HULLMI: Human vs LLM identification with explainability

Simpler models work as well as complex ones: Traditional machine learning approaches (Naive Bayes, Random Forests, XGBoost) performed comparably to advanced NLP detectors like T5-Sentinel and RoBERTa-Sentinel at distinguishing human-written text from AI-generated text, suggesting that expensive, complex models may not be necessary for this task.

Explainability is often overlooked in AI detection: Most existing AI detection tools prioritize accuracy without explaining why they classify text as human or AI-generated, creating a "black box" problem that limits trustworthiness and adoption in sensitive domains.

LIME can make detection tools transparent: By using the explainable AI technique LIME, the researchers identified which specific parts of text most influenced detection decisions, providing a pathway to make AI detection tools more interpretable and trustworthy for real-world deployment.

Real-world testing matters: The study used diverse datasets including both curated corpora and real-world samples, suggesting that detection tools need robust testing across varied conditions to be reliable in practical applications like education, healthcare, and media.

Production-ready detection requires both accuracy and interpretability: For AI detection tools to be trusted and adopted in high-stakes domains, they need to combine accurate predictions with clear explanations of how those predictions were made, not just achieve high detection rates.