The rise of large language models has been extraordinary. ChatGPT, based on OpenAI’s GPT-3.5 architecture, has attracted enormous attention across education, business and entertainment for its ability to produce fluent, extended responses to human queries. But that fluency carries a risk. The authors of the study point out that the same capabilities can be misused for malicious purposes, including spreading disinformation, generating fraudulent academic work and producing deceptive content at scale. As the volume of synthetic text grows, the responsible and ethical use of ChatGPT has become paramount, and tools to distinguish original human-generated text from machine output are increasingly needed.

The new framework is built on an ensemble, a strategy in which several models are combined so that their individual strengths compensate for one another’s weaknesses. At its core are four transformer-based language models, each pre-trained on massive text corpora and each with a distinctive view of language. The first is BERT, Bidirectional Encoder Representations from Transformers, which reads text in both directions and builds deeply contextualized word representations. The second is RoBERTa, a robustly optimized variant of BERT that refines the pre-training recipe for stronger performance. The third is XLNet, a generalized autoregressive model that overcomes some limitations of standard masked language models by considering all permutations of the input sequence. The fourth is GPT-2, a generative pre-trained transformer that predicts the next token from left to right. Because these models differ in how they process context, the researchers reasoned that their combined judgment would be more reliable than any single model’s.

But the architecture does not stop there. The team recognized that while transformers excel at capturing rich semantic features, they are less sensitive to the fine-grained, local patterns that often distinguish machine text, such as characteristic word sequences and rhythm of expression. To address this, the framework passes the transformer outputs through two complementary components. A one-dimensional Convolutional Neural Network, or 1D-CNN, extracts high-level local features by sliding filters across the embedded text and picking up distinctive n-gram-like patterns. A Bidirectional Gated Recurrent Unit, or BiGRU, then considers the temporal flow of information in both directions, capturing how meaning unfolds across a passage from beginning to end and from end back to the beginning. Together, these components allow the model to process contextualized information and reveal meaningful patterns that enhance detection.

The choice to test the system on both short and long texts is one of the study’s most practical contributions. Detecting machine-generated text in short passages, such as social media posts, is notoriously difficult because there is little statistical signal to work with. Longer documents give classifiers far more evidence to examine. The researchers evaluated their model on datasets containing both human-written and ChatGPT-generated samples. Importantly, they controlled for a subtle confound: because ChatGPT’s output can be influenced by the conversation history, the team refreshed the chat thread for every text sample, ensuring that each generated passage was independent and not contaminated by prior exchanges.

The results were striking. On long texts, the ensemble achieved an accuracy rate of 97 percent, correctly identifying ChatGPT-generated text from human text in the vast majority of cases. On short texts, where the signal is weaker, the model still reached 88 percent accuracy, a level that compares favorably with earlier methods. Perhaps more importantly, the experiments demonstrated a clear finding: the transformer ensemble consistently outperformed individual transformer models when used alone. Combining BERT, RoBERTa, XLNet and GPT-2 into a single decision framework proved more robust than relying on any one of them, validating the ensemble approach for this task.

The study builds on a growing body of research into machine-generated text detection. Earlier efforts, such as work on detecting neural fake news, exposed the vulnerability of current classifiers to synthetic text and proposed defensive models. Other researchers have explored detection of short ChatGPT texts using explainable machine learning, and the field has since expanded with dedicated shared tasks on English and multilingual machine-generated content detection. The new work draws on that lineage but adds a distinctive hybrid architecture. The research team had previously developed BERT-CNN-BiLSTM models for detecting fake news on social media, and their experience with that problem clearly informed the design here. Both fake news detection and ChatGPT detection share a common structure: distinguishing content produced with deceptive intent from authentic content, a challenge the researchers note is related to established problems such as hate speech detection.

The implications of this research extend well beyond the laboratory. In education, institutions around the world are grappling with students submitting AI-generated essays, and detection tools could help preserve academic integrity. In journalism and information security, the ability to flag machine-generated disinformation could slow the spread of synthetic propaganda. The authors explicitly frame their results as a useful approach for policymakers and researchers concerned with detecting and preventing the malicious use of ChatGPT. By publishing their methodology, including comparisons against baseline models, the study gives developers a reference point for building practical detection systems.

The technical details also matter for the broader AI community. The study underscores why ensembles have become a recurring theme in modern natural language processing. Prior work, such as ensemble models combining BERT and RoBERTa for classifying idioms versus literal texts, has shown that different pre-trained models encode complementary information. This new study extends that insight to the AI-detection problem, showing that a model that effectively processes contextualized information from multiple transformer perspectives, then enriches it with convolutional and recurrent feature extraction, can reveal patterns in text that individual models miss. The generative-versus-discriminative distinction explored in classic machine learning research also echoes here, as GPT-2 and its generative cousins see language differently from encoder models like BERT and XLNet.

The road to publication was long. The paper was received in March 2024, revised in November 2025, and accepted in December 2025, reflecting the fast-moving nature of this field where the technology being studied evolves faster than the review cycle. The fact that the model achieves such strong performance on a text generated by GPT-3.5-era ChatGPT suggests the underlying patterns of machine fluency are stable enough to be learned reliably, even as chatbots grow more capable.

Of course, no detector is a permanent solution. As generative models improve, they may produce text that evades current detection methods, and the arms race between generation and detection is likely to continue. The researchers acknowledge the need for continued investigation into ways to distinguish human from machine text. But their results demonstrate that the problem is tractable: with the right combination of pre-trained transformers and sequence-aware deep learning, machine-written text leaves a detectable fingerprint. For now, this ensemble framework offers one of the most accurate available tools for finding that fingerprint, whether hidden in a long essay or a short post. The study was not funded by any organization, and the authors declare no conflicts of interest.