invisible statistical watermark, in ChatGPT and Codex outputs within the EU to comply with Article 50 of the EU AI Act. The tool for detecting this watermark will initially be available only to approved researchers.

Why it matters

OpenAI has become the first major generative AI provider to implement word-level statistical watermarking at scale in response to binding EU regulations. This move establishes a concrete template for how a 'machine-readable provenance' system works in practice. It also forces the industry to confront a difficult trade-off: detection rates drop to approximately 80% in short or edited texts, demonstrating that the compliance mechanism is inherently probabilistic rather than absolute.

Highlights

• textGrain embeds an invisible statistical signal in word selection patterns; a separate tool reads this signal.
• Detection rates drop sharply in short texts (~80%), mathematical equations, and after editing.
• OpenAI plans to release the watermark code as open source; however, access to the detection tool will be subject to researcher/institutional approval.
• The compliance deadline for existing players is December 2; the scope covers OpenAI, Anthropic, Microsoft, Google, and Meta.

Critical take

OpenAI's own warning states that performance under ideal conditions does not guarantee reliable detection in daily use. This undermines the core purpose of the watermark. If a simple copy-paste or minor rephrasing can erase the signal, the Article 50 requirement is technically met but effectively hollow. The fact that the detection tool is provided with an approval-based, closed access model also raises questions about who verifies provenance. This power could end up concentrated in a limited number of vetted institutions rather than with the public, whom the regulation is meant to protect.