NeuroAI and Beyond

Neuroscience and Artificial Intelligence (AI) have made significant progress in the past few years but have only been loosely inter-connected. Based on a workshop held in August 2025, we identify current and future areas of synergism between these two fields. We focus on the subareas of embodiment, language and communication, robotics, learning in humans and machines and Neuromorphic engineering to take stock of the progress made so far, and possible promising new future avenues. Overall, we advocate for the development of NeuroAI, a type of Neuroscience-informed Artificial Intelligence that, we argue, has the potential for significantly improving the scope and efficiency of AI algorithms while simultaneously changing the way we understand biological neural computations. We include personal statements from several leading researchers on their diverse views of NeuroAI. Two Strength-Weakness-Opportunities-Threat (SWOT) analyses by researchers and trainees are appended that describe the benefits and risks offered by NeuroAI.

💡 Research Summary

“NeuroAI and Beyond” is a comprehensive report stemming from an international workshop held in August 2025 that seeks to chart a roadmap for the emerging field of NeuroAI – the integration of neuroscience insights into artificial intelligence. The authors begin by contrasting the multi‑scale, lifelong learning processes of biological systems with the two‑stage (pre‑training and inference) paradigm of today’s large language models (LLMs). They argue that embodied interaction with the world, a hallmark of biological cognition, is a missing ingredient for achieving more general, adaptable AI.

The paper is organized around five thematic pillars, each framed by a set of open research questions.

1. Embodiment – The authors define embodiment, explain why a body is crucial for NeuroAI, and outline practical pathways (simulation, robotic platforms, virtual environments) for introducing bodily constraints at different stages of research. They discuss whether embodiment can serve as a bridge to artificial general intelligence (AGI) and identify enabling conditions such as interdisciplinary collaboration, data sharing standards, and dedicated infrastructure.

2. Language and Communication – This section examines how to build cognitively plausible language models by borrowing developmental stages, neural language networks, and plasticity mechanisms from the brain. It highlights fundamental limits of current language‑only AI (e.g., lack of grounding, bias, and brittleness) and proposes neuro‑inspired solutions such as continual learning, developmental pre‑training phases, and the incorporation of cognitive bias research into alignment strategies. The authors also explore the potential of NeuroAI for K‑12 education and for mitigating “cognitive atrophy” that may arise from over‑reliance on AI.

3. Robotics – The authors identify three core desiderata for neuro‑inspired robotics: efficiency, safety/reliability, and biological grounding. They propose three design principles—Hierarchical Stack, Distributed Control, and Full‑Stack Co‑Design—that mirror the brain’s hierarchical and distributed processing. The discussion includes the need for certifiable safety guarantees, the role of neuromorphic hardware in reducing energy consumption, and the importance of establishing standards for interfaces, data formats, and evaluation metrics.

4. Learning in Humans and Machines – Here the focus shifts to mechanisms that enable lifelong, robust learning. Topics include interaction‑driven temporal dynamics, modular learning that respects multi‑level constraints, evolutionary‑developmental strategies to avoid retraining massive foundation models from scratch, and methods for instilling common‑sense reasoning. Energy efficiency is emphasized throughout, with the suggestion that spiking neural networks and other low‑power neuromorphic approaches could dramatically cut training costs.

5. Neuromorphic AI Engineering – This pillar asks “How neuro‑like is current AI?” and probes the relevance of binary spikes versus the analog nature of biological signaling. The authors revisit Carver Mead’s original neuromorphic principles, identify under‑exploited organizing concepts, and discuss which biological components (e.g., synaptic plasticity, dendritic computation) are essential for future AI performance. They also map the hardware landscape, pinpointing memory and device technologies that could unlock gains in energy, throughput, and silicon area.

Beyond the thematic sections, the paper provides a forward‑looking outlook with 5‑, 10‑, and 20‑year milestones, ranging from establishing standards and pilot projects to delivering integrated, AGI‑level NeuroAI systems. Two SWOT analyses—one from senior researchers and another from students/post‑docs—summarize perceived strengths (multidisciplinary synergy, novel algorithmic avenues), weaknesses (lack of standards, data scarcity), opportunities (energy‑saving technologies, industry adoption, policy influence), and threats (investment volatility, regulatory uncertainty, widening skill gaps).

A substantial appendix contains personal statements from over forty participants, spanning topics such as artificial emotions, control architectures, enactivist perspectives, and the role of theory in understanding embodiment. These narratives illustrate the diversity of visions and concerns within the community.

In conclusion, the authors advocate that NeuroAI, by jointly considering brain, body, and environment, can overcome the current data‑centric limitations of AI and move toward more efficient, safe, and adaptable systems. However, they acknowledge that substantial work remains: establishing common standards, building robust interdisciplinary infrastructure, and addressing ethical, legal, and safety implications. They call for coordinated action among academia, industry, and policymakers to translate the outlined research questions into concrete experiments, prototypes, and ultimately, transformative technologies.