Sberbank and Andrey Kurpatov's team build an architecture to curb AI hallucinations

Sberbank has disclosed work on a research project called the "AI model of human mental reality." The Laboratory of Neuroscience and Human Behavior, led by Andrey Kurpatov, is taking part, and the goal is ambitious: to give AI agents a way to discuss the human psyche without hallucinations.

What the research is about

According to the published description, this is not a new chatbot feature but an attempt to describe human mental reality in a form a machine can understand. Sber talks about a model that AI systems could work with as an internal map of concepts, links and states.

This matters most for tasks where simply picking a statistically plausible answer is not enough: the system has to hold onto context, logic and cause-and-effect relationships around a person's behavior, motivation and perception.

The fact that the project sits in a lab at the intersection of neuroscience, behavioral research and AI shows the direction of the work. The aim is not just to fine-tune a 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 eloquently about people, 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 wording easily turns into an error in the conclusion.

How they want to reduce errors

The key detail is the graph architecture. This usually means a system where knowledge is represented not as continuous text but as nodes and the links between them. For AI agents, this could become a way to check answers not only against the probability of the next word, but also against an explicit structure of concepts. When a model reasons about fear, motivation, attention or perceptual distortions, it can lean on a map of relationships between these entities rather than on random associations from training data.

• record concepts and their links explicitly
• check whether a new conclusion contradicts the already known structure
• align the answers of several AI agents with each other
• reduce the risk of fabrication where causal logic, not rhetoric, is needed

Sber's wording points to another important element: the architecture is designed specifically for agents interacting with one another. That is no longer a single chatbot but an environment where several models or modules exchange judgments about a person. In this mode the hallucination problem becomes even sharper: one error can quickly multiply along the whole chain. The graph scheme is meant as a shared framework that keeps the discussion within agreed logic.

Where it could be useful

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

This is especially interesting for the corporate market: businesses need agents that do not just retell popular psychology, but can reason within a defined model and explain where a conclusion comes from.

For now, though, publicly this is only research. Sberbank has not disclosed how the quality of such a system will be evaluated, which datasets or expert frameworks are used, or in which products the result will appear. This is an important caveat, because fighting hallucinations is one of the hardest problems across the whole AI stack. Any architecture can improve the coherence of an answer, but that does not yet guarantee truth. So the main question is not only how to represent knowledge about the psyche, but also how to validate that knowledge, update it and avoid transferring the researchers' own errors into the model.

What it means

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