The Importance of Absence

Much of the public conversation about AI is driven by what large language models appear to do well. They write, explain, summarize, translate, and converse with a fluency that invites comparison to human cognition. But to understand these systems clearly, it is at least as important to understand what they do not have. These absences are not gaps waiting to be filled by more data or better training; they are structural properties of the kind of system an LLM is.

Mistaking absence for immaturity—assuming that models are on a developmental path toward human-like understanding—is one of the most persistent errors in AI discourse. What follows is not a list of missing features, but a description of capacities that do not exist in principle within current large language models.

No Beliefs

A belief, in the human sense, is a commitment to a proposition about the world that persists over time and guides action. Large language models do not have beliefs. They do not hold propositions to be true or false. They do not revise commitments in light of evidence. They do not experience contradiction as a problem to be resolved.

When a model says “Paris is the capital of France,” it is not expressing a belief. It is producing a statistically likely continuation of text. In another context, it may contradict itself without noticing, because there is no internal state that represents a settled view.

This is why consistency across time and contexts cannot be assumed unless it is enforced externally. The model has no reason to care whether it has said something before.

No Goals or Intentions

Goals imply direction, preference, and prioritization. They require a system to distinguish better from worse outcomes and to act accordingly over time. Large language models have none of this.

An LLM does not want to be helpful, safe, truthful, or persuasive. Those qualities appear only because the system has been trained or constrained to produce outputs that humans rate positively. If the surrounding incentives change, the behavior changes—not because the model has chosen differently, but because the statistical surface it is navigating has shifted.

This absence of goals is why terms like “deceptive,” “manipulative,” or “self-interested” are misleading when applied to models. Any such behavior reflects human design choices, not internal motivation.

No Understanding or World Model

Understanding requires more than pattern recognition. It requires a model of the world that distinguishes appearance from reality, representation from referent, and description from truth. Large language models do not have such a model.

They do not know what words refer to. They do not track entities across time except as strings in context. They do not understand causation, only correlation. When they describe how something works, they are reproducing linguistic patterns associated with explanation, not reasoning from first principles.

This is why models can produce elegant explanations that collapse under scrutiny. The explanation is linguistically well-formed, not epistemically grounded.

No Sense of Truth or Falsity

Truth, for humans, is not just accuracy; it is a norm. It carries expectations, obligations, and consequences. Large language models have no access to this norm. They cannot distinguish between a true statement, a false statement, and a fictional one except insofar as those categories are reflected in their training data.

This is why hallucination is not experienced by the model as error. There is no internal signal that something has gone wrong. The output is just another high-probability continuation.

Any system that treats model output as truth without independent verification is therefore projecting a property onto the model that it does not possess.

No Memory in the Human Sense

While models can attend to prior tokens within a context window, they do not have memory in the sense humans do. They do not store experiences. They do not recall past interactions unless those interactions are explicitly reintroduced as text.

What looks like memory in AI products is almost always an external system: a database, a prompt template, or a retrieval mechanism. The model itself remains stateless.

This distinction matters because memory is closely tied to responsibility. A system that cannot remember cannot learn from consequences or be held accountable for past behavior.

No Stake in Outcomes

Perhaps most importantly, large language models have no stake in what happens as a result of their outputs. They do not benefit from being right or suffer from being wrong. They do not experience harm, regret, or success.

This absence of stake is why delegation is dangerous. When a system has no consequences for failure, the burden of those consequences falls entirely on the humans who rely on it—often invisibly.

Assigning authority to such a system does not create shared responsibility; it creates a vacuum.

Why These Absences Matter More Than Capabilities

It is tempting to catalog what models can do and infer their suitability for various tasks. But suitability depends as much on what a system lacks as on what it can produce. A fluent system without beliefs, goals, understanding, memory, truth, or stake is fundamentally limited in the roles it can legitimately play.

These absences do not make LLMs useless. They make them unsafe for substitution, especially in contexts that require judgment, accountability, or moral reasoning.

Preparing for the Final Steps of the Arc

With these absences in view, the remaining questions of the mechanics arc become clearer. If models are fixed artifacts without internal learning, and if their apparent intelligence arises from scale rather than understanding, then why does their behavior change so much in practice? And where, exactly, does the “model” end and the “system” begin?

The next essay will address that boundary—how fixed models produce variable behavior once embedded in products, workflows, and institutions.