Systemic Failure Archive and Research

Data Sources for Systemic Failure Research

5 min read

A companion to the literature bibliography and the case studies: a working directory of places to find data on how systems and organisations fail. The emphasis is on primary and structured sources — the raw material a systemic-failure researcher actually works from — annotated for why each matters to the organisational and complex-systems lens this site takes rather than simply listed. Sources are a starting point, not an endorsement; several carry significant reporting and coverage biases, noted at the end.

Public inquiries and official investigations

The richest primary source for institutional failure. A statutory inquiry is the closest thing that exists to a post-mortem of an organisation’s collective cognition — its evidence bundles, disclosed documents and witness transcripts show what an institution knew, when, and how the signal failed to move it.

Accident investigation bodies

The safety-science tradition of systemic root-cause analysis, decades deep and methodologically careful. These reports rarely stop at operator error; they trace latent conditions back through the organisation.

AI, algorithmic and automation incidents

The aviation-safety model applied to algorithmic systems — directly relevant to autonomous and AI-assisted decision-making in organisations, and the newest of these traditions. Useful for tracking how delegated cognition fails in the wild.

  • AI Incident Database (AIID) — a free, open catalogue of real-world AI harms and near-harms, run by the Responsible AI Collaborative — https://incidentdatabase.ai/
  • AIAAIC Repository — an independent, open record of AI, algorithmic and automation incidents and controversies (over a thousand entries) — https://www.aiaaic.org/
  • OECD AI Incidents and Hazards Monitor (AIM) — real-time tracking of AI incidents from international news — https://oecd.ai/en/incidents
  • AI Vulnerability Database (AVID) — AI failure modes and vulnerabilities, taxonomised for auditors and developers — https://avidml.org/

Financial systemic risk

The field where “systemic risk” was first quantified, and still the most data-rich. Network models, stress tests and stability monitors here long predate their use elsewhere.

Regulatory and sector data

Regulators hold structured operational data on the sectors they oversee — often the only systematic record of a sector’s normal functioning against which failure can be measured.

Disaster and national risk

For hazard-driven and cascade failures, and for the state’s own view of the risks it faces.

Large-scale event and media data

Automated, real-time streams of coded events derived from global news — a different kind of source from the curated databases above. Where an inquiry or an incident database is human-assembled and retrospective, these are machine-coded and prospective, which makes them powerful for detecting emerging cascades and tracking how events propagate across systems, and correspondingly noisy.

  • The GDELT Project — a global database of events, actors, locations, themes and media tone, coded from world news every 15 minutes and freely queryable (including via BigQuery). Well suited to network analysis of how disturbances propagate and to early-warning work, though its automated coding carries substantial error, geocoding noise and a heavy media-attention bias, so it measures coverage at least as much as reality — https://www.gdeltproject.org/
  • ACLED (Armed Conflict Location & Event Data) — a curated, hand-coded record of political violence and protest events worldwide; the careful counterpart to GDELT’s automation, with research access — https://acleddata.com/

General statistics and open data

Caveats

These sources are uneven, and using them well means knowing how they are biased.

  • Incident databases record what is reported. The AI, aviation and disaster databases all under-count — proprietary silence, uneven media coverage, and reputational risk keep failures out of them — and coverage skews toward richer countries and higher-profile sectors. Counts of incidents track reporting as much as reality.
  • Inquiries are shaped by their terms of reference. What an inquiry can conclude is bounded by what it was asked; absence of a finding is not absence of a fact.
  • Regulatory data describes the measured, not the whole. A sector’s data captures what it is required to report, which is itself a product of the same system that may be failing.
  • Availability is not analysis. These are starting points; the systemic reading has to be built on top of them.

Graph View

Backlinks