MIPT Algorithm Detects AI's Logical Hallucinations by Analyzing Text Structure
A researcher at the Moscow Institute of Physics and Technology (MIPT), Boris Galitskiy, has developed a method for detecting structurally unsupported reasoning chains in large language models, including logical hallucinations.

Using a dataset of approximately 1,000 fictional customer complaints, the algorithm distinguished between sound and hallucinated AI reasoning based on text structure alone in 81% of cases. When the content was also taken into account, accuracy rose to 84%. The method could improve AI reliability in medicine, law, and science, where errors in evaluating evidence can have particularly serious consequences.
According to Galitskiy, AI can generate not only fabricated facts but also reasoning chains containing logical errors. Such errors pose risks when AI systems tackle complex medical, legal, or scientific questions that require selecting relevant evidence and integrating information from different sources.
Galitskiy suggested that these errors have structural features that distinguish them from sound analysis, making it possible to detect them before AI-generated conclusions are used. In the longer term, the approach could improve the performance of large language models in fields where AI errors carry particularly serious risks, the researcher said.








































