Risk Redistribution

In the AI Story

The engineer who had spent fifteen years developing tacit knowledge about memory allocation, error patterns, and system failure modes is not facing more risk than before; she is facing risk in places where her intuition is silent. Her fishbowl was structured around the domains she could evaluate; AI has redistributed the risk to domains the fishbowl excluded. The discomfort this produces is not irrational — it is the accurate perception that the institutional arrangements within which she navigated risk no longer match the distribution of risk she now faces.

The insight generalizes. The translator whose job is automated has not disappeared from the labor market; she has been redistributed to a position in which her competence in language is less valuable than her competence in domains (quality control, stylistic judgment, client relations) she has less developed. The lawyer whose junior work is automated faces the same redistribution — the skills that remain valuable are the skills she has had less opportunity to develop, because the apprenticeship pattern that produced them has been disrupted. The redistribution is not neutral; it systematically disadvantages those whose competence was built in the domains AI now handles.

The governance implication is that AI regulation cannot be evaluated purely by aggregate risk reduction. A regulatory regime that reduces total AI-related harm while redistributing that harm toward populations who lack the resources to absorb it has not produced safety; it has produced a more unequal distribution of unsafety. The distribution problem is therefore internal to the safety question, not a separate concern to be addressed after safety has been achieved.

Wildavsky's resilience strategy was designed in part to handle redistribution. Fast feedback, distributed detection, and rapid correction allow redistributed risk to be detected by those who inherit it and surfaced to governance institutions that can respond. Precautionary regulation, by contrast, tends to miss redistribution because it evaluates aggregate effects rather than distributional ones. An AI governance regime that measures only total capability or total incident rates will systematically miss the redistributive effects that matter most to the people who experience them.

Origin

The concept emerges from Wildavsky's broader risk framework and has affinities with Ulrich Beck's Risk Society (1986), though Beck and Wildavsky disagreed substantially on the policy implications. Wildavsky's formulation is more analytically precise and more skeptical of the precautionary responses that Beck's framework often implies.

The concept has been extended in recent work on the 'social determinants' of algorithmic harm, which documents how AI systems redistribute risk in ways that follow existing inequality patterns — producing harms concentrated on populations whose voice in governance is already limited.

Key Ideas

Technologies redistribute, not create or eliminate. Risk is conserved in aggregate; new technologies reallocate its distribution.

Vertigo is relocation. The discomfort of a technological transition is often the accurate perception that risk has moved to unfamiliar territory, not that total risk has increased.

Competence mismatch. Individuals whose competence was built in now-automated domains face risk in domains where they are less equipped.

Distribution is internal to safety. A governance regime that reduces total harm while redistributing it toward vulnerable populations has not produced safety.

Resilience handles redistribution. Fast feedback and distributed detection surface redistributive effects; precautionary regulation systematically misses them.