Axio › Volume 2 › What Beliefs Are

What Beliefs Are

From credence thresholds to models of agents

This chapter is a draft — it is readable but still changing.

A mugger hints that he has something concealed in his jacket. It could be a firearm; it could be a toy. The victim has to evaluate a proposition — the concealed weapon is real — and has to do it now, with a wallet in the balance. At what point does the victim’s suspicion become a belief? Not at certainty; certainty never arrives, in this alley or anywhere else. The natural answer is: at the point where the assessment starts driving behavior. When the victim’s confidence that the weapon is real gets high enough that handing over the wallet becomes the rational move, and the wallet is handed over, something has crossed a line. Whatever we mean by belief, that crossing is it.

That answer is nearly right, and I held it for a while in exactly that form. This chapter starts there, follows it until it creaks, and ends somewhere more radical: beliefs are not things agents have at all. They are features of models of agents — including the model each agent keeps of itself. Getting from the first picture to the second is not a retraction; it is a refinement, and the first picture survives inside the second as a special case. But the destination changes what kind of thing a belief is, so it is worth making the trip carefully.

Credence with Teeth

The word belief wanders vaguely between opinion and knowledge, and most confusion about it comes from letting it wander. The threshold account pins it down: a belief is an agent’s assignment of sufficiently high subjective probability — Credence — to a proposition, robust enough to guide practical decisions, predictions, and actions.

Three features of this definition do real work.

First, belief is quantified by Credence, not by objective frequency. In the Quantum Branching Universe (QBU) — the Everettian picture introduced in Measure and Credence — Measure is objective branch weight, a fact about the world; Credence is how strongly an agent endorses a proposition, a fact about the agent’s epistemic state. Belief lives entirely on the Credence side. It embodies high credence but never absolute certainty, which is as it should be for cognitively limited creatures whose beliefs evolved to guide action under uncertainty, not to mirror the world perfectly.

Second, belief requires actionable consequences. A proposition becomes a belief only when it demonstrably shapes decisions, predictions, or behavior. Ideas that inform no action and constrain no expectation are idle speculations, whatever their owner professes. This is the definition’s cutting edge, and it cuts in the right places: it explains why the professed “beliefs” that never touch conduct — the deathbed conversions of convenience, the opinions worn as team colors — deserve scare quotes. If your credence that the bridge will hold does not show up in whether you walk across it, you do not believe the bridge will hold; you are making noises.

Third, belief is conditional. Since all truth is conditional, every meaningful belief has the implicit form: given conditions X, proposition Y warrants sufficiently high credence. The victim’s belief that the weapon is real is conditioned on background assumptions about muggers, jackets, and the local price of bluffing. There are no unconditional beliefs for the same reason there are no unconditional truths.

So far, so good. Credence, action, conditions: the account demystifies belief, connects it to decision-making, and dissolves the false demand for certainty. As a working definition it earns its keep. But push on it and two joints start to creak.

Where the Threshold Creaks

The first creak is that the threshold moves. Whether a given credence is “sufficiently high” to warrant action depends on the stakes, so the same credence crosses the line for one decision and not another. Suppose the victim’s credence that the weapon is real is 0.6. For handing over a replaceable wallet, 0.6 is far past any sane threshold: comply. For a decision with a different payoff structure — testifying under oath that the mugger was armed — 0.6 may fall well short. Same agent, same proposition, same credence: belief for one purpose, non-belief for another. The threshold account, followed honestly, concludes that “believes P” is not a stable property of an agent at all. It is a property of an agent relative to a decision problem. That is not fatal — it may even be true — but it means belief has already stopped being the simple mental possession we took ourselves to be defining. The definition set out to find belief inside the agent and found something indexed to circumstances outside the agent.

The second creak is deeper. The threshold account still assumes there is a credence in there — a number stored in the victim’s head, waiting to be compared against a threshold. Open the head and look. You find neurons, chemistry, cascading activation: processes, not propositions. Nowhere is there a sentence with a probability stapled to it. The credence, like the belief it was supposed to constitute, is not a component of the agent that anatomy could locate. It is something we ascribe to the agent to make sense of what the agent does. The threshold account quietly presupposes an act of interpretation — someone modeling the victim as a credence-bearing chooser — and then forgets the interpreter was ever in the room.

Take that interpreter seriously and the whole picture reorganizes.

Beliefs Live in Models

Beliefs are not static propositions stored inside minds. They are models of regularity, evolved and learned to explain and predict behavior — first our own, then that of others. To believe a bridge will hold is to model it as structurally reliable, and thus to walk across. To believe someone is honest is to model their future actions as aligned with truth. Belief is not representational content sitting in a mental filing cabinet; it is functional compression — a simplified generative model that guides behavior under uncertainty. In predictive-processing terms, a belief is a prior: a probability distribution over possible world-states, continuously updated by evidence. In Dennett’s terms, it is a feature of the intentional stance: an attribution that renders an agent’s behavior intelligible and predictable. Beliefs belong to the same family as the maps and models examined in Maps, Models, and Understanding — compressions judged by their usefulness — with one twist that changes everything: the thing being modeled is an agent.

Here is the claim in full strength: beliefs exist only within models of agents, never in the agents themselves. A physical or computational agent merely enacts processes — it perceives, reacts, regulates. Nothing in the enacting is a belief, any more than anything in a falling rock is a trajectory equation. But construct a model of that agent as an intentional system — a thing with goals, expectations, and information — and belief becomes a meaningful term. It appears in the representational layer that describes how the agent relates to its environment, and nowhere else.

This applies even to the first person. An agent can maintain a self-model that depicts itself as an agent — and that self-model can contain beliefs. But the beliefs live in the model, not in the agent as a physical entity. And the rule applies recursively: models of agents can contain models of agents, each level ascribing beliefs to the one beneath.

Belief is a model of the world inside a model of an agent. Agents don’t have beliefs; models of agents do.

The Interpretation Stack

Lay the levels out and the architecture of the whole thing becomes visible:

Beliefs arise only in the modeling relation itself. They are features of how systems represent agents, not of how agents exist in the world. The stack also explains why belief-talk works at all between us: language, imitation, and theory of mind are the machinery by which agents synchronize their models across the top layers of the stack. When you tell me what you believe, you are not exporting an object from your skull; you are helping me tune my model of your self-model.

The Threshold, Reinterpreted

Now the reconciliation, and it is a reconciliation rather than a replacement. Where does the credence-threshold account live in this stack? At the self-model layer. When the victim deliberates — how likely is the weapon to be real, and is that likely enough to comply? — he is running a model of himself as a credence-bearing chooser facing a decision problem. Within that model, the threshold story is exactly correct: there is a credence, there is a stakes-sensitive threshold, and belief is the credence crossing it. The threshold account is the view from inside the self-model, and it remains the right account of what deliberation is like and of how belief guides action.

Its creaks are explained rather than merely excused. The threshold moved with the stakes because thresholds are features of decision models, and decision models are indexed to decisions — the account was tracking a real property, just not a property of the bare agent. And the missing credence-object in the skull was never a defect to be repaired by better neuroscience: the credence was always an element of a model, so looking for it among the neurons was a category error, like looking for the equator among the rocks. The two accounts I once held as rivals describe the same phenomenon from two vantage points — from within the self-model, where belief guides action, and from without, where belief explains it. The models-of-agents account is the general truth; the threshold account is what that truth looks like from the inside.

Calibration, Not Certainty

One consequence matters more than the rest. If a belief is a predictive model, its epistemic virtue is not certainty and not even truth-in-some-metaphysical-sense, but calibration: how well its predictions survive contact with reality. A well-calibrated 0.6 is a better belief than an ill-founded 0.99, for the same reason a modest map that matches the terrain beats a confident map that doesn’t. Confidence, on this view, is the expected stability of a belief under new evidence — how much you’d bet that the next observation won’t move you. And faith is confidence that refuses calibration: a model that has disconnected itself from the error signal — the fatal move examined in Against Faith. Keeping beliefs calibrated is a practice, and the discipline of updating is that practice’s name.

Beliefs are living hypotheses, not frozen propositions. Their function is to reduce surprise, not to defend identity — and most of what goes wrong in human believing is the second use crowding out the first. A belief held as a badge of belonging has stopped doing its job; it is no longer compressing the world, only signaling allegiance. What makes a belief valuable is not that it is correct full stop — nothing is correct full stop — but that it reduces error without freezing adaptation.

To believe is to model. To explain belief is to model a model. That recursive architecture is not a curiosity of philosophy of mind; it underlies cognition, communication, and social inference all the way down. Beliefs are not containers of truth but constructors of coherence, binding perception, action, and interpretation into a single predictive loop. What it takes for such a model to rise to the dignity of knowledge is the next question — taken up in What Knowledge Is.