OpenAI, the leading generative artificial intelligence (AI) research company in the world, has confirmed that AI writing detectors are ineffective.
In a recently published FAQ, OpenAI acknowledged that these detectors, which are often used to flag AI-generated content, have failed to accurately discern between human-generated and machine-generated text.
The ineffectiveness of AI writing detectors like Originality.ai, GPTZero, and ZeroGPT, have been extensively discussed and criticized by experts, who have referred to them as “snake oil.” All AI detectors often produce false positives because of their reliance on “unproven detection metrics.” It has been observed that there is no inherent quality in AI-written content that consistently differentiates it with human-written text, and detectors can be easily evaded through rephrasing. Another study revealed that AI detecting softwares are biased against non-native English writers.
The discontinuation of OpenAI’s own AI Classifier further proves the limitations of AI writing detectors. Designed to detect AI-written content, the experimental tool demonstrated a terrible accuracy rate of just 26 percent.
OpenAI’s latest FAQ also debunks a common misconception that ChatGPT, the AI language model developed by OpenAI, is capable of discerning AI-generated text. The company clarified that ChatGPT has “no knowledge” of whether the content is AI-generated or not. As a result, the model may randomly provide responses when asked about the AI nature of text.
In addition, ChatGPT has the tendency to provide convincing yet false and misleading information. Incorrect data and figures are the most common among them. OpenAI warns against relying solely on ChatGPT for research purposes, as it can fabricate quotes, citations, and other information.
Does this mean that there is undetectable AI? While automated AI detectors have proven ineffective, it is still possible for humans to detect AI writing. Educators who are familiar with a student’s writing style can notice sudden changes in style and capability, indicating the use of AI-generated content.
AI writing, as it currently stands, cannot be easily detected. AI detectors have a high tendency to produce false positive results, so they should not be completely relied upon.