Russian Officials Call for Embedding Fear and Conscience Mechanisms into Generative AI

At the BIS Summit conference focused on information security for business and IT, Russian officials proposed adding human-like safeguards to generative artificial intelligence. Deputy Minister of Digital Development Alexander Shoytov stated that current GenAI systems lack any fear of making mistakes.

"In AI, a system of fear must be embedded so that it is afraid to make an error, just like a human," Shoytov said during the panel. The concept does not involve a trembling neural network worried about mortgages or conversations with a boss. Instead, it refers to a built-in mechanism of doubt that allows the model to assess the consequences of its actions, admit when data is insufficient, and stop if the potential cost of an error is too high.

Generative AI today can confidently deliver incorrect answers, invent facts, and carry out risky commands without any internal sense of danger. For ordinary chatbots this results in hallucinations, but for AI agents connected to corporate systems the outcome can be deleted files, data leaks, or operational failures in critical infrastructure.

Another ambitious proposal came from Nikolay Lishin, Deputy Head of FMBA Russia. He suggested embedding not only fear but also a form of conscience into AI models. While fear would prevent errors, conscience would help the system judge whether an action itself is permissible.

The resulting technical portrait of an ideal AI system is one that is intelligent enough to solve tasks, sufficiently cautious to double-check everything, and conscientious enough to refuse when necessary. The remaining challenge, officials noted, is translating these requirements into concrete technical specifications.