An open-source tool designed to detect AI-generated Vietnamese text reported a ROC AUC of 0.960 during threshold calibration. At the operating point identified as optimal, its false-positive rate, the rate of incorrectly flagging human-written text, was 0.049, while its true-positive rate was 0.815. Another threshold, chosen to target a true-positive rate while keeping the false-positive rate at or below 5%, was 0.899.
Those figures come from the preprint's threshold calibration. The thresholds were updated in July 2026 using new AI-generated training datasets from OpenAI and Google language models, alongside human-written datasets from research cited by the paper. The calibration metrics do not make authorship certain for any individual document. The paper says final authenticity decisions should be made by human evaluators.
The reported benchmark was narrow. It used three out-of-domain news datasets generated by GPT-5.6 Luna, Gemini 3.6 Flash and Claude Sonnet 4.6, with N = 20 reported for each dataset. The paper does not report dataset composition or class balance for the comparison, and it provides no uncertainty estimates or statistical tests. It also says the tool has not been extensively evaluated in diverse real-world settings.
How the detector handles long documents
VietAIDetector is described as a zero-shot system. PhoGPT-4B acts as an observer and PhoGPT-4B-Chat as a performer. The detector turns their treatment of the text into a score by dividing log-perplexity by cross-perplexity, then applies a threshold selected by Youden's J, Closest Point or the true-positive-rate-at-0.05-false-positive-rate rule. The threshold then determines the detector's classification.
The design also accommodates long inputs. The interface supports raw text, common text files, scanned documents and text that exceeds the context size of the language models used. Instead of truncating such files, the system divides them into overlapping token chunks, with the window and overlap sizes configurable. A reporting layer combines the chunk-level classifications into an overall decision and produces a downloadable PDF with color-coded highlights for each chunk.
The comparison was mixed
GPTZero was used as the baseline because the authors considered its functionality similar to VietAIDetector. Across the newly generated datasets, the paper reports comparable performance overall. The Gemini comparison is mixed: VietAIDetector's average AI score was 0.81, compared with 0.7 for GPTZero, but GPTZero recorded the higher accuracy.
At one tested chunking setting, the reported results varied by dataset. With a 600-token window and 150-token overlap, the AI percentage and accuracy were 96% and 100% for Claude Sonnet 4.6, 81% and 90% for the Gemini dataset, and 100% and 100% for GPT-5.6 Luna. The broader grid search covered window sizes from 200 to 650 tokens and overlap from 50 to 150 tokens, using a Youden threshold of 0.927966.
The authors draw a firm line around the result
The authors describe the approach as state of the art on out-of-domain Vietnamese datasets, attributing that assessment to previous research. In the evidence presented here, that is an author-reported interpretation rather than an independently quantified result from the benchmark itself.
The paper says the software should not be treated as a legally authoritative decision-making tool and that final authenticity judgments belong to human evaluators. The comparison also does not establish that VietAIDetector is superior in every setting, since GPTZero had higher accuracy on the Gemini dataset.
Open code, early evidence
The software is released under the MIT License and is freely available on GitHub, with pilot deployment planned at Nha Trang University. The manuscript is a preprint submitted to Elsevier on 26 August 2026. Its acknowledgment credits Nha Trang University with providing resources and the research environment and thanks the open-source community for enabling the software.
For readers who need to screen Vietnamese text, the paper offers an implementation with multiple input formats, long-document handling and document-level reports. Its evidence supports a configurable research tool with reported calibration and benchmark results, but the authors' own warning keeps the result in perspective: an AI score is evidence for review, not a final finding about who wrote a document.
Paper data and sources
Original title: VietAIDetector: An Open-Source Zero-Shot Detector for Vietnamese AI-Generated Text
Authors: Trieu Hai Nguyen, Van-Dung Hoang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text