AEyeDE uses attention maps to detect AI text better than baselines

As language models achieve near-human fluency, traditional detectors relying on surface statistics or likelihood signals are increasingly easy to evade. A new paper by Aria Nourbakhsh, Adelaide Danilov, Christoph Schommer, and Salima Lamsiyah introduces AEyeDE, an attribution-driven approach that leverages model attention as a discriminative signal. The method uses a proxy Transformer model with white-box access to extract attention-based attribution matrices for both human- and AI-generated text. These matrices are fed into a lightweight convolutional neural network (CNN) that learns to differentiate the two classes. The key insight is that attention maps exhibit recurring local structures whose relative frequencies differ consistently between human and AI writing, providing a complementary and interpretable detection signal.

Across multiple encoder-decoder translation settings, AEyeDE consistently outperforms a text-only baseline. In decoder-only settings (e.g., GPT-style models), it performs strongly in generator-specific detection and remains competitive on standard benchmarks. Impressively, the approach shows robustness under cross-dataset transfer—meaning it generalizes to new domains—and withstands alternative-spelling perturbations that often fool other detectors. The authors plan to release the code publicly, which could accelerate further research into interpretable AI-text forensics. For professionals working on content moderation, plagiarism detection, or disinformation analysis, AEyeDE offers a principled, attention-based method that does not rely on easily manipulated surface features.