ChatGPT: The launch of OpenAI’s ChatGPT gave rise to a whole flurry of memes about how students can skirt their homework and assignments by having the chatbot do it. As it turns out, those jokes may have been more prescient than we thought. In a new study, researchers found that ChatGPT can write such convincing fake research paper abstracts that scientists cannot differentiate them.
ChatGPT is a large language model based on neural networks. It learns by processing large amounts of already-existing human-generated text and can generate realistic and intelligent-sounding text based on user prompts. California-based OpenAI released it on November 30 and it is currently free to use.
According to a Nature article on the study, the researchers asked ChatGPT to write 50 medical research abstracts based on a selection of articles published in the journals JAMA, The New England Journal of Medicine, The BMJ, The Lancet, and Nature Medicine. After this, they ran it through a plagiarism detector and an AI-output detector before asking a group of medical researchers to spot the abstracts generated by the chatbot.
The ChatGPT-generated abstracts passed the plagiarism checker with flying colours, and there was no plagiarism detector. The AI-output detector, on the other hand, was able to spot 66 per cent of the fabricated abstracts. However, human researchers didn’t fare much better–they were only able to correctly identify about 68 per cent of the generated abstracts and 86 per cent of the real abstracts. The medical researchers incorrectly labelled 32 per cent of the generated abstracts as real and 14 per cent of the genuine abstracts as being ChatGPT-generated.
“ChatGPT writes believable scientific abstracts, though with completely generated data. These are original without any plagiarism detected but are often identifiable using an AI output detector and sceptical human reviewers. The boundaries of ethical and acceptable use of large language models to help scientific writing remain to be determined,” wrote the researchers in the abstract of the research paper on bioRxiv.
According to them, journals and medical conferences should adopt new policies that include AI output detectors in the editorial process and clear disclosure of the use of these technologies to ensure scientific integrity.
But according to Aravind Narayan, a computer scientist at Princeton University, it is unlikely that any “serious scientists” will use ChatGPT to generate abstracts. Narayan told Nature “the question is whether the tool can generate an abstract that is accurate and compelling. It can’t, and so the upside of using ChatGPT is minuscule, and the downside is significant.”