IT之家 News on 10/5: OpenAI announced today that, in order to comply with the content transparency requirements of the EU AI Act, it will add machine-detectable invisible watermarks to eligible ChatGPT and Codex text outputs in the EU region over the coming weeks.
According to the introduction, the text watermarking technology is named textGrain (text grain watermark), which embeds invisible statistical signals during the model's word selection process. A detector identifies these signals to determine whether a piece of text carries an OpenAI watermark.
IT之家 noted that OpenAI's next step is to roll out text watermarking in the EU. Over the coming weeks, the team will introduce the text watermarking feature only to eligible ChatGPT and Codex users across all plans in the EU. It is worth mentioning that OpenAI will not make text watermarking the global default at launch.
OpenAI will enable an opt-in watermarking feature for API customers. Starting today, API customers worldwide will be able to choose watermarked text outputs for some models. This allows customers to decide how watermarking aligns with their transparency obligations and the experience they provide to users. The team is also working with cloud partners to offer watermarking for OpenAI model outputs accessed through their services in the coming weeks.
OpenAI will also provide detector access to researchers and expert organizations, and approved researchers and expert organizations can apply starting today. Access under the code of conduct will initially be granted on a case-by-case basis to support the evaluation and improvement of text provenance. The tool will report whether an OpenAI watermark is detected, but will not reveal user identities or the content of prompts or conversations.