A model-independent redundancy measure for human versus ChatGPT authorship discrimination using a Bayesian...

Area of Science:

• Computational Linguistics
• Forensic Linguistics
• Stylometry

Background:

• Growing concern over undeclared use of artificial intelligence (AI), such as ChatGPT, in academic and scientific writing.
• Difficulty in verifying authorship of student submissions and scientific articles.
• Need for reliable methods to detect AI-generated content.

Purpose of the Study:

• To develop an AI model-independent measure for discriminating between human-authored and AI-generated texts.
• To quantify syntactic differences between human and AI writing styles.
• To implement a probabilistic approach for authorship attribution in multilingual contexts.

Main Methods:

• Development of an AI model-independent redundancy measure.
• Quantification of syntactic differences, effective even for short texts (approx. 1800 characters).
• Implementation of a Bayesian probabilistic approach using the Bayes factor for classification.

Main Results:

• Successful discrimination between human and AI-generated texts (including ChatGPT) across English and French.
• Feasible authorship discrimination with limited misclassification rates, even on short texts.
• Satisfactory model performance demonstrated even with small sample sizes.

Conclusions:

• The proposed Bayesian probabilistic method offers an original and effective approach to stylometry.
• The redundancy measure and Bayes factor provide a robust criterion for authorship verification.
• The method is applicable to multilingual texts across diverse scientific and humanistic fields.