TELL: The AI Detector That Explains Itself

Detecting AI-generated text has been a challenge, with many models offering high accuracy. Yet, these models often fall short in real-world practicality. Why? Because users, like professors, are left with a cold, numeric score and little else to work with. Enter TELL, a new model that seeks to change the game by incorporating explainability directly into its architecture.

The TELL Model

TELL doesn't just assign a score to a piece of text. It shows the 'tells,' or indicators, that led to its decision about whether the text is AI or human-written. This empowers users to decide for themselves using their judgment and the context of the writing.

The model is trained with a custom SFT dataset that includes domain-specific authorship annotations. It refines its performance further through GRPO and curriculum learning, achieving a competitive AUROC of 0.927. But what truly sets TELL apart is its focus on native explainability.

Why Explainability Matters

Explainability isn't just a nice-to-have. It's important for user trust and understanding. TELL's approach allows users to see the reasoning behind the detection, which is vital for applications in academia and beyond. Can you really trust a system that doesn't show its work?

The model's explanations are evaluated through human annotations, boasting a mean win-rate of 72.3% in metrics like coherence and plausibility. This builds on prior work from AI ethics, emphasizing the need for transparent decision-making processes.

Implications and Future Directions

By reframing AI text detection from a human-centric perspective, TELL paves the way for a new family of detectors focused on transparency. But is this the right focus? Some might argue that accuracy trumps all else, yet without understanding, how can we trust these systems?

This approach highlights a growing trend in AI: moving beyond mere performance metrics to consider user experience and ethical implications. The paper's key contribution is its shift toward a more explainable future in AI, crucially aligning model outputs with user needs.

As this field evolves, models like TELL might become the norm rather than the exception. The question is, will others follow suit?