Sberbank and Andrey Kurpatov's team develop architecture to counter AI hallucinations

Sberbank revealed work on research titled "AI-model of human mental reality". The project involves the Laboratory of Neuroscience and Human Behavior led by Andrey Kurpatov, and the goal sounds ambitious: to give AI-agents the ability to discuss human psyche without hallucinations.

About the research

According to the published description, this is not about a new chatbot feature, but an attempt to describe human mental reality in a form understandable to a machine. Sber talks about a model with which AI-systems will be able to work as with an internal map of concepts, connections and states. This is especially important for tasks where it is not enough to simply select a statistically plausible answer: you need to maintain context, logic and causal relationships around human behavior, motivation and perception.

The very fact that the project is undertaken by a laboratory at the intersection of neuroscience, behavioral research and AI shows the direction of work. Here they want not just to further train the neural network on a corpus of psychology texts, but to rely on a more formal structure. In other words, the task is not for the model to speak beautifully about humans, but for it to invent less and not replace real connections with convenient but false interpretations.

For topics related to the psyche, this is critical: an error in formulation easily becomes an error in the conclusion.

How they want to reduce errors

The key detail is graph architecture. Usually, such an approach means a system where knowledge is represented not as continuous text, but as nodes and connections between them. For AI-agents, this can become a way to verify answers not only by the probability of the next word, but also by the explicit structure of concepts. If the model reasons about fear, motivation, attention or distortions of perception, it can rely on a map of relationships between these entities, rather than random associations from training data.

• fix concepts and their connections explicitly
• check whether a new conclusion contradicts the already known structure
• coordinate the answers of several AI-agents with each other
• reduce the risk of fantasies where causal logic is needed, not rhetoric

Another important point follows from Sber's formulation: the architecture is designed specifically for interaction between agents with each other. This is no longer a single chatbot, but an environment where several models or modules exchange judgments about a person. In such a mode, the problem of hallucinations becomes even more acute: one error can quickly multiply throughout the entire chain. A graph scheme is needed as a common framework that keeps the discussion within consistent logic.

Where this will be useful

If the approach works, it can be applied in decision support systems, digital assistants, educational products and services that analyze user behavior. This is not necessarily about making diagnoses. Much closer is the applied scenario where AI helps to analyze communication, reactions, motivational patterns or cognitive errors, but does so more carefully and consistently.

For the corporate market, this is especially interesting: businesses need agents that don't just retell popular psychology, but know how to reason within a given model and explain where the conclusion comes from. At the same time, publicly, this is still being discussed as research. Sberbank has not disclosed how the quality of such a system will be evaluated, what datasets or expert frameworks are used and in which products the result will appear.

This is an important caveat because fighting hallucinations is one of the most difficult tasks in the entire AI stack. Any architecture can improve the coherence of the answer, but this does not guarantee truth. Therefore, the main question is not only how to represent knowledge about the psyche, but also how to validate such knowledge, update it and not transfer the errors of the researchers themselves into the model.

What this means

Sberbank shows an interesting turn: instead of another universal chatbot, the company is exploring a narrower but complex area where structure, consistency and error control are important for AI. If the graph approach works, it could give the market a new class of agents that reason about humans noticeably more carefully than ordinary LLMs.