Critic picks apart Sber neuro lab's AI study, finds methodological flaws

A critical review of a paper from the Laboratory of Neuroscience and Human Behavior has appeared on Habr. The lab is a Sber research unit that studies how artificial intelligence interacts with cognitive processes and human behavior.

What the lab published

The Laboratory of Neuroscience and Human Behavior is a unit within Sber working at the intersection of neuroscience, cognitive psychology and AI. Its authors published a paper with a striking thesis: "AI can change a person's mind. We tested this on moral dilemmas."

The question they examined is not an academic abstraction. The ability of language models to influence people's moral attitudes bears directly on the regulation of AI products, on educational practice, and on how society assesses the risks of large language models. The authors argued that such influence exists and backed the claim with their own experiments.

Sber is one of Russia's key AI developers, with its flagship GigaChat, so the publication claimed serious scientific authority. That is exactly the expectation with which the independent author read it.

What the critic found

The first warning sign was errors in the text. The author initially wrote them off as typos, which happen even in large labs. But a detailed review showed the problem runs deeper than formal slips.

The central objection concerns the interpretation of the experimental data. The critic believes the paper's conclusions do not follow from the data presented with the degree of certainty the authors claim. In other words: the study showed one thing, and the authors drew conclusions as if about another.

• Substantive errors that go beyond typos
• Questionable interpretation of the results of the authors' own experiments
• A mismatch between the data and the final conclusions
• Questions about methodological rigor and reproducibility
• Rhetoric uncharacteristic of peer-reviewed science

All this led the critic to a broader question: what is Sber's neuroscience lab, what is its real scientific level, and what tasks does it actually serve?

Corporate science and its risks

The case fits a broader trend: large tech companies create their own research units that publish work under an authoritative corporate brand. Trust in the brand transfers automatically to the content of the publications, and that is where a systemic risk arises.

"Given Sber's standing as the developer of the best AI in Russia, I took the paper fairly seriously.

All the more surprising that the authors took it less seriously," writes the author of the review.

When corporate labs publish methodologically weak studies on sensitive topics, they create distorted reference points, both for the expert community and for a wider audience inclined to take corporate publications for academic truth.

What this means

Public review by an independent author is an important element of a healthy industry. Corporate AI research increasingly shapes the public narrative about AI's capabilities and risks, and independent criticism is not an attack on reputation but a necessary verification mechanism.

For Sber, the case raises a direct question about standards: to what extent do the scientific publications of corporate labs live up to the trust they expect?