Russian scientists develop system to reduce errors in artificial intelligence responses

Russian scientists are working on a new system designed to limit the generation of false or misleading information produced by artificial intelligence models, particularly large neural networks that may produce inaccurate responses in certain contexts.

The development comes amid growing global debate over the reliability of AI systems, as studies and reports have shown that neural networks can sometimes generate incorrect or misleading information, raising concerns in fields such as journalism, research, and public communication.

Experts explain that such errors occur because neural networks do not independently verify the accuracy of online sources but instead rely on large datasets that may include outdated or unreliable information.

In response to these challenges, researchers at Reshetnev University in Krasnoyarsk, eastern Siberia, have developed a methodological framework aimed at reducing the presence of fabricated or unverified content in AI-generated responses.

The approach is based on Retrieval-Augmented Generation (RAG) systems, which rely on curated knowledge bases built from high-quality and verified sources. These databases serve as reference points for AI models when generating responses, thereby improving accuracy and reducing hallucinations.

According to the researchers, while this method significantly reduces the likelihood of errors, inaccuracies can still occur due to incomplete datasets, input errors, or ambiguous queries.

A research team led by Associate Professor Anastasia Polyakova from the Department of Intelligent Systems and Automation analyzed cases of AI inaccuracies and developed a classifier to detect unreliable outputs. They also designed automated testing protocols that generate sample queries, compare responses against reference standards, and evaluate accuracy using semantic similarity metrics.

Based on initial findings, the team has developed a prototype real-time monitoring module capable of logging queries and conversation context, assessing response reliability, and assigning confidence scores to AI outputs. If a low confidence level or potential error is detected, the system generates an alert for human supervision.

Researchers note that the proposed framework is highly adaptable and can be integrated into various AI-driven applications, including chatbots and government systems, as well as sectors such as healthcare, legal services, and religious institutions.