The study tested the effectiveness of AI-detection services, including Doomet, GPTZero, and Antiplagiat (version 1.8.2), in identifying AI-generated content. Researchers evaluated texts of 11,000–12,000 characters using methods such as AI-assisted editing and translation (RU-EN, EN-RU, and double translation).
Results indicate that AI-editing and translation tools fail to mask machine-generated text as human-authored. For AI-generated content, the reduction in AI-detection scores was less than 3% on Doomet. Conversely, applying these tools to human-written texts can lead to false-positive results, with Doomet showing a 51.1% AI-share after double translation.
The analysis of three academic papers written by the same student between 2024 and 2026 revealed an increase in detected AI usage from 0% to 56% according to GPTZero, and from 12.5% to 70% according to Doomet. The author concludes that attempting to modify text with AI tools often increases the likelihood of being flagged by detection systems.
The study highlights that detection services are most sensitive to texts in their native language. GPTZero demonstrated high accuracy for English texts, while Doomet showed limited competence in detecting English-language content. The author advises against using AI to refine academic work to avoid false-positive detections.
