Gary Marcus

Gary Marcus is the most prominent public critic of the large language model paradigm. He is not opposed to AI as a field — he founded a machine learning company (Geometric Intelligence, acquired by Uber in 2016) and has contributed to cognitive science research. His critique is specifically technical: he argues that LLMs are sophisticated pattern matchers that can produce outputs that appear intelligent without the underlying capacities that biological intelligence relies on, including causal reasoning, systematic generalisation, and grounded understanding of the physical world.

His 2019 paper "Deep Learning: A Critical Appraisal" and subsequent Substack writing have documented specific failure modes in LLMs — hallucination, inconsistency, susceptibility to adversarial inputs — that he argues reveal fundamental limitations rather than engineering bugs. The debate between Marcus and LLM proponents (particularly LeCun, who largely agrees on the limitations, and scaling advocates, who argue the limitations will resolve with more data and compute) is the central technical argument in AI about whether current approaches can reach general intelligence.

The AICI regards Marcus's work as essential reading not because his predictions have always been correct — he underestimated the pace of capability improvements — but because his critical method (find the failure cases, ask what they reveal about the underlying system) is exactly what rigorous AI assessment requires. The AI governance community needs people who ask what AI cannot do, not only what it can.