OpenAI Unveils Phased Approach to EU Text Watermarking

Overview

OpenAI announced a phased strategy to implement text watermarking in compliance with the European Union’s AI Act, emphasizing the current technological limitations of such tools. The company revealed that it will begin offering opt-in watermarking for API customers globally while introducing invisible watermarks to eligible ChatGPT and Codex users in the EU in the coming weeks.

The initiative is framed around promoting transparency while acknowledging that text provenance technologies remain in early stages of development. OpenAI stated that its approach balances regulatory requirements with a clear understanding of what watermarks can and cannot achieve, aiming to provide accurate signals about content origin without overstating the reliability of detection methods.

Phased Rollout and Regulatory Compliance

In response to the EU AI Act, which mandates that generative AI providers make generated text identifiable in a machine-readable format, OpenAI has structured its implementation in three primary phases. This staggered approach allows the company to gather feedback and refine detection capabilities before broader deployment.

Starting immediately, API customers worldwide will have the option to enable text watermarking for select models. However, the feature will remain disabled by default, giving developers control over whether their generated outputs carry the invisible statistical signals.

Over the next several weeks, OpenAI will deploy invisible watermarks to eligible ChatGPT and Codex text outputs within the European Union. This move directly addresses the EU’s specific regulatory requirements while maintaining a cautious stance on the technology’s current effectiveness.

The company is also opening applications for access to its text watermark detector, though initial availability will be restricted to approved researchers and expert organizations. These selected partners will help evaluate the technology’s reliability and identify responsible use cases before wider public access is considered.

"Text watermarking and detection remain early technologies with significant limitations, and views about their benefits and responsible uses are still developing." — OpenAI Safety Team

Existing verification tools for images and audio, including the openai.com/verify web tool and the Content Provenance API, will remain publicly accessible. These established systems continue to operate independently of the new text watermarking initiatives, providing organizations with methods to determine whether media files were generated by OpenAI systems.

Technical Performance and Limitations

OpenAI’s text watermarking technology, internally known as textGrain, operates by embedding an invisible statistical signal into the model’s word selection process. The accompanying detector analyzes these signals to assess whether a passage contains an OpenAI-generated watermark, with detailed technical specifications available in an upcoming report.

Evaluations indicate that textGrain matches or exceeds the performance of competing approaches, including Google’s SynthID for text. However, strong benchmark results under ideal conditions do not guarantee reliable detection in real-world scenarios, prompting the company to adopt a measured rollout strategy.

Detection challenges vary significantly based on text length and content type. At a target false positive rate of one percent, the detector identified watermarks in approximately 80 percent of 200-token passages compared to 95 percent of 400-token passages for psychology-related content. Detection rates dropped substantially for mathematics topics, where vocabulary flexibility is inherently limited.

Editing poses another major challenge to watermark reliability. In evaluations of 400-token passages, replacing just 10 percent of words with synonyms reduced detection rates from roughly 92 percent to 66 percent. When 25 percent of words were replaced, detection fell to only 17 percent, demonstrating how easily human intervention can obscure watermark signals.

These limitations explain why initial detector access remains restricted to vetted researchers and expert organizations. The company believes that controlled evaluation by qualified partners is essential for understanding the practical reliability of watermarking technology and identifying appropriate applications.

Impact on Output Quality and Interpretation Guidelines

OpenAI conducted extensive benchmark testing to ensure that watermarking does not degrade the quality of its AI-generated text. Comparisons between watermarked and unwatermarked outputs across multiple evaluation suites for the Astra model showed no meaningful performance differences.

Benchmark Results for Astra Model

• Artificial Analysis Intelligence Index: 49.57 points (unwatermarked) vs. 49.76 points (watermarked)
• AutomationBench: 34.09% vs. 34.86%
• DeepSWE v1.1: 72.80% vs. 71.68%
• Terminal-Bench 4.0: 53.90% vs. 56.06%
• BrowseComp: 87.92% vs. 87.35%
• GPQA Diamond: 94.44% vs. 93.94%

Despite these stable performance metrics, OpenAI stressed that watermarks convey limited information about the nature and origin of content. The company published detailed guidelines explaining what users should and should not infer from watermark detection results.

A watermark does not measure human contribution to a passage. While it may indicate that an OpenAI system generated or processed part of the text, it cannot quantify the extent of human judgment, editing, or creativity involved in the final output.

A watermark does not establish ownership or legal responsibility. It provides no information about intellectual property rights, lawfulness of use, disclosure requirements, or accountability for the content’s implications.

Furthermore, watermarks do not identify users or associate texts with specific accounts, prompts, or conversations. The system simply detects the presence of OpenAI’s statistical signature without linking it to individual actors or specific generation events.

"The absence of a detected watermark does not prove human authorship. Text generated with OpenAI tools may be too short, edited, or translated for detection to work reliably." — OpenAI Transparency Guidelines

This distinction is crucial for preventing misinterpretation of watermark results. A negative detection could result from short text length, extensive editing, translation into unsupported languages, use of pre-watermarking models, or generation by competing AI systems.

Future Development and Open Source Commitment

OpenAI plans to make its text watermarking technology available as open source in the future, allowing other developers and researchers to build upon the foundation. This commitment aligns with the company’s broader strategy of contributing to the development of responsible AI tools while maintaining transparency about current capabilities.

The technical report detailing how textGrain operates will receive updates in the coming weeks, providing additional insight into the watermarking methodology and detection algorithms. Researchers and organizations interested in participating in the initial detector access program can submit applications through OpenAI’s designated portal.

As the EU AI Act’s provisions take effect, OpenAI’s phased approach demonstrates a willingness to engage with regulatory requirements while advocating for realistic expectations about watermarking technology. The company continues to emphasize that provenance tools are evolving rapidly and that ongoing collaboration with the research community will be essential for improving reliability and expanding appropriate use cases.