AI Generated Text Detection
Adilkhan Alikhanov, Aidar Amangeldi, Diar Demeubay, Dilnaz Akhmetzhan, Nurbek Moldakhmetov, Omar Polat, Galymzhan Zharas · 2026
Read AI Generated Text Detection by Adilkhan Alikhanov, Aidar Amangeldi, Diar Demeubay, Dilnaz Akhmetzhan, Nurbek Moldakhmetov, Omar Polat, Galymzhan Zharas (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-03812.txt.
Introduction
Adilkhan Alikhanov and six co-authors evaluate methods for detecting AI-generated text, a growing concern as students submit model output as their own work. They built a unified benchmark from two datasets, HC3 and DAIGT v2, and split the data by topic so no topic leaks between training and testing. That design forces detectors to generalize to unseen subjects rather than memorize themes.
A simple word-frequency logistic regression baseline reached reasonable accuracy, but deep learning models did clearly better. A bidirectional LSTM classifier and a DistilBERT model scored similarly on accuracy, while DistilBERT posted the strongest overall discrimination score. The authors conclude that contextual semantic modeling beats surface word features by a wide margin. Schools and platforms choosing detectors should expect transformer models to travel better across topics.
Catalog record
| CREATOR | Adilkhan Alikhanov, Aidar Amangeldi, Diar Demeubay, Dilnaz Akhmetzhan, Nurbek Moldakhmetov, Omar Polat, Galymzhan Zharas |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 28 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-03812
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
AI Generated Text Detection Adilkhan Alikhanov, Aidar Amangeldi, Diar Demeubay, Dilnaz Akhmetzhan, Nurbek Moldakhmetov, Omar Polat, Galymzhan Zharas Abstract The rapid development of large language models has led to the increased amount of AI-generated text and students using the LLM generated texts as their work, which violates academic integrity. This paper presents an evaluation of AI-text detection including both standard machine learning models and transformer-based architectures. We utilize 2 datasets, HC3 and DAIGT v2, and then build a unified benchmark and apply a topic-based data split to prevent information leakage. This ensures robust generalization across unseen domains. Our experiments show that tf-idf logistic regression achieves reasonable baseline accuracy of 82.87%. However, deep learning models actually outperform it. BiLSTM classifier achieves the accuracy of 88.86%, and DistilBERT achieved a similar accuracy of 88.11% with the highest ROC–AUC score of 0.96, demonstrating the highest performance. The results show that contextual semantic modeling is far superior to lexical features and that it is important to mitigate topic memorization using evaluation protocols. Our limitations are mainly related to dataset diversity and computational constraints. In the future, we plan to work on expanding dataset diversity and then utilize parameter efficient fine tuning methods such as LoRA. We can also explore smaller or distilled models and utilize more efficient batching and hardware-aware optimization. I Introduction I-A Motivation The emergence of Large language models (LLMs) has led to an unprecedented spread of AI-produced text in academic, professional, and online settings. The distinction between human and machine generated text has increasingly become a challenge, especially with modern detectors often being unstable, having biases, or easily being compromised. Continually improving the detection technologies is essential to maintain the academic integrity and transparency, as well as reducing the abuse of generative AI systems. This creates a strong motivation to create robust, data-driven approaches that can be scaled across different domains and model types. I-B Problem Statement The goal of this work is to design and evaluate a unified system for classifying whether a given text sample is AI-generated or human-written. For this purpose, we considered multiple modelling approaches—including classical machine learning, sequence-based neural networks, and transformer-based architectures and used two datasets (HC3 and DAIGT v2). The study aims to assess model performance, analyze generalization under topic-grouped splits, and identify limitations of existing detection pipelines. II Related Work Due to the rapid development of LLMs, a lot of research has been done in this area. The authors have tried a variety of methods, which we can broadly categorize into the following categories: traditional ML and statistics-based methods and Deep learning based methods. II-A Traditional ML and statistical methods In the beginning of research authors have heavily relied on extracting linguistic features from texts and then feeding them into Machine learning models. The features generally fall into lexical, structural, and complexity categories. One of the extracted lexical features is POS tags such as NOUN, VERB, ADJ, and so on. In total, 18 POS tags were extracted [ 12 , 11 ] . It was found that AI generated texts usually have more number of NOUN, VERB, ADP, AUX tags as compared to human generated texts [ 11 ] . Another lexical features are the voice usage, active vs passive voices, and text style. It was found that AI generated texts tend to have more linking words [ 12 ] . The extracted statistical features include Readability metrics, lexical analysis, predictive metrics. The readability metrics such as Flesch Reading Ease score and Gunning Fog Index evaluate how complex a text is [ 12 ] . The lexical analysis includes extracting the following features: average line length, vocabulary size, and word density [ 12 ] . As for the predictive metrics, Perplexity and Burstiness were computed for texts [ 12 ] . After extracting features Prova [ 7 ] trained and test such models as XGBoost and SVM and found that they can achieve a reasonable performance. For example, SVM and XGBoost had accuracies of 81% and 84% respectively. However, they are not good enough to trust such systems. II-B Deep learning based methods Deep learning models were primarily based on the transformer architecture and they demonstrated much superior performance. Yadagiri et al. (2024) trained RoBERTa on HC3 dataset by fusing linguistic and statistical features with the word embeddings [ 12 ] . They extracted the linguistic and statsitical features from the texts such as Average Line Length, Word Density, Part-of-Speech (POS) tags, Flesch Reading Ease score, Gunning Fog Index, Perplexity and so on [ 12 ] . Then, they concatenated these features with the word embeddings generated by the RoBERTa. This methodology helped them achieve the accuracy of 99%. They injected a small trainable low rank matrices to the model’s attention layers and fine-tune the model on the HC3 dataset training only the weights in the matrix [ 10 ] . This helped them achieve the accuracy of 91% while cutting the training time and other costs by a lot. Roand LoRA Optimization: To reduce training time and computational overhead, one study employed Low-Rank Adaptation (LoRA) to fine-tune the RoBERTa-base model on the HC3 dataset [ 10 ] . LoRA freezes the pre-trained weights of RoBERTa and injects small, low-rank trainable adapters between specific layers [ 10 ] . This strategy allowed for efficient fine-tuning while maintaining high performance [ 10 ] . Another paper used BERT model and fine-tuned it on their own private dataset for binary classification task [ 9 ] . As a result, they achieved the accuracy of 97%. Another paper that also used BERT managed to achieve the accuracy of 93% on their own private data [ 7 ] . The difference between them is most likely explained via the difference in data quality and number. Another author proposed a hybrid method involving Bi-LSTM and Attention [ 8 ] . Authors first extracted POS tags from texts and then passed through an embedding layer, which learned the grammar and other patterns. Then, they are passed through convolutional layers to extract local patterns and then are passed through Bi-LSTM layer to capture long-range dependencies. On top of that, attention layer was applied which determined which POS tags mattered more. At the end, they managed to hit the accuracy of 88% [ 8 ] . III Datasets This project relies on two public datasets often used in AI text detection studies : HC3 and DAIGT v2. These include matched samples of human writing vs. machine-made text from various areas. This way, we can test how well the tools work on different subjects or tones. III-A HC3 Dataset HC3 (Human-ChatGPT Comparison Corpus) is a high-quality benchmark designed to compare human versus machine responses under fixed conditions using side-by-side analysis. The dataset consists of questions and answers gathered from online websites like Reddit’s ELI5, plus Quora or StackExchange. With every query, one response comes from a person and another is made by ChatGPT. This results in a balanced dataset containing approximately 37k human replies and 37k ChatGPT-generated outputs. The labeling is fixed and unambiguous, ensuring high label consistency. HC3 also includes topic information: the source field allows grouping into domains such as HC3_reddit_eli5 , HC3_medicine , and HC3_finance . From a language