Unsupervised Style Representation Learning for AI-Text Detection via Paraphrase Inversion

Researchers have proposed an unsupervised method for detecting AI-generated text without relying on authorship labels, using paraphrase inversion to learn style representations.

The method, described in a paper submitted to arXiv on 8 Jun 2026^[1], involves training a style encoder to reconstruct human-authored text from its machine-generated paraphrase. This approach allows the style encoder to capture non-semantic features needed for reconstruction. The learned representations were evaluated using two detection strategies: a few-shot detector and a zero-shot DeepSVDD-based detector. Results showed that the method matches or outperforms baselines in few-shot detection and is competitive with fully supervised classifiers in the zero-shot regime. Moreover, the representations generalize to unseen tasks, achieving competitive performance on authorship verification and fine-grained style discrimination. A related study on arXiv, submitted on 7 Jul 2022^[2], introduced a method for learning representations that are equivariant to symmetries of data, which decomposes the latent space into an invariant factor and the symmetry group itself. This approach is motivated by theoretical results from group theory and has been shown to produce lossless, interpretable, and disentangled representations.