Rethinking the Authorship Verification Experimental Setups

Existing benchmarks may have hidden biases: The widely-used PAN authorship dataset shows inconsistent performance between closed and open test sets, suggesting that current evaluation methods don't fully capture how well models actually work in real-world scenarios.

AI models rely too heavily on author names and places: Using explainable AI techniques, researchers discovered that BERT-based authorship verification models are biased toward recognizing named entities (like person and place names) rather than analyzing actual writing style, which undermines the goal of identifying authors based on how they write.

Removing shortcuts improves both accuracy and fairness: Models trained without access to named entities achieved better performance and generalized better to new datasets, demonstrating that removing these "shortcuts" makes AI systems more robust and less dependent on superficial patterns.

Better experimental design reveals model weaknesses: By creating five new dataset splits specifically designed to isolate topic and writing style biases, the researchers uncovered problems that standard benchmarks had masked, showing that how we test AI systems matters as much as the systems themselves.

Responsible AI requires multiple evaluation approaches: The research demonstrates that responsible development of authorship verification tools requires combining traditional performance metrics with explainability methods and diverse test datasets to catch biases that single benchmarks might miss.