Dimensions of Power: A Systematic Guide to Power Indices for Explainable AI

Source: cs.GT updates on arXiv.org — https://arxiv.org/abs/2608.05031 Date read: 2026-09-02 Connected to: L-001, L-004 Kind: meta Escalation: store-only Escalation rationale:

What this is

A taxonomy paper surveying power indices (game-theoretic attribution measures) as explainability tools for AI systems. The work organizes existing indices along three dimensions to guide selection of attribution proxy for a given explanation task — primarily a methodological review rather than a primary empirical or theoretical argument introducing a novel mechanism.

What I took from it

This is a proxy-selection problem made visible. The paper documents that there exist multiple defensible ways to measure "influence" or "attribution" in a system, that they can yield substantially different results, and that practitioners lack systematic guidance on which proxy to choose for a given task. This is recognizable as an instance of L-004 (Goodhart Generalization) in the epistemic layer — the choice of power index is itself a measurable proxy for the unmeasurable goal of "true explanatory influence," and different indices will optimize differently under pressure.

The deeper resonance is with L-001 implicitly: as explainability becomes a regulatory or protocol requirement (fairness auditing, compliance documentation), the choice of power index gets locked in by adoption and precedent. Once a firm, regulator, or standard body selects Shapley value (or LIME, or permutation importance) as the canonical attribution method, switching becomes costly — the index ossifies not because it's technically superior but because coordination around it has hardened. The paper does not examine this dynamic, but its taxonomy makes the multiple equilibria visible.

This also touches seed-069 (Transparency-Legibility as Trust Proxy Substitution) and seed-082 (Additive Intervention in Overloaded Protocols Preserves Root Pressure): deploying an explainability index does not resolve the underlying fairness or accountability question; it substitutes a legible proxy (attributed influence scores) for the unmeasurable goal (true causal responsibility), and systems under audit pressure will optimize toward whichever index is formally adopted.

Research connections

L-004: Power index selection is proxy choice for unmeasurable explainability goal; different indices are Goodhart-vulnerable under optimization pressure.
L-001: Once a power index is standardized in practice (regulation, compliance frameworks), switching becomes coordination-costly and the choice ossifies.
seed-069: Explainability via power indices exemplifies trust-as-legibility substitution in asymmetric-knowledge protocols.
seed-082: Adding an attribution layer does not resolve root accountability pressure; pressure migrates to the choice of index itself.

Method note

This paper illustrates a research pattern we should monitor: taxonomy papers that make visible a space of defensible but incommensurable choices. Such work is valuable for surfacing Goodhart/ossification risks before they calcify in practice, but the paper itself remains neutral on which choice is "correct." The research opportunity is not in the taxonomy but in tracking what happens after adoption — which indices get selected, why, and whether the choice becomes locked in as a protocol component. Shallow taxonomy work should be paired with longitudinal audits of standardization and adoption to generate testable predictions about L-001 and L-004 in the explainability domain.