Latent Trajectory Discrimination for AI-Generated Text Detection

Most AI-generated text detection (AIGTD) methods treat documents as static objects, relying on global embeddings or aggregate statistics. This paper investigates whether the distinctive signature of AI-generated text lies instead in the dynamic, sequential evolution of latent representations—a byproduct of the autoregressive generation process—and whether modeling these trajectories can improve detection accuracy.

The authors propose the Geometric Trajectory and Contrastive Learning (GTCL) framework. Instead of compressing a document into a single vector, GTCL segments text into overlapping windows, encodes each window into a latent space, and treats the sequence of these embeddings as a discrete trajectory. To focus on the generative dynamics rather than static content, the model computes first-order differences between consecutive windows. It then uses supervised contrastive learning to structure the projection space, training the model to distinguish between the geometric patterns of human-written versus AI-generated text. Finally, a k-nearest neighbors classifier is used for detection.

GTCL consistently outperforms existing detection baselines across multiple benchmarks, including the RAID and NYT-AI datasets. The results demonstrate that explicitly modeling the sequential dynamics of latent states provides a more robust discriminative signal than static document representations. This suggests that the autoregressive mechanism leaves a structural, geometric footprint in the latent space that is largely independent of the specific topic or lexical choices, making it a powerful feature for detection.

This work shifts the paradigm of AIGTD from static classification to dynamic trajectory analysis. By focusing on how semantic representations evolve, the approach offers a more nuanced understanding of the differences between human and machine-generated content. This methodology provides a promising path toward more robust detection systems that are less susceptible to the limitations of traditional, static-based approaches.