A web tool called can identify whether an essay was generated by the artificial intelligence chatbot ChatGPT with high accuracy. This could help identify cheating in schools and misinformation, but only if OpenAI, the company behind the popular chatbot, continues to give access to the underlying AI models.
OpenAI is reportedly working on inserting a watermark to text that its models generate. But in the time since became publicly available in December 2022, millions of people have tried it and there have been reports of students using the tool in schools and universities to do their homework.
at Princeton University has developed a that uses ChatGPT鈥檚 own architecture to indicate the likelihood that a chunk of text came from ChatGPT. 鈥淭he methods here basically use the model itself to evaluate whether [the AI鈥檚 output] is plagiarised,鈥� says Tian.
Advertisement
When someone enters text into GPTZero, it runs it through an older version of the model behind ChatGPT, then rates how likely it is that the text was produced by the AI and how much this likelihood varies over the length of the script. Human-generated text can alternate between seeming to be generated by an AI and not, whereas AI texts are more constant, says Tian.
Tested on a data set of AI-written articles based on BBC News headlines, GPTZero got the answer right about 98 per cent of the time, having a false positive rate of less than 1 per cent.
Sometimes, tools like GPTZero can only work for so long before more advanced models learn how to circumvent their detection techniques, but Tian thinks the system is robust because it uses the AI鈥檚 own systems. 鈥淎s long as OpenAI releases its models publicly, or as long as developers of a language model are transparent with their methodologies and model, detection is always one step ahead,鈥� says Tian.
A spokesperson for OpenAI says: 鈥淲e鈥檝e always called for transparency around the use of AI-generated text. Our policies require that users be upfront with their audience when using our API and creative tools like DALL-E and GPT-3. We don鈥檛 want ChatGPT to be used for misleading purposes in schools or anywhere else, so we鈥檙e already developing mitigations to help anyone identify text generated by that system.鈥�
Tools like these aren鈥檛 new, says at King鈥檚 College London. A similar tool from researchers at聽MIT, IBM and Harvard to detect individual words that seem to imply the text was artificially generated, but ChatGPT鈥檚 success has made them more relevant.
GPTZero relies on having access to an AI model (GPT-2) that is similar to the one behind the chatbot to work accurately, she says. But if the model behind ChatGPT is updated to become more different, it may lead to inaccurate results, she adds.
