OpenAI postpones release of new detection tool for AI-generated content

OpenAI appears to be holding back a new tool that can detect content generated by ChatGPT, which it claims is “highly accurate.”

The decision comes amid concerns that the tool could be manipulated and that non-English speaking users would not have access to reliable text generation using AI models.

Working on methods to identify AI content

In May, the company said in a blog post that it was working on several methods to identify content generated specifically by its products. However, according to an exclusive report from the Wall Street Journal (WSJ) on August 4, plans to release these tools have been stalled by internal debate over the potential implications.

Following the WSJ report, OpenAI updated its May blog post with new information about the detection tools. The company noted that there is still no timetable for release, despite saying at least one tool is “highly accurate and even effective against local tampering.”

Detection can still be bypassed

OpenAI stated that there are still ways for malicious actors to bypass detection, which is why the company is unwilling to release the tool publicly. Additionally, the company suggests that non-English speakers could be “stigmatized” when using AI products for writing functions, due to an exploit that translates English text into another language to evade detection.

“Another important risk we weigh is that our research suggests that the text watermarking method has the potential to disproportionately impact some groups. For example, it could stigmatize the use of AI as a useful writing tool for non-English speakers,” OpenAI said.

While there are currently several products and services available that claim to be able to detect AI-generated content, none are known to have demonstrated high levels of accuracy in peer-reviewed research. OpenAI’s detection tool is said to be the first internally developed system to rely on invisible watermarks and proprietary detection methods for content generated specifically by the company’s models.