Computing machinery and causal intelligence

This paper develops an argument inspired by Judea Pearl for the view that deep learning models are incapable of causal reasoning. The argument embraces Pearl’s conclusion but rejects his approach, which focuses on a mini-Turing test restricted to causal questions. The argument presented draws on the traditional debate between empiricism and nativism, presenting a challenge for the sort of empiricism that has been associated with deep learning. The paper then considers an objection based on the fact that large language models perform fairly well on various causal benchmarks. In response to this point, a diagnosis is offered of how such models’ performance might be explained without attributing to them genuine causal reasoning abilities. The paper concludes by connecting the discussion to the limitations of the original Turing test and familiar critiques of behaviorism.