OpenAI is set to introduce invisible watermarks for text generated by ChatGPT and Codex in the European Union in the coming weeks, making AI-generated content easier for detection tools to identify.
The move is aimed at meeting the EU AI Act requirement that machine-generated content be identifiable in a machine-readable form.
OpenAI’s textGrain Watermarking System
OpenAI’s system, called textGrain, creates an invisible statistical pattern through the model’s word choices.
The company said its performance matched or exceeded other watermarking methods, including Google DeepMind’s SynthID for text, and plans to open-source the underlying technology.
API customers worldwide can opt in to watermarking on select models, while researchers and expert organisations may also get access to the detector.
OpenAI said the detector identified about 80% of watermarked 200-token passages and around 95% of 400-token passages at a 1% false-positive rate, though performance was weaker in areas such as mathematics.
Editing also reduces detection: replacing 10% of words with synonyms cut detection from about 92% to 66%, while replacing 25% reduced it to 17%. Translated, shorter or highly constrained text can also be harder to detect.
The watermark does not reveal the user, prompt, ownership or level of human input, and its presence is not proof of authorship. The absence of a watermark also does not prove human authorship.
OpenAI will also test a new visual advertising format later this month in the US with an initial group of advertisers.
Ads will appear alongside images generated by ChatGPT, remain clearly labelled and separate from the generated content.
OpenAI said advertising will not influence ChatGPT’s answers, while advertisers will be able to promote products, their uses and related experiences.
