OpenAI is preparing to add invisible watermarks to texts generated by ChatGPT and Codex in the European Union.
The watermark is not visible when you read or copy the text. Instead, OpenAI says it is new textGrain The technology slightly alters the model’s word selection to create a statistical pattern that can later be recognized.
“In the coming weeks, we will add an invisible watermark to permitted ChatGPT and Codex text output in the European Union,” said OpenAI explained.
OpenAI does not yet make this a global standard.
Starting today, API developers around the world can enable watermarks for supported models, but they will remain disabled by default.
The company is also opening applications for its watermark detector, but access is initially limited to approved researchers and expert organizations.
OpenAI admits that its AI watermark may disappear if you edit the text
Text watermarking is far from perfect, and OpenAI’s own testing shows that fairly normal editing can significantly impact the ability to recognize AI-generated text.
“In an evaluation of 400 token passages, replacing 10% of the words with synonyms reduced recognition from about 92% to 66%. Replacing 25% of the words reduced it to 17%, OpenAI noted.
With a false positive target of 1%, OpenAI detected the watermark in about 80% of psychological responses with 200 tokens, compared to about 95% when the text reached 400 tokens.
For subjects like math, recognition was even worse because the model has less freedom to choose different words.
“The lack of a detected watermark does not prove human authorship,” OpenAI warned. “Text generated with OpenAI tools may be too short, edited or translated for recognition to work reliably.”
OpenAI also says that a detected watermark does not reveal who created the text, their account, prompt or conversation, and it cannot tell how much of the final work was written or edited by a human.
Interestingly, OpenAI states that watermarks do not have a significant impact on the quality of GPT-6 Astra and benchmark results remain largely similar when textGrain is enabled.
