How do humans and AI generalise differently?

Humans and AI generalise differently primarily in their methods and outcomes. Human generalisation often involves abstraction and concept learning, allowing individuals to learn from a few examples, leverage common sense, and apply robust reasoning even in novel contexts. They excel in dealing with noise and out-of-distribution data through causal inferences^[1].

In contrast, AI typically relies on data-driven statistical learning, which struggles to generalise beyond its training distribution. AI systems often derive patterns based on correlations rather than causal relationships, limiting their ability to handle unforeseen contexts effectively^[1]. As such, humans achieve a more flexible and context-aware form of generalisation than current AI models.