OpenAI Introduces Text Watermarking to Comply with EU AI Act Regulations
OpenAI has outlined its new text provenance strategy, introducing opt-in watermarking for global API users and rolling out invisible watermarks for EU users.
Content provenance plays a crucial role in helping individuals determine where information originates, how it was developed or modified, and whether it carries signals tied to specific language models. While audio and image identification tools have been publicly available for some time, OpenAI has now detailed its strategy for text watermarking. This rollout aligns with the European Union Artificial Intelligence Act, which mandates that generative AI creators make machine-readable identifiers for generated text.
Text watermarking and detection remain in early development stages, presenting distinct technical boundaries and differing opinions regarding their practical benefits. Because of these challenges, the organization is rolling out a phased strategy that respects current limitations while meeting regulatory obligations in the European Union.
Implementation and Global Availability
Starting immediately, API customers globally can choose to enable text watermarking for selected models, though the feature remains turned off by default within the API. In the coming weeks, an invisible signal will be integrated into eligible ChatGPT and Codex text outputs generated inside the European Union.
To better understand the technology, applications are now open for the text watermark detector. Initial access is restricted to approved researchers and expert organizations capable of assisting with evaluation and enhancement. Meanwhile, verification mechanisms for audio and images—such as the openai.com/verify portal and the Content Provenance API—remain fully accessible to organizations verifying media files.
The underlying technology, named textGrain, embeds an invisible statistical signal into word selections made by the model. A specialized detector scans for this pattern to determine if a text segment contains the watermark. OpenAI plans to publish the technical report and release the technology in open source so others can build upon it. Evaluations show that textGrain matches or exceeds alternative approaches like SynthID for text.
What it means for developers
Developers working with generative AI must account for the inherent limitations of text watermarking. Testing demonstrates that shorter or highly constrained text is much harder to detect accurately. For instance, detection rates drop significantly in mathematical content compared to broader subjects like psychology, due to reduced flexibility in word choice. Furthermore, manual or automated editing easily weakens the signal; replacing even a fraction of words with synonyms drastically reduces detection success.
API users can opt in to these watermarked outputs to align with transparency standards and regional obligations. For those building applications requiring multi-model access, developers can try top AI models cheaply through one API at https://apixoai.online. However, developers should note that watermarks do not measure human contribution, prove authorship, establish ownership, verify factual accuracy, or tie text back to specific user accounts and prompts.
Performance and Broader Ecosystem
Benchmarks evaluating the frontier model Astra indicate no meaningful performance differences between unwatermarked and watermarked text outputs across various standardized tests, including AutomationBench and GPQA Diamond.
No single provenance method is sufficient on its own, prompting a layered strategy that combines open standards, resilient watermarks, and verification tools. OpenAI continues to use Content Credentials for images, C2PA compliance, and SynthID watermarks for multimedia files. As standards and regulatory demands shift, the organization intends to adapt its text provenance framework and expand detector access responsibly over time.
Source: Our approach to EU text provenance rules — OpenAI News. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
