08 Jun AI Content Increasingly Detected in ASCO Oncology Abstracts

MedicalResearch.com Interview with:

Dr. Howard

Frederick Howard MD
Assistant Professor of Medicine
Section of Hematology / Oncology
University of Chicago

MedicalResearch.com: What is the background for this study?

Response: With the advent of AI language models like ChatGPT, these tools may be used to generate scientific literature or abstracts. Indeed, a survey conducted by Nature in 2023 found that nearly 30% of scientists were using AI tools to aid in the writing of scientific manuscripts. The use of AI in scientific literature can be difficult to identify, and previous studies suggest that human reviewers cannot distinguish between AI generated and human written scientific abstracts. Commercial tools designed to identify AI content may have a higher degree of accuracy, but the optimal approach to applying such tools to detect AI content within scientific literature is uncertain.

MedicalResearch.com: What are the main findings?

Response:  We evaluated text from over 15,000 abstracts from the ASCO Annual Meeting with three commercial AI content detectors (GPTZero, Sapling, and Originality.ai), and found there was an increasing signal of AI content for the year 2023 as compared to 2021/2022, with approximately twice as many abstracts characterized as containing AI content in 2023. All AI detectors achieved a 99% or higher accuracy of distinguishing text generated by GPT3.5 models from human written text, but were less accurate in identifying text from the newer GPT4 model or mixtures of human written and AI generated text. Many translation tools rely on similar principles as AI language models, and we also found that translating human written abstracts to Spanish and back to English resulted in an up to 33% false positive rate for AI content by one of the detectors (Orignality.ai). Of note, abstracts with a high likelihood of AI content were more likely to be presented in an online only format, suggesting that scientific reviewers found the abstracts to be of lower quality. 

MedicalResearch.com: What should readers take away from your report?

Response: Although AI content cannot be identified with perfect accuracy, our study suggests a significant rise in the use of AI to aid in the writing of scientific abstracts. AI models are prone to errors – such as referencing a scientific study which does not exist, or confidently stating an incorrect fact. Although most scientists may use these models responsibly and rigorously review AI generated text included in their studies, standards are needed to identify and verify the accuracy of any AI content included in the literature to ensure that readers are not misled by inaccurate statements from AI models. These AI detectors could serve as screening tools to identify abstracts with the highest likelihood of AI content to undergo more careful review for content accuracy.

MedicalResearch.com: What recommendations do you have for future research as a results of this study?

Response: Although AI content cannot be identified with perfect accuracy, our study suggests a significant rise in the use of AI to aid in the writing of scientific abstracts. AI models are prone to errors – such as referencing a scientific study which does not exist, or confidently stating an incorrect fact. Although most scientists may use these models responsibly and rigorously review AI generated text included in their studies, standards are needed to identify and verify the accuracy of any AI content included in the literature to ensure that readers are not misled by inaccurate statements from AI models. These AI detectors could serve as screening tools to identify abstracts with the highest likelihood of AI content to undergo more careful review for content accuracy.

Disclosures: Consulting fee from Novartis.

Last Updated on June 8, 2024 by Marie Benz MD FAAD