Improved detection methods for generated texts

Artificial text detection is the task of classifying natural language text into two classes: generated by a language model and written by a human. Such detectors should be resistant to changes in the semantic domain, generating language model, and language. Currently, some of the most effective detectors consist of one or more pre-trained LLMs, fine-tuned for the text classification task. The models with the best accuracy additionally use statistical and contextual features of the text. In this paper, we propose a new detector model that achieves the highest accuracy among known solutions while using much smaller computational resources for training and inference. This detector uses the internal representations of the pre-trained language model without further training as features for classical classification algorithms. In the course of the work, new results were also obtained showing high efficiency for this task of using internal hidden states of pre-trained language models and reducing the dimensionality of these states.