Testing for Causality in Artificial Intelligence (AI)

Abstract

In the 1950 in a landmark paper on artificial intelligence (AI), Alan Turing posed a fundamental question “Can machines think?” Towards answering this, he devised a three-party ‘imitation game’ (now famously dubbed as the Turing Test) where a human interrogator is tasked to correctly identify a machine from another human by employing only written questions to make this determination. Turing went on and argued against all the major objections to the proposition that ‘machines can think’. In this chapter, we investigate whether machines can think causally. Having come a long way since Turing, today’s AI systems and algorithms such as deep learning (DL), machine learning (ML), and artificial neural networks (ANN) are very efficient in finding patterns in data by means of heavy computation and sophisticated information processing via probabilistic and statistical inference, not to mention the recent stunning human-like performance of large language models (ChatGPT and others). However, they lack an inherent ability for true causal reasoning and judgement. Heralding our entry into an era of causal revolution from information revolution, Judea Pearl proposed a “Ladder of Causation” to characterize graded levels of intelligence, based on the power of causal reasoning. Despite tremendous success of today’s AI systems, Judea Pearl placed these algorithms (DL/ML/ANN) at the lowest rung of this ladder since they learn only by associations and statistical correlations (like most animals and babies). On the other hand, intelligent humans are capable of interventional learning (second rung) as well as counterfactual and retrospective reasoning (third rung) aided with imagination, creativity, and intuitive reasoning. It is acknowledged that humans have a highly adaptable, rich, and dynamic causal model of reality which is non-trivial to be programmed in machines. What are the specific factors that make causal thinking so difficult for machines to learn? Is it possible to design an imitation game for causal intelligence machines (a causal Turing Test)? This chapter will explore some possible ways to address these challenging and fascinating questions.