АВТОМАТИЧЕСКОЕ РАСПОЗНАВАНИЕ ТЕКСТОВ, СОЗДАННЫХ ИСКУССТВЕННЫМ ИНТЕЛЛЕКТОМ. СРАВНИТЕЛЬНЫЙ АНАЛИЗ МОДЕЛЕЙ

9 7

Аннотация

. This paper investigates the effectiveness of machine learning methods in automatically distinguishing artificial intelligence (AI)-generated texts from human-written texts. The study was conducted on a balanced dataset (2,750 essays; 1,375 entries per class). 14 linguistic-statistical features were extracted from the text, among which vocabulary_richness, word_count, text_length, sentence_count, and complex_word_ratio were found to have high discriminative value using Cohenʼs d. The features were vectorized using TF-IDF and embeddings, and algorithms such as RandomForest, GradientBoosting, XGBoost, LightGBM, LogisticRegression, SVM, KNN, DecisionTree, AdaBoost, and MLP were evaluated using stratified cross-validation. The results showed that gradient boosting models (especially XGBoost) and transform methods performed well; the classification score on the test set reached very high values. Cluster analysis showed a correlation between thematic structure and class division. However, the generalizability of the obtained high scores requires further testing in the case of cross-domain evaluation, adversarial attacks, and manipulations such as reduction/paraphrasing. Future research is recommended to focus on transformer fine-tuning, adversarial stability, and multilingualism.