The smartest AI "understands nothing": what that means

Modern language models show remarkable abilities: they write code, compose poems, and even pass difficult legal exams. Yet behind these achievements lies a fundamental paradox: they do not understand a single word of what they "say." Their operation rests on nothing more than statistical prediction of the next word in a sequence, based on enormous volumes of data.

Unlike artificial intelligence, human cognition relies on an understanding of how the world works, on a grasp of how objects move and interact in space. We build mental models of reality that let us not merely predict but also explain what happens. Because of their architecture, current language models lack this ability.

This fundamental limitation, long drowned out by a stream of enthusiastic coverage, is becoming harder to ignore. It shows up as an inability to be truly creative, a susceptibility to hallucinations and biases, and brittleness when facing unfamiliar situations. A model trained on texts describing cats may be completely thrown by a photo of a cat lying in an unusual pose.

Growing awareness of these limits heralds a paradigm shift in AI development. The era of unrestrained hype, fed by promises of imminent "strong AI," is gradually giving way to a more realistic view of what the technology can and cannot do. Investors and researchers are starting to refocus on building more reliable, explainable and controllable systems that can solve specific tasks in narrow domains.

According to a number of experts, 2026 will be a turning point, a year of sobering up and reassessing priorities in the AI industry. Practical usefulness and reliability will matter more than impressive demos. Building AI capable of real understanding and learning will require new approaches and breakthrough discoveries. The key may be integrating symbolic and neural-network methods, as well as creating systems that can interact with the physical world.

The consequences of this "sobering up" will be felt by every player in the market. Startups betting on the "magic" of large language models will struggle to raise funding. Large companies investing in AI will assess risks and potential returns on their projects more carefully. Users disappointed by unfulfilled promises will view AI's capabilities more critically.

In short, recognizing the fundamental limits of today's AI systems marks the start of a new stage in the technology's development. The era of unrestrained hype is ending, replaced by a more realistic and pragmatic approach. The future of artificial intelligence belongs to systems capable of real understanding, learning and interaction with the world.