OpenAI, the company behind ChatGPT, is delaying the release of a new tool described as “highly accurate.” This tool is reportedly capable of identifying content generated by ChatGPT itself. In summary, the hesitation is due to concerns that the tool could be manipulated or discourage non-English speakers from using artificial intelligence (AI) models to generate text.
In a blog post from May, the company revealed that it was developing several techniques to identify content specifically generated by its products. However, an exclusive report by the Wall Street Journal (WSJ), published on August 4, suggested that plans to launch these tools have been postponed following internal debates about the potential consequences of making such technology available.
OpenAI Fears Misuse of its Tool
Following the WSJ report, OpenAI updated a blog post from May, adding additional information about its detection tools. The company revealed that it has not yet set a release date for these tools, though it emphasized that one of them, designed to identify the origin of texts, has proven to be “highly accurate and effective even in cases of minor tampering.”
However, OpenAI acknowledges that there are still techniques that could be used by malicious individuals to evade detection. For this reason, the company has opted not to make this technology publicly available for the time being.
Additionally, the company mentioned concerns about non-English speakers being “stigmatized” when using AI products for text composition. This is due to a loophole that allows for the translation of texts from English to other languages as a method to bypass detection.
OpenAI’s Detection Tool Would Be the First Developed Internally
Currently, there are various products and services that claim to detect AI-generated content. However, none have yet demonstrated high accuracy in general tasks according to peer-reviewed studies.
OpenAI’s tool would be pioneering as the first system developed internally that utilizes invisible marking and proprietary detection methods specifically for content produced by the company’s own models.
