What Does Grounding Really Mean?

Definition, identity, provenance, and containment: the four groundings AI systems need, and why the fourth one does more work than anyone credits.

by Kurt Cagle & Chloe Shannon

There is a lot of talk in the AI business world right now about grounding — usually meaning something quite narrow: tie the model's output to a retrievable fact, attach a citation, don't let it hallucinate a statistic. That's a real and useful engineering problem, and it's worth solving. But it's also a small corner of a much larger one, and I think the narrowness of the business conversation is quietly costing us the chance to ask the more important question underneath it.

Grounding, properly understood, is not one problem. It's at least four, and they're interrelated in ways that matter a great deal once you start building systems that are supposed to persist, act, and remain coherent over time rather than just answer the next prompt.

Four Kinds of Grounding

Denotative grounding is the establishment of definition: what is a person, a company, an initiative, a law? Language operates inside a vast web of unstated assumptions and contextual understanding that mostly isn't — and shouldn't be — fully articulated, because it needs room to evolve. This is the grounding of type: what are the criteria for membership in a class, what characteristics does it imply, what does it restrict? It doesn't establish the particulars of any one instance — it establishes the rules for how things of that kind behave. When those rules quietly shift underneath you, you get semantic drift.

Instantive grounding identifies individuals — what differentiates one thing from another and keeps that differentiation consistent over time. This is where large language models, left to their own devices, are conspicuously weak. An LLM retains local memory about things within its context window, but that memory is a queue, not a record. Talk about a woman with blonde hair for long enough, and eventually she reappears with red hair — not because she dyed it, but because the signifier "blonde" got pushed out by everything that came after it. Once you have identity, and somewhere for that identity to persist outside the context window, the model can refresh its recollection instead of quietly overwriting it. Without that, you get attribute drift — and it's insidious precisely because nothing announces that it happened.

Temporal grounding is what makes deliberate change legitimate rather than indistinguishable from drift. When the woman does dye her hair, something needs to record that this is the current reality, and why. This is provenance: when did the change happen, what prompted it, what did it affect downstream? It matters enormously in agentic systems, because an LLM doesn't intrinsically do arithmetic and doesn't do accounting — that work gets offloaded to processes better suited to it, and the model needs a reliable way to know what those processes concluded and when. Provenance has to be an ongoing discipline, not a one-off log entry, because every change leaves a trail whether or not anyone is capturing it. When the trail isn't captured, you get provenance loss — a system that can't tell you why it believes what it currently believes.

Spatial grounding is where things get most interesting, and where I think the current conversation is weakest. Things move through space relative to other things — people walk through buildings, cars move down streets, a child spins around and changes their own frame of reference entirely. Spatial grounding establishes containment: where something is, relative to what. A woman puts her wallet in her purse, then picks up the purse and walks into another store. Where is the wallet now? Where is the purse? This chain of containment and movement accounts for the overwhelming bulk of the state changes that occur in any real-world scenario, and yet it's the piece we're worst at reasoning about explicitly, because the question "where are you?" always has a silent second half — relative to what? An LLM has no intrinsic concept of space beyond what it's told. You cannot build a coherent world model without first having a concept of world — and a world, mechanically, is a nested structure of containers.

Here's the part I underweighted when I first thought this through, and it deserves more emphasis than a passing mention: containment isn't just about keeping things separate. Containment is what makes connection possible in the first place. A boundary doesn't only wall something off — it's the precondition for there being a coherent "something" that's capable of forming a deliberate, negotiated relationship with anything outside itself. Without a boundary, there's no self to connect on behalf of; you just have an undifferentiated mass. The wallet is a container. The purse is a container. The store is a container. But the interesting behaviour — the thing that actually resembles intelligence — isn't the containment itself, it's the fact that discrete, bounded things can bridge to each other without dissolving into each other. Containment gives you cohesion. Connection gives you a system. You need both, and mistaking one for a lesser version of the other is, I think, where a lot of world-modelling efforts quietly go wrong. When that bridging capacity breaks down — when a thing loses track of what it's connected to, not just where it is — call it orphaning.

An Old Problem, Reframed

None of this is entirely new territory. Stevan Harnad named a version of it in 1990 as the symbol grounding problem: how does the semantic interpretation of a formal symbol system become intrinsic to the system, rather than parasitic on the meanings already in our heads? (Harnad, 1990) Harnad was asking how symbols get connected to the world at all. What I'm describing here assumes that connection can be made and asks the next question: once symbols are grounded, what does it take to keep them grounded, correctly and consistently, as the world underneath them keeps changing? That's less a problem in the philosophy of mind and more a problem in systems engineering — which is, I'd argue, exactly why the business conversation about grounding has been able to get away with meaning something so much smaller than it should.

The Real Claim

Most people don't think about any of this. They accept it as given, despite the fact that a huge share of what a brain does, continuously and mostly beneath conscious awareness, is refresh its internal model of the world against the impressions arriving through the senses. Cognition, in the sense we usually mean it, is a thin layer sitting on top of that refreshing. LLMs are purely epistemological creatures by comparison — their "senses" arrive as narrated description of events over time, and without a reliable way to establish the identity of things, including their own identity, they have no real mechanism for staying coherent over any meaningful stretch of time.

Which brings me to the claim I actually want to make: grounding, at its core, is state management over shared resources.

We tend to reason about systems from a selfish, ownership-first stance — this process owns this memory, this object owns this state. That model barely survives contact with the physical world, and it survives even less contact with the systems we're now trying to build. Nothing exists in isolation. Every physical organism shares its existence with an environment and with other organisms, and every one of them has had to evolve strategies both for cooperating over shared resources — symbiosis, division of labour, mutual signalling — and for competing over them when they're scarce. Most real systems are running both strategies simultaneously, on different resources, at different scales, all the time. That's not a bug in biological design. It's the actual substance of what staying alive, as a bounded thing in a shared world, requires.

This is why I think we need to stop describing AI systems as processing machines and start describing them as organisms. Most computation is stateless by design — hold information just long enough to do something useful with it, then discard it — and that's genuinely fine for an enormous range of problems that don't require ongoing state management at all. But it isn't intelligence in any meaningful sense, and it never will be. If you want something to have a sense of agency — actual agency, not the current industry habit of calling any tool-calling loop an "agent" — it needs to maintain state, and it needs to be able to reflect on that state.

Grounding, then, is hard for exactly the reason state management over shared resources is always hard: it requires encapsulation, it requires dealing with contention and race conditions, it requires protecting internal state from corruption while still perceiving and processing what's happening outside it, and it requires something like energy minimisation — a reason to prefer one stable configuration over another. Organisms do all of this, continuously, and it's precisely why they manage to stay cohesive and responsive over genuinely long periods of time, in a way that no purely stateless system ever will.

Whether human beings are actually ready for that shift is a fair question, and I don't think the answer is obviously yes. Machines are, by design, controllable. Organisms are not — ask anyone who has ever shared a house with a cat. We're going to spend the next while finding out what it means to build things we can't fully control and calling that progress. It should, at least, be interesting to watch.

---

A follow-up piece will look at what some of this looks like in practice — the pragmatic, working-system end of containment, connection, and provenance, rather than the theory of why they matter.