OpenAI has announced plans to introduce invisible digital watermarks for text content generated by its popular ChatGPT chatbot and the coding tool Codex. This initiative will initially target users within the European Union, specifically to comply with the forthcoming EU AI Act.
The EU AI Act mandates that artificial intelligence companies ensure their AI-generated content is identifiable in a machine-readable format. OpenAI's new watermarking system, dubbed 'textGrain,' is designed to meet this requirement. According to the company, these watermarks are embedded subtly within the text, making them imperceptible to the human eye but detectable by specialized AI tools using a secret key.
How OpenAI's Text Watermarking Works
The 'textGrain' method subtly influences the word choices ChatGPT makes during text generation. This allows an AI detector, when provided with the text and the secret key, to verify if the content originated from an OpenAI system. The company emphasizes that while the watermarks are designed to be persistent, certain modifications can impact their detectability.
Limitations and Challenges
OpenAI acknowledges that editing the generated text can make the watermark harder to detect. For instance, altering just 10% of the words to synonyms can reduce detection accuracy from approximately 92% to 66%. The system also faces challenges with very short text passages, mathematical answers, and translated content. Due to these limitations, initial access to the detection tools will be provided to approved researchers and expert organizations to evaluate reliability and responsible use.
It's important to note that these watermarks can indicate that an OpenAI system generated or processed a portion of a passage, but they do not quantify the amount of human judgment, editing, or creativity involved. The watermarking system is not expected to impact the performance of the AI models themselves.
Rollout and Future Plans
The 'textGrain' watermarking system is slated to roll out to ChatGPT users in the EU over the coming weeks. While the initial focus is on European compliance, OpenAI is also making the technology available globally to API customers as an opt-in feature for certain models, effective immediately. This move follows similar steps by other AI developers, such as Anthropic, which released a watermarking system for its Claude AI texts after the EU AI Act came into force in August 2026.