Mirror Minds?

Pezzulo, G., Parr, T., Cisek, P., Clark, A., & Friston, K. (2024). Generating meaning: active inference and the scope and limits of passive AI. Trends in Cognitive Sciences, 28(2), 97-112.

This paper argues that while generative AI systems share similarities with active inference models of the brain in their use of generative models and prediction, they fundamentally differ in how they acquire and use these models. Biological systems learn through embodied interactions with the world, grounding their understanding in sensorimotor experiences and causal consequences of actions. This active engagement allows them to develop genuine understanding and meaning. In contrast, generative AI learns passively from large datasets of curated text and images, lacking the direct connection to action and the world that living organisms possess. Consequently, generative AI systems may excel at generating content that mimics understanding but lack the true, grounded understanding developed through active inference and embodied experience, leading to concerns about their limitations and potential "inverse phylogeny."