TL;DR
The comparative analysis with a traditional NLP-based discriminative neural network model revealed that direct text piece classification outperforms perplexity-based methods, although the latter still demonstrate practical utility.
The aim of the research. The rapid advancement of generative artificial intelligence language models has introduced new complexities in discerning the authorship and quality of textual content. In this paper, we explored the feasibility of using perplexity – a measure of token predictability – as the only discriminative feature for classifying AI-generated versus human-written texts in Ukrainian within the IT domain. Our approach employed small language models to calculate perplexity and detect content generated by state-of-the-art models, evaluating the potential for lightweight solutions. Research results. Initial experiments using a single perplexity threshold across Gemma 3 / Llama 3.2 1B models yielded classification accuracies around 0.70. The full token-level probability sequences were proposed as feature vectors, enabling us to achieve an accuracy of 0.68 via simple KNN classification. Finally, the convolutional neural network architectures trained on these features allowed us to obtain 0.82–0.87 accuracy. Conclusions. The comparative analysis with a traditional NLP-based discriminative neural network model revealed that direct text piece classification outperforms perplexity-based methods, although the latter still demonstrate practical utility.
A hybrid detection framework which combines semantically deep embeddings from the RoBERTa transformer with a set of carefully designed language statistics and linguistic statistics and shows excellent resistance to the surface-level adversarial paraphrasing strategy.
The key idea is to smooth adjacent token scores to reduce their variability, while using an adaptive Lepski-type rule to select the bandwidth according to the local authorship structure, and the proposed method achieves favorable mean square error performance in estimating the underlying signal.
A comprehensive review of the evolution of NLP from traditional rule-based approaches to modern transformer models including BERT and GPT demonstrates that NLP continues to transform intelligent systems and is expected to play an increasingly significant role in the development of next-generation AI technologies.
The fast development of Large Language Models is a problem for keeping academic integrity in scientific publishing. The usual tools that detect this kind of thing use statistics like perplexity and linguistic features. These tools demonstrate limited effectiveness against sophisticated domain-specific AIgenerated text. This paper presents ResearchNet, a hybrid detection framework for identifying whether scientific text is human-authored or LLM-generated. ResearchNet uses a kind of encoder called Frozen SciBERT and a Graph Convolutional Network, which looks at text as a graph where the sentences are connected by logical transitions. It also uses something called DeepScientificAttention to combine information about terminology and citations. The structure of the text to make a strong classification. ResearchNet was evaluated on the AIGTxt dataset across ten scientific fields such as Astrophysics, Medicine and Social Sciences ResearchNet achieves a ROC AUC of 90.02 percent and the highest Mixed-class F1 of 0.69 which is better than all the other models we compared it to and it was 2.7 percent points better, than the next best model, which was SciBERT+GCN.