Written by Martin Kardzhilov
NeuroAI is used to describe everything from brain-inspired chip design to neural decoding. Until the field agrees on what it is, it cannot be bought, regulated, insured or evaluated. Definitions are not semantics; they are market infrastructure.
I have watched three different things be called NeuroAI in the same week. A review of brain-inspired approaches to AI energy efficiency. A preprint on networks of astrocytes. A funding round for an implantable interface. All three are legitimate work. They are not the same industry.
This matters more than it sounds. Categories are not marketing conveniences; they are the infrastructure that lets buyers, regulators, insurers and investors make decisions. Until a category is defined, it cannot be procured. Procurement requires a specification. Specifications require a boundary. Without a boundary, every conversation about NeuroAI becomes a conversation about which NeuroAI.
So let me be specific about the distinction that matters commercially.
There is a body of work that uses AI to understand the brain: neural foundation models that generalise across species, brain regions and tasks, suggesting a set of machine-learnable rules govern neural population activity. That is neuroscience with better tools, and it is valuable.
There is a body of work that uses the brain to improve AI: the argument that the brain’s roughly 20 watts of power, with about 0.1 watts attributed to cortical computation, is evidence that competent cognition persists under tight constraints. A review published this month maps seven interacting energy regimes in large-scale AI and concludes the answer is cross-layer co-design rather than any single trick. That is architecture research, and it is valuable.
And there is a third thing, which is where Brainlayer operates: models of human state during real-world operation — attention, readiness to intervene, correction behaviour, recovery — used to make autonomous systems work better alongside people. This is not neuroscience and it is not chip design. It is the measurement layer for physical AI, and it is the one with an immediate commercial buyer.
Conflating these three costs everyone. It lets a hardware company market a neuroscience adjacency it does not have. It lets a research lab’s genuine result be read as a product claim. It makes due diligence nearly impossible, because there is no shared vocabulary in which to ask what a company actually does.
The governance conversation has the same problem. The OECD’s recent report on neurodata governance, built on interviews with 16 neurotech companies across seven countries, found that the hardest issue is definitional: nobody agrees what counts as neural data, so nobody can say which protections apply. Reuse rules for research and AI training remain unsettled. That is a definitional gap producing a regulatory one.
My proposal is unglamorous. The field should adopt three named categories and use them consistently: neural decoding (reading signals), brain-inspired computing (borrowing principles), and human-state modelling (measuring the person in the loop). Different buyers, different regulators, different evidence standards, different risks.
I am aware that arguing for taxonomy is a strange way for a chief executive to spend a column. But I have sat in too many meetings where two parties used the same term for different products and left without a deal. Clarity is a commercial asset. NeuroAI will get a budget when it gets a definition.
