Detecting AI-generated text is increasingly important to prevent misuse in education, journalism, and social media, where synthetic fluency can obscure misinformation. Existing detectors often rely on likelihood heuristics or black-box classifiers, which struggle with high-quality outputs and lack interpretability. We propose DivEye, a novel detection framework that leverages surprisal-based features to capture fluctuations in lexical and structural unpredictability, a signal more prominent in human-authored text. DivEye outperforms existing zero-shot detectors by up to 33.2%, matches fine-tuned baselines, and boosts existing detectors by up to 18.7% when used as an auxiliary signal. DivEye is also robust to paraphrasing and adversarial attacks, generalizes across domains, and offers interpretable insights into rhythmic unpredictability as a key indicator of AI-generated text.
Examples
This demo illustrates how DivEye detects AI-generated text across diverse examples. Select different text samples to see how our robust detection system performs against various AI-generated content and human-written text.
We introduce Moving Symbols, a customizable synthetic dataset that facilitates the objective evaluation of unsupervised representation learning in video prediction models. Unlike real-world videos, Moving Symbols permits the isolation of specific representational complexities, including the number and nature of moving objects, their spatial connections, and the temporal dynamics of object movement. Through Moving Symbols, we can thoroughly assess the representational prowess of various video prediction models and meticulously quantify their strengths and shortcomings. Our studies demonstrate that Moving Symbols enables the uncovering of model-specific prejudices and limitations, and proposes avenues for further exploration in unsupervised video representation learning research.
High confidence that this text was generated by an AI model. The linguistic patterns and structure are consistent with machine-generated content.
Warning: This is for representational purposes only. For real testing, visit our HF Space page.
Citation
@inproceedings{
diveye25,
title={Diversity Boosts {AI}-Generated Text Detection},
author={Advik Raj Basani, Pin-Yu Chen},
booktitle={Data in Generative Models - The Bad, the Ugly, and the Greats},
year={2025},
url={https://openreview.net/forum?id=QuDDXJ47nq}
}