OpenAI Confirms: AI Writing Detector is Not Accurate

2023-09-11

No detector can "reliably distinguish between AI-generated content and human-generated content".

Last week, OpenAI released some tips for educators in a promotional blog post, demonstrating how teachers can use ChatGPT as an educational tool and providing suggestions on how to get started. In a related FAQ, they officially acknowledged what we already know: AI writing detectors, although frequently used to check students' assignments, are not accurate.

OpenAI wrote in the FAQ "Can AI detectors work?": "In short, no. While some companies (including OpenAI) have released tools claiming to detect AI-generated content, these tools have not been proven to reliably distinguish between AI-generated content and human-generated content."

Due to reliance on unverified detection metrics, detectors like GPTZero often produce false positives. This ultimately leads to no distinction between AI-written text and human-written text, and users can deceive the detectors by rephrasing. In the same month, OpenAI discontinued its AI classifier, an experimental tool that detects AI-written text. Its accuracy was a dismal 26%.

OpenAI's new FAQ also addresses another major misconception, that ChatGPT itself can know whether the text is AI-written. OpenAI wrote, "Additionally, ChatGPT doesn't know what content is AI-generated. Its response to questions like 'Did you write this article?' or 'Is this written by AI?' is random and lacks factual basis."

Following this line of thought, OpenAI also addresses the tendency of its AI model to fabricate false information. The company wrote, "Sometimes, ChatGPT's answers can sound persuasive, but it may provide incorrect or misleading information (often referred to as 'hallucinations' in the literature)." "It can even make up citations, so don't rely on it as your sole source of research."

In May of this year, a lawyer got into trouble for doing just that - he extracted 6 non-existent cases from ChatGPT.

Although automated AI detectors are ineffective, it does not mean that humans cannot discern AI-generated content. For example, teachers familiar with students' typical writing styles can sense sudden changes in their style or ability. Additionally, some careless individuals may leave traces when passing off AI-generated work as human-written work, such as the phrase "as an AI language model," indicating that they accidentally copied and pasted ChatGPT's output.

For now, completely avoiding the use of automated AI detection tools is the safest choice. "AI writing is undetectable at the moment and likely to remain so," said Ethan Mollick, a professor at the Wharton School. "AI detectors have a high error rate, so they should not be used."

RECENT AI NEWS

RECENT AI TOOLS