OpenAI introduces ‘TextGrain’, an AI-generated text identification technology… Announcement of pilot introduction of advertising

“There is virtually no difference in the quality of the answers”… A pilot introduction of advertising when creating images was also announced.

(San Francisco = Yonhap News) Correspondent Kwon Young-jeon = Following Antropic, OpenAI also applies a text identification pattern (watermark) to articles generated by artificial intelligence (AI).

Open AI announced on the 5th (local time) that it will introduce 'Text Grain', a technology to identify whether a text was generated by AI, in response to the mandatory generative AI transparency regulations of the European Union (EU) AI law.

Textgrain is a technology that inserts subtle statistical signals in the process of selecting words when an AI model generates text.

OpenAI explained that even if this technology is applied, there is no real difference in the quality, performance, or readability of AI answers.

However, it was also revealed that if the sentences are short or the content has little flexibility in word selection, such as mathematical formulas, there are limitations, such as a low identification rate of whether AI was created or not.

It was added that if a user replaces some words in the generated text with synonyms, the identification rate may drop significantly, which may be affected by post-editing.

For example, if one-tenth of the words in a text of 400 tokens are replaced with synonyms, the identification rate drops from 92% to 66%, and if one-quarter of the words are replaced, the identification rate plummets to 17%.

This technology will be applied to ChatGPT and Codex output in the EU region in the coming weeks.

By default, the watermark function is disabled in application program interface (API) results for developers, but the watermark function can be turned on voluntarily depending on the needs of each organization.

The tool to detect the identification pattern will be provided first only to researchers and professional organizations that have received access permission requests and been approved.

Previously, Antropic also applied a similar AI content identification function last August.

As backlash grew, with users protesting that they would cancel Claude's paid subscriptions, Antropic explained that it would also apply to other corporate AI models as an obligation under EU regulations.

A similar function called ‘Synth ID-Text’ is also applied to Google Gemini.

Meanwhile, OpenAI also announced a new advertising policy on this day.

When using the image creation function within ChatGPT, visual advertising will be pilot tested with some advertisers in the United States.

OpenAI emphasized that advertisements do not affect ChatGPT answers, and the content of the conversation remains private.