OpenAI is preparing to introduce invisible watermarks in ChatGPT-generated text across the European Union to comply with the EU AI Act, which requires AI-generated content to be identifiable.
The technology, called textGrain, will gradually roll out to eligible ChatGPT and Codex outputs in the coming weeks. Instead of inserting hidden characters or visible labels, it subtly influences the words selected by AI, creating statistical patterns that specialized detectors can recognize.
These watermarks are designed to remain detectable when text is copied and pasted without modification. However, substantial editing, paraphrasing, or translation can weaken or remove the signal.
OpenAI says the watermark contains no personal information about users, their organizations, or their prompts. The company also plans to offer optional watermarking to API customers worldwide.
Access to the detection tool will initially be limited to approved researchers and expert organizations, as the technology is not completely reliable. Short responses, code, mathematical answers, and highly factual content are particularly difficult to identify accurately.
Under the EU's transparency guidelines, text shorter than 200 tokens and code snippets are exempt from watermarking requirements.
OpenAI claims textGrain has no significant impact on response quality or generation speed.
The move follows the company's unsuccessful attempt to detect AI-written text in 2023, when its classifier correctly identified just 26% of AI-generated samples during testing.
With textGrain, OpenAI is taking another step toward making AI-generated writing traceable, although the technology cannot reliably determine how much of a finished text was written or edited by a human.