Taxonomy Bias Audit¶
Audit — instantiates Implicit Bias in Knowledge Structure
A deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes.
A Taxonomy Bias Audit takes one classification scheme and works it over in depth: it reads every label, weighs each boundary, measures where cases pile up, and then follows the taxonomy's choices out to the decisions they change. Where a screen asks "does anything look off?", the audit asks "what, specifically, is off, on what evidence, and with what effect?" — and it will not close on a hunch. Its defining move is the pairing of internal structural inspection (names, defaults, residual categories, missing distinctions) with external consequence evidence (what actually happens to the cases the taxonomy handles badly). An audit that only critiqued labels would be a relabeling exercise; an audit that only traced outcomes would be a fairness report. This mechanism is the one that does both on a single named taxonomy.
Example¶
A national job board classifies every posting into an occupation taxonomy modeled on the public ONET framework. An audit is commissioned after recruiters complain that certain roles never surface in the right searches. The auditor starts *inside the structure: she inventories the top-level families, notes that "care work" is split across three unrelated branches (health, personal services, and education), and finds that a single residual node — "other services" — holds 12% of all live postings. Then she pulls the misfit cases: home health aides tagged under "personal services," doulas with no code at all, hybrid "community health worker" roles forced into whichever branch the poster guessed. Finally she traces the consequence: because the board's recommender ranks by within-family similarity, roles scattered across three families are systematically under-surfaced to the workers searching for them, and the overloaded "other" bucket is effectively invisible to search.
The audit's output is not "the labels are bad." It is a documented chain: this split + this residual overload → these misfit cases → this retrieval failure for these workers. That chain is what lets the taxonomy's owners decide whether to add a cross-cutting care dimension or accept the cost of leaving it — a decision they could not have made from the complaint alone.
How it works¶
The audit runs three passes that must connect:
- Structural inspection. Enumerate names, boundaries, defaults, sequence, and cross-references; flag skewed subdivision, missing distinctions, and any residual bucket, then measure how full each bucket is. A residual category's fill rate is the audit's single most diagnostic number.
- Misfit pull. Deliberately retrieve the cases the structure classifies badly — residual-bucket entries, forced translations, cases with no home — and read them as evidence of the hidden assumption's shape, not as noise to be cleaned.
- Consequence trace. Follow the flagged categories to the decisions they feed (search, ranking, eligibility, routing) and show, case by case, what changes because the taxonomy cuts the world this way.
The three passes are chained, not parallel: a structural flag is only sustained if a misfit pull and a consequence trace back it. That chaining is what separates a real audit finding from an aesthetic objection to a label.
Tuning parameters¶
- Depth vs. coverage — audit one high-stakes branch thoroughly or the whole tree shallowly. Deep-on-one finds real chains; broad-on-all finds candidates for deep audit later.
- Residual threshold — the fill rate at which a catch-all bucket counts as "overloaded" and triggers investigation. Set it low and everything is a finding; set it high and genuine dumping grounds pass.
- Misfit sampling reach — how hard to hunt for cases with no code at all (the invisible misfits) versus miscoded ones (the visible misfits). The invisible ones are the costlier to find and usually the more revealing.
- Consequence horizon — how many decision hops downstream to follow a category before stopping. One hop is cheap and often enough; several hops catch laundered effects but risk over-attribution.
When it helps, and when it misleads¶
Its strength is that it produces defensible findings — a structural flaw tied to real cases and a concrete effect — which is exactly what a maintainer needs to justify the cost and disruption of changing a live taxonomy. It is the mechanism that turns "this feels biased" into "here is the chain."
Its failure mode is cosmetic capture: the audit names lopsided labels, the owners rename them, and the boundaries and consequences underneath are left untouched — the same exclusion in nicer words. Bowker and Star's study of how classification schemes bury their "residual categories" and quietly torque the lives sorted through them is the standing warning here.[n1] The classic misuse is stopping the audit at the structural pass because it is the cheap one, and declaring victory on vocabulary. The guarding discipline is the chaining rule: no finding ships without a misfit pull and a consequence trace behind it, so a relabel that doesn't move the misfits or the outcomes is visibly not a fix.
How it implements the components¶
The audit realizes the evidence-gathering core of the archetype — the components that turn suspicion into a documented chain:
category_audit— the structural pass is this component: a systematic read of names, boundaries, defaults, and residual buckets on one named taxonomy.excluded_case_review— the misfit pull tests the taxonomy against exactly the cases it handles worst, reading them as evidence of the hidden assumption.downstream_consequence_trace— the consequence pass links each flagged category to the concrete decisions it changes, distinguishing structural bias from vocabulary preference.
It does not gather affected people's own accounts (stakeholder_perspective_set) — that is inclusive_classification_review.md — nor does it write the change or preserve continuity (revision_proposal, legacy_category_history), which category_revision_log.md records once a decision is made.
Related¶
- Instantiates: Implicit Bias in Knowledge Structure — the audit is the archetype's core evidence engine on a single taxonomy.
- Consumes: Excluded Case Sampling can supply the misfit cases the audit reads; the audit interprets them rather than collecting them.
- Sibling mechanisms: Bias Review Checklist · Boundary Critique Session · Category Impact Assessment · Category Revision Log · Excluded Case Sampling · Inclusive Classification Review · Ontology Critique Workshop · Red-Team Schema Review · Stakeholder Category Review
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: Taxonomy Bias Audit operates as a bounded evaluation of existing evidence or work that produces a finding or disposition because it a deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes.
Independent corroboration: The frozen evidence defines Taxonomy Bias Audit as 'A deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes', so its operative form is Assessment, Review & Assurance.
Nearest alternative: Analysis, Modeling & Optimization — Taxonomy Bias Audit includes features of an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution, but its defining operation is a bounded evaluation of existing evidence or work that produces a finding or disposition.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Library & Information Science
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Universal
Rationale: The defining operation is: A deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes. In the library_information_science lineage, that operation is specifically evidenced by authoritative or primary work that requires examining classification categories, missing groups, contextual bias, and downstream impacts of taxonomy choices. This makes library_information_science the best historical origin, while the retained alternates document contributing methods and later applications rather than being mistaken for coequal origins.
Related originating lineages:
- Computer Science & Software Engineering — Computer science and software-engineering practice supplies a parallel or contributing lineage for the mechanism's defining operation: a deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes.
- Law & Governance — Legal doctrine, regulatory governance, and procedural accountability supplies a parallel or contributing lineage for the mechanism's defining operation: a deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes.
- Linguistics & Semiotics — Linguistics and semiotics' terminology, meaning, and sign-system tradition provides a formative adjacent lineage for the same taxonomy bias audit operation.
- Sociology & Anthropology — Sociology and anthropological study of institutions and social relations supplies a parallel or contributing lineage for the mechanism's defining operation: a deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes.
- Ethics of Technology & AI Governance — tech_ethics_ai_governance supplies a historically relevant parallel or contributing practice for the defining operation—A deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes—but the evidence does not make it the best primary lineage.
Review resolution: The blind reviewers disagree on primary lineage (library_information_science versus tech_ethics_ai_governance), so I adjudicated the mechanism rather than inheriting either label. The defining operation is: A deep, evidence-driven investigation of a single taxonomy — its labels, residual buckets, and missing distinctions — that traces how its category choices shape downstream outcomes. In the library_information_science lineage, that operation is specifically evidenced by authoritative or primary work that requires examining classification categories, missing groups, contextual bias, and downstream impacts of taxonomy choices. This makes library_information_science the best historical origin, while the retained alternates document contributing methods and later applications rather than being mistaken for coequal origins. The cited NIST, Identifying and Managing Bias in AI directly supports the mechanism-specific operation and its disciplinary lineage. I retain all independently explained historical alternates without a numeric cap. origin_mode=cross_disciplinary_synthesis records how the mechanism arose; domain_reach=universal separately records how broadly it can now be applied.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
Notes¶
[n1] Geoffrey Bowker and Susan Leigh Star, Sorting Things Out: Classification and Its Consequences, documented how working classifications hide "residual categories," accrete invisible assumptions, and impose "torque" on the people and cases forced to fit them. It is the standard reference for why auditing a taxonomy means reading its residual buckets and misfit cases, not just its labels. ↩