Don’t be fooled—LLMs don’t reason

A recent article from MIT Technology Review argues that current Large Language Models (LLMs) lack genuine reasoning capabilities, despite their impressive ability to mimic human-like responses. The author contrasts modern AI with AlphaGo, the system that famously defeated Lee Sedol in Go. While AlphaGo utilized a dual-system approach—combining intuitive pattern recognition with a deliberative search mechanism—LLMs rely almost exclusively on next-token prediction. This process, described as a form of 'System 1' thinking, excels at pattern completion but fails to engage in the step-by-step, verifiable deliberation required for true reasoning. The author contends that even techniques like 'chain of thought' are merely extensions of this probabilistic prediction rather than a distinct reasoning engine. Consequently, the article warns that without developing systems capable of explicit, inspectable, and persistent deliberation, AI will remain limited in its ability to provide trustworthy insights in complex fields like science and medicine.