August 05, 2024 – According to The Wall Street Journal, OpenAI has developed a tool capable of high-precision identification of texts generated by ChatGPT, yet the company has been holding off on its release. In response to this, OpenAI has acknowledged that they are indeed researching text watermarking technology, but emphasized that the technology still faces numerous challenges.
It is understood that through text watermarking, OpenAI aims to exclusively detect texts originating from ChatGPT, rather than those from models of other companies. This will be achieved by making subtle alterations to the word choices made by ChatGPT, essentially creating an invisible watermark within the text that can later be detected by a separate tool.
OpenAI stated that text watermarking is one of the multiple solutions they are exploring, along with classifiers and metadata, to determine the source of texts. Although text watermarking technology performs well in certain scenarios, its effectiveness diminishes when faced with tampering behaviors such as translation, rewriting, or the insertion of special characters. Additionally, this technology may disproportionately affect certain groups, such as non-native English speakers.
Due to these complexities and their potential impact on the entire ecosystem, OpenAI has expressed caution in advancing the research of text provenance technology. Instead, the company has prioritized the development of authentication tools for audio-visual content.
