Deep Learning Revives Associationism Theory of Human Learning

A new theoretical paper by Daniel Rothschild, published on arXiv (2606.20600), argues that the success of modern AI systems—from large language models to game-playing agents—offers surprising support for a classic psychological theory: associationism. Rothschild proposes a "new associationism" grounded in supervised learning, where learning is driven by evaluative feedback. He contends that this mechanism is surprisingly uniform across a wide range of AI systems, differing primarily in how the feedback signal is generated. This defuses the once-influential criticism that associationist mechanisms are too limited to account for human cognitive capacities. By taking AI seriously as a model of human learning, Rothschild suggests that a modest but genuine form of associationism is not only plausible but empirically supported by deep learning's successes.

However, Rothschild is careful to note that deep learning's architectures—such as transformers and deep neural networks—go far beyond anything classical associationists like Hume or Mill envisioned. Supervised learning operates as one component within these systems, not a complete account. The paper bridges AI research and cognitive science, offering a framework for understanding how error-driven, gradual learning could underpin complex cognition. For tech professionals, this reinforces the idea that the algorithms powering today's AI may also reflect fundamental principles of how humans learn, with implications for AI alignment, education, and our understanding of intelligence itself.