Comparing scientific abstracts generated by ChatGPT to original abstracts using an artificial intelligence output detector, plagiarism detector, and blinded human reviewers

Abstract

Background: Large language models such as ChatGPT can produce increasingly realistic text, with unknown information on the accuracy and integrity of using these models in scientific writing. Methods: We gathered ten research abstracts from five high impact factor medical journals (n=50) and asked ChatGPT to generate research abstracts based on their titles and journals. We evaluated the abstracts using an artificial intelligence (AI) output detector, plagiarism detector, and had blinded human reviewers try to distinguish whether abstracts were original or generated. Results: All ChatGPT-generated abstracts were written clearly but only 8% correctly followed the specific journal's formatting requirements. Most generated abstracts were detected using the AI output detector, with scores (higher meaning more likely to be generated) of median [interquartile range] of 99.98% [12.73, 99.98] compared with very low probability of AI-generated output in the original abstracts of 0.02% [0.02, 0.09]. The AUROC of the AI output detector was 0.94. Generated abstracts scored very high on originality using the plagiarism detector (100% [100, 100] originality). Generated abstracts had a similar patient cohort size as original abstracts, though the exact numbers were fabricated. When given a mixture of original and general abstracts, blinded human reviewers correctly identified 68% of generated abstracts as being generated by ChatGPT, but incorrectly identified 14% of original abstracts as being generated. Reviewers indicated that it was surprisingly difficult to differentiate between the two, but that the generated abstracts were vaguer and had a formulaic feel to the writing. Conclusion: 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 skeptical human reviewers. Abstract evaluation for journals and medical conferences must adapt policy and practice to maintain rigorous scientific standards; we suggest inclusion of AI output detectors in the editorial process and clear disclosure if these technologies are used. The boundaries of ethical and acceptable use of large language models to help scientific writing remain to be determined.

Can artificial intelligence help for scientific writing?

“…To detect the author is a complex task and requires thorough critical reading to reach a conclusion. However, a few characteristics might reveal that a paper was written by a chatbot, such as the lack of nuance, style, or originality, which could allow the identification by AI output detectors and skeptical human reviewers [ 6 ]. Interestingly, the same writing traits could be found in texts written in a language that is not an individual's native tongue.…”

Artificial Intelligence Can Persuade Humans on Political Issues

“…This demonstration of AI's persuasive capability thus presents regulators with new urgency in addressing AI-based misinformation. Potential responses include requiring AI chatbots to reveal themselves to be AI, embedding identifiers for AI-generated text content, training AI models to detect AI-generated content 12 , and installing guardrails on AI models to refuse tasks such as generating arguments in favor of unfounded claims. It is important for future research to explore the feasibility and efficacy of these and other potential solutions.…”

ChatGPT: Bullshit spewer or the end of traditional assessments in higher education?

“…Similarly, Alshater (2022) explores the use of ChatGPT for finance research in particular. Gao et al (2022) compare scientific abstracts generated by ChatGPT to original abstracts using an artificial intelligence output detector, plagiarism detector, and blinded human reviewers.…”