TELL: Explainable AI Text Detection System
Researchers have developed TELL, a novel architecture for AI-generated text detection that prioritizes explainability. Unlike existing systems that output only a numeric score, TELL identifies specific textual 'tells' that indicate whether content is AI or human-written, empowering users like professors to make their own judgments. The system is trained on a custom supervised fine-tuning dataset of domain-specific authorship annotations. This approach addresses the misalignment between current detection outputs and real-world user needs, where a score without explanation is insufficient.
Key facts
- TELL is a new architecture for explainable AI-generated text detection.
- It provides users with specific 'tells' indicating AI or human authorship.
- Current detectors offer only a numeric score without explanation.
- TELL is trained on a custom SFT dataset of domain-specific authorship annotations.
- The system aims to empower users to decide authorship using their own judgment.
- Real-world applicability of current detectors has stalled due to lack of explainability.
- TELL still outputs a numerical score for comparability with other detectors.
- The research is published on arXiv as 2605.27921.
Entities
Institutions
- arXiv