Wordvice’s Large-Scale Study of False Positives in AI Detection Accepted to EMNLP 2026 Industry Track

Analysis of more than 135,000 pairs of original and edited academic manuscripts finds that professional human editing alone can change AI detection results

Seoul, Korea (South), September 17, 2026 -- “AI detector scores alone cannot establish whether AI was used. Manuscript evaluation should consider the quality of the work and how it was written and edited.”

Wordvice, a global provider of proofreading and editing services and AI writing solutions, has announced a study examining the reliability of AI text detectors and their false-positive rates using more than 135,000 pairs of real-world academic manuscripts. The study was accepted to the EMNLP 2026 Industry Track, part of an international conference on natural language processing.

The study, “Style as a Confound: False Positives in AI Detection of Non-Native Academic Writing,” analyzed 135,389 pairs of academic manuscripts, collected between 2018 and 2025, before and after professional editing. By comparing original and edited versions with the same authorship and research content, the researchers examined how changes in grammar, fluency, and academic style affect AI detection results.

The researchers evaluated 13 AI text detectors using various approaches, including methods based on token statistics, zero-shot detection, and classifiers. They found that false-positive rates for the same set of human-written academic manuscripts varied widely across detection methods, ranging from 0% to 100%.

In particular, AI detection results changed significantly after these manuscripts underwent professional human editing. For several major classifier-based detectors, including MAGE and RADAR, professional editing reduced the false-positive rate at which human-written text was classified as AI-generated. However, other detection methods exhibited the opposite pattern. In addition, more extensive editing was associated with greater changes in AI detection scores.

According to the researchers, these findings suggest that AI detectors can be influenced not only by who or what produced a text but also by various characteristics, such as writing style and linguistic quality. Because legitimate language improvements, such as professional English editing routinely used in academic publishing, can change detection results, the findings call for caution in treating AI detector scores as definitive evidence of AI use.

“An AI detector score reflects how a particular detection method evaluates a text. It cannot, on its own, determine how that text was actually written,” a Wordvice spokesperson stated. “Rather than revising their writing to lower AI detector scores, researchers should focus on improving the clarity, accuracy, academic quality, and publication readiness of their manuscripts.”

The study was conducted during Hyeonchu Park’s internship at Wordvice. Park is a researcher at Chung-Ang University’s Graduate School of Artificial Intelligence, supervised by Professor Bugeun Kim. The research provides empirical evidence to inform the interpretation of AI detector results, particularly in academic writing by nonnative English-speaking researchers. Previous studies have compared different groups of writers, whereas this study directly compared the same academic manuscripts.

Wordvice is a global education technology company built on academic English editing services provided by native English-speaking editors with expertise across numerous research disciplines. In addition to professional editing, the company offers a range of AI writing solutions for researchers, including a grammar checker, translation, paraphrasing, summarization, and AI detection. Drawing on its expertise in professional human editing and AI technology, Wordvice helps researchers improve the quality of their academic writing and more efficiently prepare manuscripts for publication, while also studying the reliability and fairness of AI writing and detection technology.

The preprint is available on arXiv.

Paper title: “Style as a Confound: False Positives in AI Detection of Non-Native Academic Writing”

Accepted to EMNLP 2026 Industry Track

Contact Info:
Name: Gahye Jeong
Email: Send Email
Organization: Wordvice
Website: https://wordvice.ai

Release ID: 89203597

In case of detection of errors, concerns, or irregularities in the content provided in this press release, or if there is a need for a press release takedown, we strongly encourage you to reach out promptly by contacting [email protected] (it is important to note that this email is the authorized channel for such matters, sending multiple emails to multiple addresses does not necessarily help expedite your request). Our efficient team will be at your disposal for immediate assistance within 8 hours – resolving identified issues diligently or guiding you through the removal process. We take great pride in delivering reliable and precise information to our valued readers.