Last updated: 2026-10-08
Belief Without an Inner Believer? Dennett and Artificial Agency
This series opened with a thermostat that seems to want the room warmer and believe it's currently too cold, and the feeling that this description is too generous without an easy account of why. Dennett's answer doesn't settle whether the thermostat really has beliefs. It changes the question: instead of asking whether a system has an inner belief in some deep metaphysical sense, ask whether describing it as having beliefs is the most useful way to predict what it will do next — and then ask what "most useful" is doing in that sentence, because the answer turns out not to be a cop-out.
Three Stances, Not One FoundationalKnowledge that endures for decades — core principles
Dennett's 1971 paper distinguishes three strategies for predicting a system's behaviour, each costing more effort than the last to set up and each buying more predictive leverage once it's running1. The physical stance predicts from the system's physical composition and the laws governing it — exact, but for anything complex, computationally hopeless; nobody predicts a chess program's next move from the voltages in its transistors. The design stance predicts from what the system was built to do, assuming it works as designed — a calculator will add correctly because that's what calculators are for. The intentional stance predicts by attributing beliefs and desires and assuming the system acts rationally in light of them — treating a chess program as though it wants to protect its queen and believes a particular move achieves that, without tracing a single line of its source code.
For a simple thermostat, the design stance is already cheap and exact enough that the intentional stance buys nothing extra — which is the actual answer to the opening puzzle. For a chess program, a weather system, or another person, the design stance either isn't available (we don't have the source code) or isn't tractable (there's too much of it), and the intentional stance becomes not just convenient but the only strategy that predicts well at a cost anyone can afford. Legitimacy, on this account, is not a yes/no fact about whether a system secretly has beliefs. It's a graded fact about how much predictive work the intentional stance does for a given system that the cheaper stances can't do as well.a cheaper lie that predicts as well as the expensive truth
Competence Without Comprehension FoundationalKnowledge that endures for decades — core principles
A further distinction, developed across Dennett's later work, keeps the intentional stance from collapsing into "anything that behaves usefully has a mind": a system can be competent at a task — reliably producing the right output — without comprehending anything about why that output is right2. A bacterium competently navigates toward a nutrient gradient with no comprehension of chemistry, and a thermostat competently holds a temperature with no comprehension of heat. Comprehension, on this view, is not a separate ingredient bolted onto competence — it's a further, harder-won achievement built out of layers of competence that can themselves explain, represent, and reason about their own operation, rather than merely executing it. The question "does this system understand what it's doing" is really several questions stacked on top of each other, and a system can clear the lower ones without clearing the higher ones.
Does the Intentional Stance Reveal Real States, or Only a Useful Strategy? FoundationalKnowledge that endures for decades — core principles
The obvious worry: if attributing a belief is just whatever predicts well, intentional-stance talk is a convenient fiction, true of nothing in particular. Dennett's own answer resists both horns of that dilemma. A pattern is real, on his account, if it affords a genuine compression of the data — if treating a system as having beliefs and desires lets you predict its behaviour with fewer bits of description than tracking its physical state would require, and that compression actually works, repeatedly, across new cases3. A belief attributed this way is real in the same sense a constellation, a wave, or a species is real: not an illusion, not a fundamental physical particle either, but a genuine pattern that a coarser level of description would miss and a finer one would drown in noise.
What Ordinary AI Language Reveals and Conceals Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
Phrases already in everyday use around deployed systems can be read through this framework directly, rather than dismissed as sloppy or defended as literal:
| Phrase | What it plausibly tracks | What it risks implying |
|---|---|---|
| "The model knows X" | X is reliably retrievable from training or context in a way that predicts correct output | A stored, justified belief the system could explain how it acquired |
| "The agent decided to Y" | Y was selected over alternatives by a process sensitive to some goal representation | A deliberated choice weighing reasons the way a person would |
| "The system wants Z" | Behaviour reliably moves toward outcomes resembling Z across varied situations | An experienced desire, felt as a pull toward Z |
| "The model is confused" | Output is inconsistent in a way that tracks a recognisable failure mode | A subjective state of not-understanding, rather than a pattern of error |
| "The assistant remembers" | Earlier context measurably affects later output | An episodic recollection, experienced as recall rather than computed as context |
The middle column is where the intentional stance is doing real predictive work and the pattern is arguably real in Dennett's sense. The right-hand column is where usage quietly imports claims — about experience, justification, deliberation — that the intentional stance's success never established and was never designed to establish.
Strongest Objection: This Makes Attribution Too Cheap FoundationalKnowledge that endures for decades — core principles
If legitimacy is just "whatever predicts well," a sufficiently elaborate thermostat description could eventually earn intentional-stance treatment too, and the account seems to have no principled floor. Dennett's reply is that the floor is empirical, not stipulated: the intentional stance earns its keep only where it demonstrably out-predicts the design stance at a comparable cost, and for genuinely simple systems it never does, no matter how elaborately the description is dressed up. The test isn't whether intentional language can be applied — it almost always can, to almost anything, as the opening thermostat case shows — it's whether applying it buys anything the design stance didn't already give for free. That's a real, checkable fact about a specific system in a specific predictive task, not a matter of descriptive taste.
Two Competing Pressures FoundationalKnowledge that endures for decades — core principles
Eliminativism pushes from one side: folk-psychological vocabulary (belief, desire, wanting) is a pre-scientific placeholder that a mature neuroscience or computational theory should eventually replace outright, the way phlogiston talk was replaced rather than reinterpreted. Robust realism about mental states pushes from the other: some theories hold that beliefs are discrete, physically implemented structures (a language of thought, on one well-known proposal) rather than patterns that are real only in Dennett's deflationary sense. Dennett's instrumental realism sits deliberately between them — patterns are real, but not in the way either eliminativism or a language-of-thought view would want. "The intentional stance" names one specific position in that three-way dispute, not the only serious option.
Provisional Conclusion FoundationalKnowledge that endures for decades — core principles
Whether "the model knows X" is a legitimate description doesn't turn on peering inside the system for a hidden belief-token. It turns on whether the attribution compresses and predicts the system's behaviour better than cheaper descriptions do, and on tracking which half of that attribution — the predictive pattern, or the experiential furniture ordinary language quietly drags along with it — is actually doing the work in a given sentence.
Questions for Further Thought
- Is there a current AI system for which the design stance (knowing exactly how it was built) still out-predicts the intentional stance, despite its complexity?
- Does a system's own ability to use intentional-stance language about itself change anything about whether the stance applies to it?
- If "real pattern" doesn't require a physical token corresponding to each belief, what distinguishes a real pattern from a merely convenient one?
- Which of the five phrases in the table above would you be most comfortable using about a system you'd built yourself, and why that one specifically?
Related Topics
- What Is the Philosophy of Artificial Intelligence? — the thermostat this page's opening directly answers.
- The Deception Criterion — a system successfully adopting the intentional stance toward itself is a different question from a system successfully inducing a false belief in an observer; the two are easy to conflate.
- Does a Self-Model Get You Anywhere Near an Analogue of Consciousness? — a related but distinct use of Dennett's work (his argument against classical qualia) applied to a specific self-modelling architecture.
Further Reading
- Dennett, D. C. (1971). Intentional systems. The Journal of Philosophy, 68(4), 87–106.
- Dennett, D. C. (1987). The Intentional Stance. MIT Press.
- Dennett, D. C. (1991). Real patterns. The Journal of Philosophy, 88(1), 27–51.
- Dennett, D. C. (2017). From Bacteria to Bach and Back: The Evolution of Minds. W. W. Norton.