There is a foundational idea in psychology — drawn from cognitive therapy, constructivism, and the work of thinkers like George Kelly and Alfred Korzybski — that the mental models we carry around in our heads are not reality itself. They are maps. And like any map, they can become outdated, inaccurate, or simply wrong.
The problem is that we rarely notice when our map has stopped matching the territory. We keep navigating with it anyway, interpreting new experiences through its lens, discarding information that doesn't fit, and clinging to patterns that no longer apply. This is not a flaw unique to troubled minds — it is how human cognition works by default. The psychological work is precisely about building the habit of noticing the discrepancy and being willing to revise the map.
"The map is not the territory." — Alfred Korzybski
AI models are frozen maps
A large language model, at its core, is also a map. It is a compressed representation of a vast body of text — a snapshot of what was written, thought, and published up to a certain point in time. The moment training ends, the map is frozen.
The world does not freeze with it. Scientific consensus shifts. Social norms evolve. Events happen. New technologies emerge that redefine the meaning of older concepts. But the model continues to navigate using its static snapshot, confidently routing you through streets that may no longer exist.
This is not merely a technical limitation — it is a structural one, and it has direct parallels to the psychological patterns that cause the most damage in human thinking: the refusal (or inability) to update the model of the world when evidence demands it.
Three patterns worth noticing
1. Hallucination as map-territory mismatch
When a model generates a plausible-sounding but false statement, it is doing exactly what a person does when they fill gaps in their mental model with confident inference. The map has no entry for the territory in question, so it extrapolates — and presents the extrapolation as fact. The problem is not the extrapolation itself; it is the absence of a signal that says this is uncertain ground.
2. Catastrophic forgetting as the cost of updating
In psychology, one reason people resist revising their maps is that updating one belief threatens to destabilise many others. The same is true in machine learning: when you fine-tune a model on new information, it tends to overwrite what it previously knew — a phenomenon called catastrophic forgetting. The field has developed strategies to mitigate this (LoRA, rehearsal methods, elastic weight consolidation), but the fundamental tension remains: updating is costly, and preserving coherence is hard.
3. Retrieval-augmented generation as external memory
One of the most practical responses to the frozen-map problem has been Retrieval-Augmented Generation (RAG): instead of trying to update the model itself, you give it access to an external knowledge base it can consult at inference time. This is, interestingly, close to how healthy human cognition works — we don't rewrite our entire worldview every time we learn something new. We annotate it. We keep notes. We develop habits of checking rather than relying solely on memory.
What this suggests about design
The psychological parallel points to something important for how we think about AI systems beyond the purely technical. A model that cannot express uncertainty, that has no mechanism to flag when it is operating near the edges of its training, and that has no path to graceful update, is not just technically limited — it is epistemically fragile in a way that mirrors the most brittle patterns of human thinking.
The goal in psychology is not to make someone certain about more things. It is to make them better at sitting with uncertainty, distinguishing what they know from what they infer, and remaining open to revision without losing coherence. That seems like a reasonable design aspiration for AI systems too.
We are building tools that will shape how millions of people think and make decisions. The least we can do is make sure those tools model the epistemic virtues we would want in a thoughtful collaborator — curiosity about what they don't know, humility about the limits of their map, and a genuine capacity to update it.