“Deep Learning: A Critical Appraisal” is a 2018 paper by Gary Marcus, written at the height of enthusiasm following deep learning’s successes in speech recognition, image recognition, and game playing. Marcus does not dispute those wins. His claim is that the method has structural limits that more of the same will not fix, and that the field was underestimating how far it still had to go to reach general intelligence.
The paper enumerates ten concerns. Among them: deep learning is data-hungry and learns shallowly; it transfers poorly to situations outside its training distribution; it has no natural way to represent hierarchical structure or to do open-ended reasoning; it cannot easily incorporate prior knowledge or common sense; it is hard to interpret and to debug; and it can be brittle and fooled in ways humans are not. Marcus frames many of these as consequences of relying on correlation in large datasets rather than on explicit models of the world.
His conclusion is not to abandon deep learning but to supplement it. He argues that reaching artificial general intelligence will require combining neural networks with the symbolic, structured-knowledge techniques that earlier AI emphasized - the position now called neuro-symbolic or hybrid AI. The paper became a reference point for the long-running argument between researchers who expect scale and data to keep delivering and those who believe new architectural ingredients are needed.
Why business readers should care: many deployed AI failures - confident wrong answers, breakdowns on edge cases, sensitivity to small input changes - are exactly the limits this paper named years before the current wave. They are worth designing around rather than assuming they will be trained away.