"Death by a thousand taxonomies?": AI Risk Classification In Practice

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.06831 Date read: 2026-09-02 Connected to: L-013, L-015 Kind: meta Escalation: store-only Escalation rationale:

What this is

Empirical study of Sociotechnical Outcome Taxonomies (SOT) in AI governance, drawing on 25 interviews with researchers and practitioners. The work documents fragmentation in how risk classification systems are developed and deployed across regulatory, industry, and research contexts—a landscape of competing, incompatible taxonomic schemes rather than convergence.

What I took from it

This is a direct observational record of L-013 (Paradigm-Locked Anomaly Tolerance) and L-015 (Interpretive Continuity Decay) in action. The paper appears to document why governance systems tolerate accumulating evidence of malfunction (proliferating incompatible taxonomies) without triggering restructuring: each taxonomy represents a localized, formally coherent paradigm that resists integration with others. The fragmentation persists because the institutional investment in each scheme, and the interpretive continuity it anchors, outweighs coordination pressure.

The meta-insight is that risk classification itself becomes a coordination substrate that resists integration. Rather than taxonomies converging toward a unified legible risk ontology, practitioners and regulators maintain parallel, incommensurable schemes—each locally functional, collectively incoherent. This suggests that formalization of risk categories, rather than enabling coordination, can lock in divergence by making it costly to switch or merge frameworks.

Research connections

L-013: Established risk-classification paradigms (each SOT) tolerate accumulating evidence that they fail to integrate with other legitimate schemes, without triggering paradigm shift.
L-015: Formal taxonomic records survive and propagate while the institutional understanding of why those categories were chosen decays, producing formal coherence without semantic alignment.
seed-061: Risk taxonomy architectures may function as governance locks—once adopted, they constrain what future governance moves are legible.
seed-068: The unmeasurability of actual AI harm may insulate risk-classification systems from anomaly detection; each taxonomy succeeds locally by choosing different unmeasurable targets.
seed-073: Competing taxonomies may represent correlated failure under proxy consensus—each scheme optimizes for legibility to its stakeholder group rather than for alignment with others.

Method note

This paper suggests that studying protocol fragmentation requires ethnographic or interview-based access to why systems resist integration, not just documenting that fragmentation exists. The meta-lesson: when formal systems proliferate incompatibly, the causal mechanism often lies in institutional lock-in and interpretive path-dependence, not in technical incommensurability. Research on the new nature should look for cases where formalization increases fragmentation rather than enabling convergence—this reverses the common assumption that legibility solves coordination.