Machine Learning Techniques for Detecting AI-Generated Content on Social Media

Authors

  • Emmanuel C. Omeye Tansian University Umunya, Anambra State
  • Chuka Charles AGUBOSIM Chukwuemeka Odumegwu Ojukwu University, Anambra, Nigeria
  • Ekene Josiah Olum Raindrop institute of Management and technology

Keywords:

AI-generated Content, Machine-generated Text, Social Media, Natural Language Processing, Machine Learning, RNN-LSTM

Abstract

The growing use of language-generation systems on social media creates a need to distinguish machine-generated messages from human-authored content. This study evaluated Logistic Regression, Random Forest, Support Vector Machine (SVM), K-Nearest Neighbors (KNN) and a recurrent neural network with long short-term memory (RNN-LSTM) using the Twitter Deep Fake Text Dataset. The Kaggle source describes 25,836 tweets, comprising 12,918 human-authored and 12,918 machine-generated tweets from 17 human and 23 automated accounts. User identifiers, account metadata and target-derived variables were excluded from predictive modelling. Accounts were partitioned into training, validation and testing groups of 28, 6 and 6, respectively. RNN-LSTM achieved 92.80% accuracy on 4,428 test tweets, correctly classifying 4,109 observations. Its weighted precision, recall and F1-score were 92.81%, 92.80% and 92.80%, respectively. SVM achieved the highest accuracy among the traditional classifiers at 88%, followed by Random Forest at 85%, Logistic Regression at 81% and KNN at 76%. Lexical diversity and tweet length received the largest reported feature-importance scores. The findings support content-provenance classification within this dataset; they do not establish identity impersonation or performance across other platforms and unseen language generators.

DOI: https://doi.org/10.5281/zenodo.24161043

Published

2026-11-12