OpenAI Implements Invisible Watermarks on ChatGPT Content

OpenAI has begun implementing invisible digital watermarks in text generated by ChatGPT and its Codex development environment. This move is a direct response to the European Union's AI Act, specifically Article 50, which mandates that AI providers enable unambiguous computer identification of synthetic content. The technology, named textGrain, embeds subtle statistical signals based on a secret key that influences the model's word choices, creating a unique mathematical signature. This signature allows specialized detection tools to identify machine-generated text with high probability, similar to Google's SynthID and Anthropic's recent solution.

Extensive performance tests on OpenAI's GPT-6 Astra model indicate that the textGrain implementation does not degrade creative output or accuracy. The model achieved nearly identical scores on complex programming, organizational automation, and professional knowledge tests with and without the watermarks. However, OpenAI acknowledges limitations; detection effectiveness varies with paragraph length and subject matter. For instance, detection rates dropped from 95% for 400-token paragraphs in psychology to 80% for 200-token paragraphs, and significantly lower in fields like mathematics where word choice is restricted.

A major limitation is the impact of human editing. The company's tests showed that replacing just 10% of words with synonyms reduced detection accuracy from 92% to 66%, and replacing a quarter of the words dropped it to 17%. Due to the risk of falsely identifying human text as AI-generated (false positives), OpenAI will not release the watermark detector publicly at this stage, offering controlled access only to researchers and accredited testing organizations.

For developers, OpenAI is enabling opt-in watermarking via its global API, with the default setting remaining off. This allows third-party companies and applications using OpenAI's engines to decide whether to embed watermarks to comply with EU law. This initiative complements OpenAI's existing transparency measures for images and audio. Currently, the watermarking mandate for ChatGPT and Codex will be rolled out in the coming weeks exclusively for users within EU countries. Israeli users are not affected at this stage, but the European adoption signals a growing global trend towards AI regulation and transparency that could eventually influence Israeli laws and educational systems.