OpenAI Introduces Invisible Text and Code Watermarking for EU Users
The system, dubbed textGrain, embeds statistical markers into model outputs to meet Article 50 compliance requirements under the EU AI Act.
OpenAI has announced plans to embed invisible digital watermarks into text and code generated across the European Union to satisfy incoming regional transparency rules, according to a report by Engadget (https://www.engadget.com/2277866/openai-will-add-a-digital-watermark-to-text-and-code-generated-in-the-eu/).
The company is deploying a proprietary technology called textGrain, which embeds a subtle statistical pattern into a language model's word selection. A companion detector tool is designed to analyze text samples and identify whether the watermark is present.
The initiative responds to Article 50 of the European Union's AI Act, which requires providers of generative AI systems to ensure text outputs are identifiable in a machine-readable format. While new market entrants are subject to immediate compliance, established operators—including OpenAI, Anthropic, Google, Microsoft, and Meta—face a December 2 deadline. Anthropic rolled out a comparable watermarking initiative earlier this year.
OpenAI stated that textGrain matched or exceeded the performance of alternative approaches, such as Google's SynthID for text. However, the company cautioned that benchmark results in controlled settings do not ensure consistent detection in real-world scenarios.
Detection accuracy declines on shorter text passages, where the success rate drops to approximately 80 percent, as well as on mathematical content. Editing generated text also diminishes detection reliability. OpenAI said it plans to release the watermarking codebase as open source for outside modification.
Watermarking will be enabled by default for eligible ChatGPT and Codex outputs within the EU. Outside the bloc, customers will be permitted to opt in across select, currently unspecified models.
Access to the detection software will initially remain restricted to approved researchers and expert organizations. OpenAI noted that the detector does not identify individual users or expose prompts and underlying conversations.
Sources
Written by
The Company Wire
Inside the companies building what’s next. Reporting on startups, technology, funding and the people shaping them.
