Public AI-harm incident registry building

Construct and sustain a public, structured, longitudinally-updated database of AI-system harms and near-misses — the AI Incident Database at Responsible AI Collaborative, OECD.AI's Incidents Monitor, the AI, Algorithmic, and Automation Incidents and Controversies repository, sector-specific ledgers (facial-recognition misidentifications, hiring-tool bias findings, generative-AI hallucination incidents in medicine and law) — that other actors in the movement can index, cite, and reason from. The vehicle is a small standing organisation running a permanent submission-triage-classification pipeline, with a public taxonomy, permanent identifiers, and citation-ready records that survive individual news cycles.

An actor chooses this strategy because episodic documentation of AI harms — a single news story, a single audit — cannot support the arguments the movement needs to make against a class of technology. The frequency-and-pattern argument (this harm has happened seventy-three times, in these sectors, with these root causes) requires a cumulative record that no single actor's investigative capacity produces, and the registry is that record. The strategic ancestor is the aviation-safety incident database (NTSB, FAA, ASRS) whose forty-year accumulation is exactly what turned aviation safety from an art of individual courage into a governable engineering discipline; the movement's argument that AI needs comparable governance rests structurally on the existence of an incident record comparable to the one aviation had. The registry also gives litigation, journalism, and rulemaking a shared factual substrate they can cite in common terms, converting the movement's fragmentation on evidence into an interoperable knowledge commons.

It trades speed and salience for durability. The registry's value compounds over years — a single incident entry is not persuasive; the seventy-third entry citing the same failure mode is — and the strategy pays the cost of a slow-moving infrastructural project whose returns accrue after the movement has been running the registry for a decade. The strategy is also structurally exposed to classification politics: what counts as an AI incident, what harm-taxonomy the entries are coded into, and which incidents are surfaced or under-weighted are all political choices the vendor and the state have strong incentives to influence, and a registry captured by well-intentioned methodological choices that under-count structural harms is worse than no registry at all.

This strategy differs from strat-empirical-audit-and-expose by unit of production — audit-and-expose produces a finding about one system in one study, registry-building accumulates entries across many systems over time — and from strat-map-the-ai-supply-chain-of-state-harm by object (harms observed vs. supply-chain relationships between actors). It is the movement's slowest-compounding infrastructure play and one of the highest-yield if sustained; the strategy the ten-year-forward version of the movement will most wish had been started earlier.