Introduction

Artificial intelligence holds great promise for automating insights from vast datasets, but these data-driven systems can sometimes make bizarre errors. This article explores the faulty reasoning and limitations in current AI that can lead to inaccurate causal conclusions and incorrect facts. We’ll cover issues like correlations versus causation, overfitting on sparse data, goal misalignment, and the need for hybrid reasoning systems.

Artificial intelligence systems like deep neural networks excel at discerning patterns and correlations in huge training datasets. However, this prowess at finding patterns in data does not always translate accurately to the real world. AI systems often make logically unsound inferences, presenting false conclusions and erroneous facts. There have been a number of cases where AI has got causality very wrong, by assuming correlation implies causation without deeper reasoning.

AI Falls Into the Correlation-Causation Trap

A core machine learning technique used by AI systems is detecting correlations – when two variables appear statistically related in data. However, the old adage applies: correlation does not equal causation. Just because variables are correlated does not imply one actually directly causes the other. AI systems lack the basic common sense to intrinsically understand this distinction.

For example, an AI model may analyse data showing that ice cream sales and drowning deaths both increase sharply in the summer. It could potentially draw the conclusion that increased ice cream consumption causes more drowning fatalities with the modelling falling into the classic correlation-causation trap. Hot summer weather drives both factors, but no logical causality exists between them. However, and here is the core issue, AI has no inherent ability to step back and apply this basic rationale.