Azerbaijani scientists develop AI-based method to detect disinformation

A study by Azerbaijani scientists revealed major limitations in AI-based disinformation detection tools. The findings were recently published in a prestigious international scientific journal.

AZERTAC reports that the study found the 98–99% accuracy demonstrated by AI systems detecting disinformation on benchmark fake texts does not fully reflect their reliability in real-world conditions. It revealed that, in many cases, these systems rely on stylistic features that identify whether a text was written by a human or generative AI, rather than determining whether the news itself  is real or fake. 

According to the study, when generative AI systems such as ChatGPT, Claude, and others rephrase only the writing style of a text without altering its content, the performance of existing fake news detection systems degrades significantly. This suggests that these systems learn mostly stylistic features rather than the content's semantic characteristics.

To address this issue, the researchers designed a two-regime evaluation. As part of the experiment, both real and fake texts were passed through paraphrasing chains of Claude, GPT-4o, and Gemini models to standardize them into a unified writing style, thereby eliminating the impact of stylistic variations on the evaluation.

Based on the results obtained, the authors propose implementing the style control approach as a new methodological standard when evaluating disinformation detection systems. They believe this approach could contribute to the development of more reliable and resilient AI systems in the future.

The study was conducted by Ramiz Aliguliyev, head of department at the Institute of Information Technology under the Ministry of Science and Education; director of the Institute of Data Science and Artificial Intelligence at Azerbaijan Technical University; and corresponding member of ANAS, alongside Jalal Mehdiyev, a doctoral student at the institute.

The findings of the study were summarized in a paper titled “Cross-LLM paraphrase laundering: a register-controlled evaluation of fake news detectors,” published in an international AI journal with an impact factor of 6.5 and indexed in Q1.

It should be noted that the research was supported by a joint grant project (Grant No. AzTU-DQL-2025-M01/60182631) between the Institute of Information Technology and Azerbaijan Technical University.

© All rights reserved. When using these materials, reference to www.iсt.az is required.