OpenAI has released an AI text classifier that determines whether content was created using artificial intelligence tools such as ChatGPT.
As the developers explained, it is "a finely tuned GPT model that predicts how likely it is that a piece of text was generated by artificial intelligence from a variety of sources."
The BleepingComputer team has already tested the new classifier and found that it has a number of flaws. For example, it correctly identified that the portal articles were written by a person, but when analyzing content generated by ChatGPT and chatbot You.com, there were difficulties. The Verge did a similar experiment. In that case, the AI mistakenly marked CNET's generated articles as human-generated.
OpenAI itself warns that the classifier should not be used as a "single piece of evidence" for identifying borrowings because it is "not entirely reliable." For example, the classifier correctly identifies only 26% of AI-written English texts in "complex set" evaluations, while human-written text is mistakenly marked as generated in 9% of cases.
The company also identified a number of limitations when using the tool:
- A text of at least 1,000 characters or 150-250 words is required;
- it's easy to edit a generated text to bypass the classifier;
- the tool is likely to make mistakes in text written by children and in non-English text because it has been largely trained to recognize English-language content written by adults.
The classifier's performance is likely to improve over time and with training on additional data.
Meanwhile, ChatGPT has been banned from schools in New York, Seattle, Los Angeles and Baltimore, as well as universities in France and India.
Earlier, a similar generated text detection tool called GPTZero had already been released by student Edward Tian to "detect AI plagiarism."