Researchers want to fix AI's worst habits by borrowing from the brain.
A new paper, grounded in a National Science Foundation workshop held in August 2025, lays out a roadmap for NeuroAI - shorthand for neuroscience-informed artificial intelligence. The authors identify three gaps in today's AI: it can't interact with the physical world, its learning is brittle, and it burns through energy and data at an unsustainable rate. For each gap, they point to a matching principle from neuroscience: co-designing bodies and controllers together, learning through prediction and interaction, multi-scale learning guided by neuromodulatory control, hierarchical distributed architectures, and sparse event-driven computation. They organize all of this into near-term, mid-term, and long-term research goals.
This isn't just a wish list - it's an argument that today's dominant approach, scaling up models and burning more compute, has a ceiling. The paper says getting past that ceiling requires researchers trained across both neuroscience and engineering, plus new infrastructure: shared hardware access, interdisciplinary training, community standards, and ethics guidelines.
It's a sweeping vision, but like most roadmaps born from a single workshop, the real test is whether funding and hardware access follow the paper rather than just citations.