Big Update For ChatGPT Users: Your AI-Generated Text Could Soon Carry A Hidden Watermark; Here's Why

The new textGrain technology operates by creating a statistical pattern in word selection, rather than using visible symbols or hidden characters. This pattern can then be detected by a specialized AI text detector. OpenAI states the watermark should not significantly affect the performance or quality of its AI models, ensuring the user experience remains consistent.

While the initial rollout targets eligible users within the European Union, OpenAI has indicated that API customers globally have the option to implement text watermarking for select models. The company is also developing a corresponding detector to identify the presence of its watermark, suggesting a potential for broader, opt-in adoption beyond the EU's regulatory scope.

OpenAI acknowledges specific limitations of the watermarking system. Testing showed detection rates of approximately 80% for 200-token passages and 95% for 400-token passages in psychology-related text, at a 1% false-positive rate. However, shorter content, mathematical text, translations, and heavily edited text can be more difficult to detect. Critically, the system cannot determine the proportion of human-written content within a final piece, nor does the absence of a watermark automatically confirm human authorship.