Real Patterns
In his 1991 paper Real Patterns, Dennett addressed a long-standing objection to his intentional stance: if beliefs and desires are merely useful predictive postulates, are they real? His answer introduced a general criterion: a pattern is real if describing it compresses the data — if treating the system as having the pattern gives you predictive traction beyond what the raw physical description provides. On this criterion, belief-desire psychology is real (because it compresses enormous amounts of behavioral data), the stock market is real, species are real, and, when we notice that the intentional stance works on language models, the patterns it describes in the models are also real. The paper supplies the ontological footing for everything else in Dennett's framework.
In the AI Story
The paper was Dennett's response to critics — W.V.O. Quine, Paul Churchland, Jerry Fodor — who had pressed him from opposite directions. Quine and Churchland wanted to eliminate intentional talk as unscientific; Fodor wanted to make it metaphysically robust with a language of thought. Dennett's middle path was that intentional patterns are as real as any patterns ever are, and that the criterion of reality is not metaphysical substance but predictive compressibility.
The AI application is direct and consequential. When builders observe that a language model has beliefs, goals, preferences, and habits, the question is not whether these are really there in some deep sense. The question is whether the intentional description compresses. If it does — if treating the system as believing X predicts its outputs more efficiently than raw weight analysis — then the belief is a real pattern in that system, in exactly the sense that beliefs are real in human subjects.
This reframes the AI consciousness debate. Demands that the pattern be 'grounded' in something beyond predictive utility are, in Dennett's view, demands for skyhooks. Real patterns are all the reality ontology needs, and the criterion applies symmetrically across substrates.
The framework is also practically useful for interpreting AI behavior. Patterns like 'the model is more agreeable in this domain,' 'the system tends toward confident-sounding output when uncertain,' or 'Claude tracks Segal's style when drafting' are real patterns if they predict. Their reality does not require solving the hard problem; it requires doing the interpretive work that the intentional stance licenses.
Origin
Dennett published Real Patterns in the Journal of Philosophy in 1991. The paper became one of the most cited in late-twentieth-century philosophy of mind and underwrote his position against eliminativism on one side and mental realism on the other.
Its application to AI emerged in the 2010s, most explicitly in Dennett's discussions of interpretability. If the patterns the intentional stance picks out are real, then interpretability research is not merely instrumental — it is discovering real features of the systems it studies.
Key Ideas
Compressibility is the criterion. A pattern is real if describing it compresses the data better than physical description alone; metaphysical substance is not required.
Beliefs are real in this sense. The belief-desire patterns the intentional stance describes pass the criterion in virtue of their predictive power, which is enormous.
Substrate-neutral. The criterion applies identically to biological, social, and computational systems; patterns in language models can be as real as patterns in brains.
Ontology without inflation. The view avoids both eliminativism (which denies too much) and substantive realism (which demands too much) by locating reality at the level of predictive pattern.
Appears in the Orange Pill Cycle
Further reading
- Daniel Dennett, 'Real Patterns' (Journal of Philosophy, 1991)
- Daniel Dennett, The Intentional Stance (MIT Press, 1987)
- Andy Clark, Mindware (Oxford University Press, 2014)