Table 1 Analysis of existing studies (Results in acc %).
References | Model | Dataset | Feature | Results | Strength | Limitation |
|---|---|---|---|---|---|---|
RNN, Markov model | Reddit, Yahoo Answers | N-Gram, TF-IDF | 88 | Combines sequential and probabilistic models for robust pattern detection | Limited contextual understanding; struggles with long-range dependencies | |
SVM, KNN, Decision Tree | Research paper content | POS tagging | 85 | Leverages syntactic cues via POS features for clear interpretability | Classical classifiers may overfit on limited syntactic patterns; poor semantic generalization | |
CNN, RNN | Research paper content | N-Gram, POS | 85 | Utilizes convolutional layers for local pattern extraction and recurrent memory for context | High computational cost; sensitivity to hyperparameter tuning | |
Naïve Bayes, LSTM | WordNet and PAN human-written texts | Textual feature sets | 90 | Combines probabilistic baseline with deep sequence modeling for balanced performance | Naïve Bayes overly simplistic; LSTM requires extensive training data | |
RoBERTa | Yelp user reviews | Default transformer encoding | 91 | State-of-the-art contextual embeddings capture nuanced sentiment and style | Large model size leads to high inference latency and resource demands | |
RoBERTa WordNet ontology | Tweets, Reddit comments, Yahoo answers, and Yelp user reviews. weets, Reddit comments, Yahoo answers, and Yelp user reviews. tweets, Reddit comments | Default feature encoding | 91 | Subword‐level embeddings handle misspellings and rare words effectively | Ontology reliance may introduce bias; limited to covered semantic relations | |
BERT | Essays | TF-IDF | 79 | Fine-tuned transformer demonstrates baseline applicability to structured essay texts | TF-IDF lacks semantic depth; model underperforms on free-form or noisy inputs | |
BERT | Essays | Default feature encoding | 66 | Leverages pretrained contextual knowledge | Low accuracy indicates overfitting to training domain; limited feature adaptation | |
GRU | Essays, Tweets, Yelp | Count vectorization | 87 | Gated units capture sequence dynamics with moderate resource usage | Simpler than LSTM; may miss very long-term dependencies | |
SVM, Logistic Regression, RF, DT | BBC News | Textual features | 89 | Comprehensive comparison of multiple shallow classifiers highlights best performer | Shallow methods struggle with semantic nuances; inconsistent performance across topics | |
SVM, GBM, DT | Online text corpus (AI vs. Human) | Linguistic feature fusion | 87 | Ensemble and single models compared on AI-human task shows versatility | Performance gains marginal; feature engineering intensive | |
Random Forest, SVM | 1500 Human texts | Word embeddings | 92 | Embedding-based features significantly boost classical models | Small dataset limits generalization; embedding quality dependent on pretraining data | |
SVM + LSTM | Mixed Genre Corpus | Word embeddings | 85 | Hybrid approach balances interpretability and sequence modeling | Complexity in combining models; tuning both components is challenging | |
Hybrid Detection Framework | GPT vs. Human news articles | Word embeddings | 88 | Unified pipeline demonstrates end-to-end applicability across news domains | Framework complexity may hinder real-time deployment | |
RoBERTa | Human vs. LLM Text Corpus | Pretrained embeddings | 93 | Achieves state-of-the-art results with minimal feature engineering | Large model footprint; sensitive to domain shift without further fine-tuning | |
BERT | Essays | Pre-Trained Embeddings | 90 | Strong baseline with deep contextual representations | High inference time; less efficient compared to distilled variants |
