AI Understands. Differently.

In our article "When Expectations Outrun AI", we explored the fluency trap - how sophisticated language triggers our inference that full human-like intelligence exists behind it. We identified architectural constraints that reveal where that inference breaks down.

Yet one question persists beneath those constraints: do they understand?

This matters. If LLMs merely simulate understanding, then the constraints we identified - the frozen knowledge, the missing memory, the lack of grounding - all point in the same direction. Engineering would be building scaffolding around behavior, not extending understanding itself.

But if some form of understanding is present - even partial or ungrounded - then everything changes. The constraints become boundaries around an existing capacity. We would be extending something real.

How we answer this question shapes how we build these systems - and what we believe intelligence itself requires.

The Persistent Objections

Two objections recur whenever the question of understanding arises.

First: these systems are merely sophisticated auto-complete. They predict the next word based on statistical patterns.

Second: they hallucinate - confidently inventing facts, people, and citations that do not exist. If they truly understood what they were saying, how could they be so wrong?

Both objections sound compelling. Both reveal something important - just not what they first appear to.

The Auto-Complete Argument

The comparison sounds intuitive: both systems predict the next word, so they must work the same way, right?

Not quite. The difference lies in what happens before that prediction.

When you read, “He blarfed the sandwich in one bite”, you immediately infer what blarfed must mean - even though you have never heard the word before. You do this by recognizing relational structure: a subject performs an action on food, completed quickly. Context constrains the plausible meanings.

LLMs can perform a similar kind of inference. From sandwich, one bite, and the sentence structure, they can infer that blarfed likely refers to eating quickly. This is not simple lookup. It reflects sensitivity to contextual and relational patterns learned across many examples.

Traditional auto-complete cannot do this. If a word is not in its lookup table, it fails. It extends text by replaying surface-level associations with limited context, without the capacity to infer new relationships or constrain meaning in novel situations.

Yes, both predict the next word. But treating large language models as “better auto-complete” is like calling translation "better spell-check". What has changed is not prediction itself, but the richness of the internal structure that informs it.

The Hallucination Argument

The second objection points to their mistakes. If they sometimes invent details, the argument goes, then surely they do not understand what they are saying.

The term "hallucination" is misleading. In psychology, a closer analogue is confabulation - the reconstruction of information rather than its retrieval. Humans do this constantly. We rebuild what we recall, filling gaps with plausible details, often confidently even when wrong.

A well-known example comes from cognitive psychologist Ulric Neisser's analysis of Watergate testimony. John Dean, President Nixon's counsel, delivered remarkably detailed accounts of conversations. But when compared to the actual tapes, many specifics - exact phrases, timing, sequencing - did not match. Yet his testimony preserved the relational structure: who held power, who was implicated, what pressures shaped decisions. His memory reconstructed events in a coherent narrative rather than replaying them verbatim.

When LLMs invent details, something similar is happening. They generate plausible continuations based on learned structure rather than retrieving stored facts. Often, the shape of what belongs in a given context - the type of entity, its role, its relationships - is correct even when the particulars are not.

The similarity has limits. Human reconstruction is grounded in lived experience, consequence, and accountability. Model-generated errors are not. But error alone does not cleanly separate relational competence from understanding. Mistakes reveal the process of meaning-making in both biological and artificial systems - not its presence or absence.

What Both Arguments Reveal

Taken together, these objections point to something deeper - how we recognize understanding at all.

We rely on familiar cues: fluent explanation, confident recall, reasoning that resembles our own. In humans, inference from context and reconstruction under uncertainty are not taken as evidence against understanding. In LLMs, the same behaviors are often treated as proof that understanding is absent.

Perhaps the deeper limitation is not in the models alone, but in assuming understanding must resemble our own.