Feigenbaum test¶
A proposed domain-specific variation of the Turing test in which a computer's performance is judged by whether it can reproduce the behavior or output of a recognized human subject-matter expert.
Core Idea¶
The Feigenbaum test is a proposed subject-matter-expert variant of the Turing test. Instead of asking whether conversation appears human in general, it asks whether a computer can reproduce the task performance or outputs of a recognized expert in a field such as chemistry or marketing. A meaningful test needs a bounded domain, representative tasks, competent human comparators, controlled access to outputs, and explicit scoring. A meaningful test needs a bounded domain, representative tasks, competent human comparators, controlled access to outputs, and explicit scoring.
Scope of Application¶
Use Feigenbaum test for the named proposal and state field, tasks, expert selection, blinding, scoring, and inference limits. Use Feigenbaum test for the named proposal and state field, tasks, expert selection, blinding, scoring, and inference limits.
- Artificial intelligence. Evaluates domain performance.
- Expert systems. Compares encoded expertise.
- Professional assessment. Defines expert benchmarks.
- Human-computer studies. Tests evaluator judgments.
- AI history. Tracks proposed intelligence tests.
Clarity¶
Expert-like output can reflect imitation, retrieval, or task competence without establishing the same internal understanding. The closest near miss sets the boundary: The ordinary Turing test is closest: it evaluates human-like conversational indistinguishability broadly, while the Feigenbaum variant narrows the comparator to field expertise. A positive case must satisfy this test: An evaluation is a Feigenbaum test when a machine and recognized expert are compared on controlled tasks within a declared specialty for expert-level behavioral or output equivalence.
Manages Complexity¶
Results depend strongly on case sampling, expert disagreement, evaluator blinding, tool access, and scoring. One narrow pass should not be generalized beyond the tested field and conditions. The central observable equivalence–internal competence tradeoff is this: Matching outputs does not reveal reasoning mechanism. A second expert standard–expert disagreement tension matters because Recognized specialists can disagree on hard cases.
Abstract Reasoning¶
Use three linked moves: bound the subject-matter domain; select qualified human comparators; build representative controlled tasks. As a collapse test, the case exits when no expert comparator, bounded subject domain, or controlled equivalence judgment is present. A fourth check is to blind and score outputs under one rubric. A final check is to limit conclusions to observed expert-level equivalence.
Knowledge Transfer¶
Comparator-based imitation tests transfer across professions, but the named expert-domain framing and historical proposal delimit the Feigenbaum test. The nearest stopping boundary is explicit: The ordinary Turing test is closest: it evaluates human-like conversational indistinguishability broadly, while the Feigenbaum variant narrows the comparator to field expertise. The inclusion test remains: An evaluation is a Feigenbaum test when a machine and recognized expert are compared on controlled tasks within a declared specialty for expert-level behavioral or output equivalence. The structure no longer applies when the case exits when no expert comparator, bounded subject domain, or controlled equivalence judgment is present. No canonical parent prime is currently asserted; broader structural comparisons remain related-prime analogies until separately adjudicated in the DAG. The proposal narrows imitation to expertise. A task set operationalizes the comparison.
Neighborhood in Abstraction Space¶
Feigenbaum test sits in a crowded region of the domain-specific corpus (37th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.
Family — Group Dynamics & Collective Behavior (19 abstractions)
Nearest neighbors
- Clinical Equipoise — 0.89
- Face validity — 0.89
- Preventive action — 0.88
- Social comparison bias — 0.87
- Showdown Cooperative Learning — 0.87
Computed from structural-signature embeddings · 2026-10-08