Last updated: 2026-10-08

U
Undergraduate level

Intentionality: What Makes a State About Something?

A weather vane turns to face the wind. A thermostat's bimetallic strip bends with temperature. A sentence, "it's windy today," is also reliably produced when it's windy. All three co-vary with some state of the world in a lawful, trackable way — and only one of them is naturally described as being about the wind, in the sense a belief is about something. The weather vane doesn't represent the wind being from the north; it just physically responds to it. Finding the actual difference between a state that merely correlates with something and a state that is genuinely about it — call this the problem of intentionality — turns out to sit directly underneath several of this series' other questions, including what Dennett's intentional stance (the previous page) is actually attributing when it attributes a belief.

Brentano's Mark of the Mental FoundationalKnowledge that endures for decades — core principles

Brentano revived a medieval scholastic term for exactly this property and proposed it as what separates mental phenomena from physical ones: every mental state is directed at an object — a belief is about something, a desire is for something, a fear is of something — and, strikingly, that object need not exist. A person can fear a monster under the bed that isn't there, or believe something false, and the fear or belief is still directed at that (non-existent, or false) object as determinately as if it were real1. No ordinary physical relation works this way — a weather vane cannot point at a wind direction that doesn't exist, because physical causation requires an actual cause. That one asymmetry is the whole puzzle: intentionality is a relation that behaves as though it reaches an object, even when there's no object there to reach.

Original Versus Derived Intentionality FoundationalKnowledge that endures for decades — core principles

Searle's development of the problem distinguishes two very different ways a thing can be "about" something2. A person's belief that it's windy has original (or intrinsic) intentionality — it's about the wind in its own right, not because anyone else interprets it that way. The word "windy," printed on a page, has derived intentionality — it's about wind only because speakers of English have agreed to use it that way; the ink marks themselves carry no aboutness at all, and a reader who didn't know English would see only marks. A photograph, a map, a weather vane's reading, and — Searle argues directly — a running computer program all sit on the derived side of this line: whatever meaning they carry is meaning assigned by an interpreting mind standing outside them, not meaning the artefact generates for itself.

This is the same distinction doing the work in Searle's Chinese Room argument, already discussed on this site in connection with a different self-modelling question3: a system that manipulates symbols according to rules, however complex those rules and however fluent the output, has only derived intentionality by construction, because nothing in syntactic symbol-manipulation adds up to the symbols being about anything to the system itself. The aboutness, on this view, is supplied entirely by the external interpreter reading meaning into the input and output.syntax, however elaborate, is not semantics

A Direct Reply From the Previous Page FoundationalKnowledge that endures for decades — core principles

Dennett's account of intentional systems can be read as a direct challenge to the original/derived distinction itself, not merely as a different theory sitting alongside it. If a belief is real exactly when attributing it affords a genuine predictive compression — the account Belief Without an Inner Believer? develops in full — then "original" intentionality may not be a different kind of thing from "derived" intentionality at all. It may just be derived intentionality that has been bootstrapped through enough layers of interpretation, feedback, and self-correction that no single external interpreter is doing identifiable work anymore. On this view, asking whether a sufficiently sophisticated system has "real" (original) aboutness or "merely" derived aboutness presupposes a sharp line that a graded, bootstrapping process may not actually respect.

Note well. Whether original intentionality is a distinct kind of thing, or derived intentionality elaborated far enough to look different, is not settled by either side simply restating its own position with more confidence. It is the actual disagreement.

A Naturalistic Alternative: Content as Causal Correlation FoundationalKnowledge that endures for decades — core principles

A third family of views tries to ground intentionality in ordinary physical relations after all, rather than treating it as a special, irreducible mental fact or deflating it into a predictive pattern. Causal and informational theories of content propose that a state is about X when it stands in the right kind of reliable, lawful causal relationship to X — roughly, a state that reliably occurs when and only when X is present, under the conditions the state was "designed" (by evolution, or by training) to track. This would make a thermostat's bimetallic strip genuinely about temperature after all, in at least a thin sense, while still needing extra machinery to explain why a belief can be about something that doesn't exist, which a purely causal story struggles with directly — there is no actual monster under the bed to stand in a causal relation to the fear of it.

Why This Isn't Merely Academic for AI Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

A language model's internal representation of "Paris" reliably activates in the presence of text about Paris, reliably supports correct answers about Paris, and reliably co-occurs with related representations (France, Eiffel Tower, capital city). By a purely causal-correlational standard, that representation is about Paris in exactly the thin sense a thermostat's strip is about temperature. By Searle's standard, it has at most derived intentionality — about Paris only because human readers interpret its outputs that way, with nothing on the inside for which Paris is genuinely a topic. Which standard is the right one to apply is not a settled question this page can close, and later pages — directly, when this series takes up what a trained representation is actually latching onto — return to it with the vocabulary this page has now put in place.

Strongest Objection: The Whole Distinction May Be Doing No Work FoundationalKnowledge that endures for decades — core principles

A deflationary critic can reasonably ask what empirical difference "original versus derived intentionality" is supposed to predict, if any. If no observable test distinguishes a system with original aboutness from one with derived aboutness that's merely very well bootstrapped, the distinction risks being unfalsifiable in exactly the way Chalmers' hard problem of consciousness is sometimes accused of being. This page does not resolve that worry. It notes that the same structural shape — a distinction that feels obviously real from the inside and resists being cashed out in a decisive external test — recurs across this series, and is worth tracking as a pattern in its own right rather than refighting separately each time it appears.

Provisional Conclusion FoundationalKnowledge that endures for decades — core principles

A weather vane, a word, and a belief all manage to be "about" something, but not in the same way or by the same mechanism, and collapsing the three together is where casual talk about what an AI system "represents" usually goes wrong first. Whether a trained representation's aboutness is more like the word's (derived, interpreter-dependent), the thermostat's (thin, causal), or something that doesn't fit either template cleanly, is a live question this series treats as open rather than settling by definition.

Questions for Further Thought

  • Could a causal-correlational account of content ever explain a state being about something that doesn't exist, the way Brentano's examples require?
  • If "original" intentionality turned out to be derived intentionality with enough bootstrapping, would that change anything practical about how AI systems should be built, evaluated, or described?
  • Does a language model's representation of "Paris" more closely resemble a word, a weather vane, or neither?

Further Reading

  • Brentano, F. (1874). Psychology from an Empirical Standpoint. (English translation: Routledge, 1995.)
  • Searle, J. R. (1983). Intentionality: An Essay in the Philosophy of Mind. Cambridge University Press.
  • Jacob, P. (2023). Intentionality. In E. N. Zalta & U. Nodelman (Eds.), The Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/intentionality/

References

  1. Brentano, F. (1874). Psychology from an Empirical Standpoint. (English translation: Routledge, 1995.) ↩

  2. Searle, J. R. (1983). Intentionality: An Essay in the Philosophy of Mind. Cambridge University Press. ↩

  3. Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–457. ↩