The Memory Paradox: Why Our Brains Need Knowledge in an Age of AI (arxiv.org)

This paper argues that widespread reliance on generative AI and digital tools creates a "memory paradox": as external systems become more capable, human declarative and procedural memory — the neural machinery for facts, skills and intuition — risks atrophying because key learning processes (retrieval practice, error correction, schema-building and consolidation) are being short‑circuited. Drawing on neuroscience and cognitive psychology, the authors review empirical evidence that premature offloading to AI during learning inhibits proceduralization and intuitive mastery, and they highlight parallels between deep learning phenomena (notably "grokking" and overfitting/overlearning) and how humans form robust, implicit knowledge via overlearning and repeated error-driven refinement. For the AI/ML community the paper has practical and design implications: effective human-AI interaction requires users to possess strong internal models (biological schemata or neural manifolds) so they can evaluate, refine and direct model outputs. That suggests rethinking interfaces, curricula and evaluation: scaffolded AI use that preserves retrieval practice, delayed or constrained automation during skill acquisition, and tools that promote errorful practice rather than effortless answers. The work raises priorities for ML research on human-in-the-loop workflows, trust and alignment, and for policymakers crafting education and workforce training in the age of large language models.