The Intentional Stance

In The You On AI Encyclopedia

For AI, the intentional stance dissolves the question "does this system really think?" and replaces it with the more tractable question: "is treating it as if it thinks the most useful model?" For large language models, the intentional stance is how most users actually engage: we ascribe beliefs and goals to the model because that's what makes its behavior predictable, even though we know the underlying mechanism is statistical text prediction.

The intentional stance has become an unusually productive framework for thinking about LLMs. When a user treats a language model as if it "believes" facts and "wants" to be helpful, they are taking exactly the stance Dennett described. The practical results vindicate the approach: users who ascribe intentionality to their models manage the interaction more effectively than users who treat them as purely mechanistic. Whether the ascriptions are metaphysically accurate is, in Dennett's own framework, the wrong question.

Origin

Dennett introduced the concept in papers through the 1970s, collecting the mature form in The Intentional Stance (MIT Press, 1987).

Key Ideas

Three stances. Physical, design, intentional — each a prediction strategy, each appropriate for different systems.

Real patterns. Intentional ascriptions are not fictions: they track real regularities in behavior.

Mild realism. Dennett's position: beliefs and desires are real in the sense that intentional-stance patterns are real, even if no specific neural state maps to a specific belief.

Implication for AI. A system that rewards intentional-stance prediction is, in the sense that matters for practice, rational.

A stance is a prediction strategy, not a metaphysical claim. Dennett insists the intentional stance does not imply that the system really has beliefs; it implies only that treating it as if it does is the best available predictive move. This distinction separates Dennett from stronger realist positions about mental states — and makes his view unusually compatible with a behaviorist approach to AI mind.