OpenAI has developed a highly accurate tool capable of detecting content generated by its AI models, like ChatGPT. Despite its effectiveness against localized tampering, the company has decided not to release it due to concerns about potential misuse and the risk of stigmatizing non-English users. Specifically, translating English text to another language could allow users to bypass detection, disproportionately affecting non-native English speakers.
In a blog post updated following a Wall Street Journal report, OpenAI explained that while the detection tool is proficient, it remains vulnerable to certain exploitations. The company is weighing the ethical implications of its release, particularly how it might impact various user groups. OpenAI is cautious about unintentionally discouraging non-native English speakers from using AI for writing due to these detection methods.
OpenAI’s tool utilizes invisible watermarking and proprietary methods to identify AI-generated content. While there are existing third-party tools for detecting AI-generated text, none have achieved high accuracy in peer-reviewed research. OpenAI’s internally developed system holds significant potential but remains under wraps as the company navigates the complexities of ensuring fairness and preventing misuse across diverse user demographics.