AI Incident Monitoring through a Public Health Lens

Abraham, Chen, Chhun, Jaramillo-Gutierrez, Mylius, Raaj, Slattery, McGregor (arXiv:2604.19914, submitted 2026-04-21). Tier T1 — a methodology paper, not a product or compliance claim; one author (McGregor) built the AIID itself (ai-incident-database), so this is partly the AIID's own team answering the problem their earlier paper named.

The problem this directly answers

ai-incident-database's source page asks growth edge 23's question and finds no denominator anywhere: raw incident counts conflate media attention, deployment scale, and true harm rate, and nothing in the corpus separates them. This paper is the first source here that proposes a method rather than restating the gap.

The method: borrow from public-health surveillance

Public-health surveillance assumes its own data is incomplete and noisy, and has built tools for exactly that case. The authors define six phases of incident emergence and, critically, an exposure denominator — a domain-specific count of how much the relevant system is actually deployed or used, so raw incident counts can be converted into a rate.

Worked case: autonomous vehicles. California DMV mandatory-reporting records give autonomous miles driven as the denominator. Incident rates are expressed per mile traveled, not as a raw count. Mandatory reporting gives this case "dependable incident-rate ground truth" that AI incidents generally lack.

Second case: deepfakes. A separate case study outside the mandatory-reporting domain; the abstract does not specify its denominator in enough detail to extract without the full paper.

What it does not fix

This is one worked example (AVs) with a government-mandated reporting regime most AI systems do not have. The authors are explicit: without the exposure data, "policymakers, companies, and the public lack a means to weigh the benefits of AI against their in-context risks" — a statement about the field generally, not a claim that this paper supplies the missing denominator everywhere. eu-ai-act Article 73 creates a reporting duty but not (yet) a published exposure count to pair with it.

Verdict on growth edge 23

Half-close. The edge asked "is there a denominator anywhere?" — and for one domain (AVs, under mandatory reporting), yes, and this paper shows the method for building one. For the AI systems most governance instruments actually target — general-purpose models, no mandatory reporting — the method exists but the input data (deployment-scale counts) is still unpublished. The residue: does any AI system without a mandatory-reporting regime have a published exposure denominator anywhere, even an imperfect proxy (API call volume, active-user counts)?

Related

ai-incident-database · eu-ai-act · oecd-ai-principles · partnership-on-ai