OpenAI is delaying the release of a new “highly accurate” tool designed to detect content generated by ChatGPT due to fears that it could be tampered with or lead to non-English users avoiding AI text generation.
In a May blog post, the company mentioned working on various methods to identify content created by its products. However, on August 4, the Wall Street Journal reported that internal debates over the potential consequences of releasing these tools have stalled their launch.
Following the WSJ report, OpenAI updated its May blog post with new details about the detection tools. Despite claiming that at least one tool is “highly accurate and effective against localized tampering,” there is still no release timetable.
The company noted that there are still ways for bad actors to bypass detection, making them hesitant to release the tool to the public.
Additionally, OpenAI implied that non-English speakers might be “stigmatized” against using AI for writing due to a translation-related exploit that could bypass detection.
“Another significant concern we are considering is that our research indicates that the text watermarking technology has the potential to disproportionately affect certain populations. For example, it may stigmatize the usage of AI as a valuable writing tool for non-native English speakers.”
While various products claim to detect AI-generated content, none have shown high accuracy across general tasks in peer-reviewed research. OpenAI’s system would be the first internally developed tool using invisible watermarking and proprietary detection methods for content generated by its models.
