Critic analyzed Sber neuroscience lab study on AI and found methodological errors

On Habr, a critical analysis of an article by the Laboratory of Neuroscience and Human Behavior — a scientific subdivision of Sber studying the interaction of artificial intelligence with human cognitive processes and behavior — was published.

What the laboratory published

The Laboratory of Neuroscience and Human Behavior is a structure within Sber working at the intersection of neuroscience, cognitive psychology, and AI. The authors published an article with a notable thesis: "AI can change human opinion. We tested this on moral dilemmas."

The question they investigated is not an academic abstraction. The ability of language models to influence people's moral positions directly relates to AI product regulation, educational practices, and how society assesses the risks of large language models. The authors claimed that such influence exists, and supported this claim with their own experiments.

Given that Sber is one of the key AI developers in Russia with its flagship GigaChat, the publication claimed serious scientific authority. It was with this expectation that an independent author read it.

What the critic found

The first warning sign was errors in the text. The author initially attributed them to typos — such things happen in large laboratories. But upon detailed analysis, it turned out that the problem ran deeper than formal mistakes. The central criticism is the interpretation of experimental data. The critic believes that the article's conclusions do not follow from the presented data with the degree of certainty with which the authors present them. In other words: the study showed one thing, but the authors drew conclusions as if about something else.

• Substantive errors going beyond typos
• Questionable interpretation of their own experimental results
• Discrepancy between data and final conclusions
• Questions about methodological rigor and reproducibility
• Rhetoric uncharacteristic of peer-reviewed science

All of this prompted the critic to ask a broader question: what is Sber's Laboratory of Neuroscience, what is its actual scientific level, and what tasks does it actually solve?

Corporate science and its risks

The situation fits into a broader trend: large technology companies create their own research divisions publishing works under an authoritative corporate brand. Trust in the brand is automatically transferred to the content of publications — and this is where a systemic risk arises.

"Given Sber's authority as the developer of Russia's best AI, I approached the article quite seriously.

What was surprising was that the authors approached it less seriously," writes the author of the analysis.

When corporate laboratories publish methodologically weak research on sensitive topics, they establish distorted benchmarks — both for the expert community and for a broad audience inclined to accept corporate publications as academic truth.

What this means

Public analysis by an independent author is an important element of industry health. Corporate AI research is increasingly shaping the public narrative about AI's capabilities and risks, and independent criticism is not an attack on reputation, but a necessary mechanism for verification. For Sber, this case raises a direct question about standards: how well do the scientific publications of corporate laboratories meet the trust they expect?