Could an International AI Incident Database Work Like Aviation Accident Reporting?

Photo: Airman 1st Class Kathrine McDowell, U.S. Air Force, Wikimedia Commons, Public domain

01 What aviation actually built

Commercial aviation is the one industry that turned failure into a public good. After the 1956 Grand Canyon mid-air collision killed 128 people, the United States created what became the NTSB model: an investigator independent of the operator, the regulator and the manufacturer, empowered to publish findings that blame no one but change everything. The international layer, ICAO Annex 13, harmonizes accident investigation across borders; confidential no-blame reporting systems — from the 1970s onward — let pilots confess near-misses without fearing their licenses. The result is a system where every crash anywhere makes every flight everywhere safer.

Three design principles do the work. Independence: investigators do not answer to the people whose product failed. No-blame: the report's job is probable cause and safety recommendations, not liability — which is why pilots, dispatchers and now mechanics self-report. Publication: findings are public, searchable and usable by competitors. Fatal commercial accidents fell from roughly 40-plus a year in the early 1970s to single digits today — a decline that tracks the maturing of exactly this reporting loop.

Analysis — not prediction. N43 and Hermes AI grounds every scenario in the documented record and verified reporting as of September 21, 2026; where evidence is incomplete we say so.

02 The closest thing AI has today

AI's nearest equivalent is the AI Incident Database — incidentdatabase.ai, run by the Responsible AI Collaborative with roots in Harvard's Berkman Klein Center — which has collected more than 3,000 reports since 2019: biased hiring screens, chatbot harms, wrongful arrests driven by facial recognition, model-assisted fraud. It is a serious piece of civic infrastructure, and it is entirely volunteer-curated. Nobody must file anything. Coverage skews toward harms that reach the press, and almost nothing arrives from inside the companies that build frontier systems.

As the anchor video for this analysis puts it, the database undercounts harms, and the fix is better reporting standards — the exact gap the aviation model closes with mandates. The formal scaffolding is starting to appear: the NIST AI Risk Management Framework gives organizations a vocabulary for cataloging AI failures, and the EU AI Act imposes the first genuine legal incident-reporting duties on providers and deployers of high-risk systems, with serious-incident reports required to the authorities and enforcement phasing in through 2026. But a duty to report to a regulator is not yet an independent, public investigation of the NTSB kind.

03 Why the NTSB model is hard to copy for AI

Aviation investigations work because crashes have physical evidence: wreckage, recorders, radar, bodies. AI incidents are different in kind. A model's failure often leaves only a text file, a log line and an argument about what the system intended. The same model behaves differently across prompts, users and updates — the “airframe” changes silently every time weights are refreshed, which is as if Boeing re-welded every 737 overnight and told no one. Reconstructing probable cause after the fact is not harder; it may be impossible.

Independence collides with the industry's defining structure. Aviation had airlines that could survive a public finding against a plane they bought; frontier AI has a handful of labs whose competitive position rests on undisclosed model details. No-blame reporting collides with litigation: the moment an incident report names a system, it is discovery material. And publication collides with national security, because the most consequential AI deployments — cyber, military, intelligence — are precisely the ones no government will investigate in public. Each principle survives; each survives in a weakened form.

04 What a workable regime would need

A plausible AI-incident regime looks less like the NTSB and more like a hybrid. Standardized report formats: ICAO-style taxonomy — what system, what task, what harm, what deployment context — so that incidents are comparable across companies and years; the NIST AI RMF and emerging ISO work provide raw material. Confidential channels with public aggregate findings: individual reports protected, patterns published, borrowing from aviation's protected safety data rather than its crash docket. Independent technical investigators with cleared access: small standing bodies able to receive model weights, logs and deployment records under protective order, as financial regulators already do with bank books.

The EU AI Act is the live experiment: its serious-incident reporting duty is the first time any jurisdiction has obliged AI providers to file at all. Whether those reports stay sealed compliance artifacts or become the raw material for a genuinely public safety literature is the decision that determines if the aviation analogy holds. The precedent for hope is that aviation reporting, too, began as proprietary and liability-terrified — and the industry discovered that shared failure data was cheaper than shared crashes.

05 The liability freeze problem

The deepest obstacle is not technical, it is legal. Aviation's no-blame principle required carving investigators' findings out of the courtroom — in the U.S., NTSB probable-cause findings are largely inadmissible in civil litigation. AI has no such carve-out. Every incident report is a plaintiff's exhibit, which is why companies describe failures in carefully hedged post-mortems, if at all. Until reporting is legally safe, the honest comparison is: AI today is aviation in an alternate history where every cockpit voice recorder is turned over to the injured passengers' lawyers within 30 days.

There are emerging patches. The EU AI Act's reporting duties come with regulatory confidentiality regimes; some U.S. legislative proposals have floated limited liability protection for safety filings; and insurance — as the parallel N43 analysis on AI underwriting explores — could price the risk of honest disclosure. None is yet the clean legal wall aviation built. The question for policymakers is blunt: without one, a mandatory AI incident database collects exactly the incidents that are already public, and none of the ones that matter.

06 What to watch next

Watch the EU AI Act's first serious-incident enforcement: the volume and quality of reports filed under the high-risk-system duties will be the first real data on whether mandated AI reporting produces signal. Watch the AI Incident Database's growth curve: if volunteer reports keep doubling even as formal duties arrive, that says the official channels are capturing little. Watch NIST and ISO taxonomy work, because standardized incident classes are the precondition for any international comparability, and watch whether any government creates an independent AI investigation body with cleared access — the single clearest marker that the NTSB model is being attempted rather than admired.

The aviation analogy is seductive because it is true: reporting is what made the safest transport system in history. It is also a reminder that the regime took a Grand Canyon-scale catastrophe to get started. AI is trying to build the same machinery before its equivalent event — the harder path, and the wiser one.