Show, Don't TELL: Explainable AI-Generated Text Detection

Explainability matters more than raw accuracy for real-world use: The paper found that existing AI-detection systems achieve high technical performance but fail in practice because they only provide numeric scores without explanation. Users like professors need to understand why a text is flagged as AI-generated, not just receive a confidence number.

The TELL system identifies specific "tells" of AI writing: Rather than a black box score, TELL highlights particular linguistic features and patterns that suggest AI authorship, enabling users to evaluate the detector's reasoning against their own knowledge of the student or author and the writing context.

High-quality explanations are achievable alongside competitive accuracy: TELL maintained state-of-the-art performance (AUROC 0.927) while providing human-validated explanations that scored 72.3% on quality metrics including concreteness, falsifiability, and plausibility—demonstrating that explainability doesn't require sacrificing detection performance.

Human judgment remains essential in the detection process: By design, TELL empowers users to make final decisions themselves rather than blindly trusting the system, acknowledging that context about an author and their writing situation is crucial information that an AI detector alone cannot possess.

Responsible AI requires alignment with user needs: The paper demonstrates that technical excellence is insufficient; AI systems must be designed with the actual end-users' requirements in mind, shifting from "here's a score" to "here's what we found—now you decide."