Sharon Lee Armstrong, Lila R. Gleitman, Henry Gleitman
5 min
Cognitive psychologists have long used prototypicality effects—such as the finding that some members of a category are judged to be 'better' examples than others or are verified more quickly—to argue that human concepts are organized as fuzzy, non-definitional clusters. This paper investigates whether these effects truly reveal the structure of concepts or if they are merely artifacts of how we identify category members.
The authors conducted three experiments comparing 'prototype' categories (e.g., fruit, vehicle) with 'well-defined' categories (e.g., odd numbers, plane geometry figures). In Experiment 1, they replicated a classic exemplar-rating task. In Experiment 2, they used a verification-time paradigm to see if subjects responded faster to 'good' exemplars of well-defined categories. In Experiment 3, they explicitly asked subjects whether membership in these categories was a matter of degree or all-or-none, and then had those same subjects perform the rating task again.
Across all experiments, the results were consistent: well-defined categories yielded the same graded responses and verification-time patterns as prototype categories. Even when subjects explicitly stated that membership in a category like 'odd number' is an all-or-none matter, they still produced graded ratings when asked to judge how 'good' an example a specific number was. This indicates that the experimental paradigms used to support prototype theory are not sensitive enough to distinguish between different conceptual structures.
This study challenges the common interpretation of prototypicality effects in cognitive psychology. By showing that these effects occur even for concepts with clear, logical definitions, the authors argue that researchers must distinguish between the 'core' conceptual structure (the definition) and the 'identification procedures' (the heuristics used to quickly categorize items). This distinction is crucial for developing a valid theory of human conceptual organization.
Sam: And the behavioral dissociation they find in the final experiment really closes that off. Subjects explicitly insist membership is all-or-none, yet still provide graded exemplariness ratings in the same session.
Alex: [precise] That's the crucial observation. It proves category membership and exemplariness are psychologically distinct. People hold a firm categorical definition while simultaneously using a graded heuristic for quick-reference tasks. The "fuzziness" isn't in the concept—it's in the interface. We've been measuring the efficiency of our mental indexing systems and attributing those performance characteristics to the logical structure of the concepts themselves.
Sam: Which raises the harder question. If prototype theory is just a study of performance artifacts, what does that leave us with for mental representation? [[RP_SECTION:dual-process-conceptual-models|Dual process conceptual models]]
Alex: [measured] It forces a dual-process picture. The core handles logical entailments—the things that follow necessarily from category membership. The identification procedure handles the efficiency of everyday interaction. They serve different functions and are answerable to different constraints. The problem is the field has been using evidence from one to make claims about the other. [[RP_SECTION:critique-of-feature-decomposition|Critique of feature decomposition]]
Sam: And the authors are skeptical there's even a clean feature decomposition waiting to be found for those cores. They make the point that dictionaries define words using other words—there's no bedrock of primitive semantic atoms you eventually hit. The search for a universal feature set might be a category error.
Alex: [steady, analytical] That's the crux of their skepticism. If concepts don't decompose into simple features, then the research program built around finding those features isn't just incomplete—it's aimed at the wrong target. What they advocate instead is investigating specific, highly structured domains. They point to universal grammar as a model: rich, domain-specific, and not reducible to a simple feature list.
Sam: So rather than a grand unified theory of categorization, we might be looking at a collection of distinct, complex systems with no single underlying architecture.
Alex: [calm] That's the reality they present. The search for a master theory built on feature structures likely underestimates the actual complexity of human conceptual knowledge. It's a sobering conclusion—not because the questions go away, but because the methodology we've been using to answer them turns out not to be fit for purpose.
Sam: A 1983 paper, and the methodological point still stands as a live challenge for anyone running typicality paradigms. Thanks for walking through it.
Alex: Thanks for listening to ResearchPod.