Russian Researchers Build Real-Time Monitor to Catch AI Hallucinations

A team at Reshetnev University in Krasnoyarsk has developed a method to detect and flag when neural networks generate false or fabricated information—a persistent problem known in the industry as hallucination. The approach could be applied across education, government services, healthcare, and legal tech.

Current solutions like Retrieval-Augmented Generation (RAG) reduce hallucinations by pulling from a knowledge base before answering, but they still stumble on typos, contradictory queries, or incomplete data. The Reshetnev team, led by associate professor Anastasia Polyakova, tackled this by first building a classifier that identifies common hallucination patterns. They then created an automated stress-testing pipeline that generates test queries, compares outputs against reference answers, and measures accuracy using semantic similarity metrics.

The key deliverable is a prototype real-time monitoring module. It logs every incoming query and dialog context, assigns a confidence score to each model response, and alerts an operator if the risk of an unreliable answer is high. Because the module is model-agnostic, it can be dropped into any neural network—from university chatbots to government portals and medical or legal assistants—without requiring a full system overhaul.