Detection of ChatGPT fake science with the xFakeSci learning algorithm

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Scientific Publishing

Background:

• Generative AI tools like ChatGPT are increasingly prevalent.
• AI-generated content may possess distinct characteristics separable from authentic scientific articles.
• Prompt engineering enables the creation of AI-generated articles on various diseases.

Purpose of the Study:

• To investigate the distinct behaviors of AI-generated content compared to scientific articles.
• To develop and validate a novel algorithm, xFakeSci, for distinguishing AI-generated from scientist-authored publications.
• To assess the efficacy of xFakeSci against traditional data mining algorithms.

Main Methods:

• Articles were generated using prompt engineering for cancer, depression, and Alzheimer's.
• The xFakeSci algorithm was trained on network models derived from both AI-generated and authentic articles.
• A calibration step using data-driven heuristics (proximity, ratios) was incorporated to mitigate overfitting.
• Classification involved 300 articles per condition, balanced with authentic PubMed abstracts.
• Performance was evaluated against Support Vector Machines, Regression, and Naive Bayes algorithms.

Main Results:

• The xFakeSci algorithm achieved high F1 scores ranging from 80% to 94%.
• Classical data mining algorithms scored significantly lower F1 values, between 38% and 52%.
• The enhanced performance of xFakeSci is attributed to its calibration step and proximity distance heuristic.

Conclusions:

• AI-generated content exhibits detectable differences from scientific publications.
• The xFakeSci algorithm represents a significant advancement in identifying and combating AI-generated fake science.
• The developed algorithm demonstrates superior performance in distinguishing synthetic scientific literature.