Attribute Substitution

In the AI Story

Kahneman formalized attribute substitution with Shane Frederick in 2002. The mechanism has three requirements: the target question must be computationally difficult; a heuristic attribute (an easier related question) must be cognitively accessible; and System 2 must fail to override the substitution. All three conditions are routinely satisfied in everyday judgment.

In human-AI collaboration, a taxonomy of substitutions emerges. The question is this argument valid? gets substituted with does this argument sound persuasive? The question is this the right decision? gets substituted with does this decision feel comfortable? The question does this code work correctly? gets substituted with does this code look clean? The question have I understood this concept? gets substituted with can I recognize this concept when I see it?

The machine contributes its own substitution. Claude frequently reframes the question implicitly in its response — narrowing scope, shifting emphasis, replacing the original question with one the model can address more completely. The reframing is usually subtle and embedded in the response rather than announced. The human absorbs the reframing along with the content, and the question asked diverges from the question answered, with the divergence concealed by the fluency of the response.

Detection requires a specific meta-cognitive operation: stopping after receiving the response and asking what question did I ask, and what question was actually answered? This question-about-questions is effortful and uncomfortable. System 1 prefers the satisfying coherence of the received answer to the unsettling possibility that it answers something slightly different than what was asked.

Kahneman's observation that machines could in principle address hard questions directly — rather than substituting — is undercut when the human evaluates the machine's answer using substitution heuristics. The benefit of the machine's literalism is lost in the human's evaluative step.

Origin

The concept crystallized from experimental work in the 1990s on preference construction, framing, and the surprising ease with which people's answers to difficult questions tracked whatever easier question happened to be accessible. Kahneman and Frederick's 2002 chapter "Representativeness Revisited" provided the formal account.

The name is deliberately technical. Kahneman chose the term to emphasize that the phenomenon is structural — an operation the cognitive system performs — rather than a failure of motivation or intelligence.

Key Ideas

Seamless swap. The substitution happens without awareness; the answerer experiences herself as having answered the original question.

Easier attribute available. The substituted question must be cognitively accessible through heuristics the mind already uses.

Fluency conceals the swap. A well-written answer to an adjacent question feels like a well-written answer to the asked question.

AI reframes implicitly. The machine often converts the question in its response, and the human absorbs the conversion.

Detection is meta-cognitive. Only the question-about-questions — what was asked vs. what was answered — reveals the substitution.