Estimating AI Harm Rates for Germany: Applying Epidemiological Methods to Incident Monitoring
Mentor: Branwen Nia Owen
Project area: The epidemiology of AI risk
Project Language
Minimum Time Commitment
8 hours per week.
Project Abstract
AI incident databases are gaining traction, but raw counts don't tell policymakers much. A rise in reported harms could mean genuine risk escalation — or just improved reporting, or increased AI deployment. To distinguish these, you need both a numerator (how many harms actually occurred, adjusted for reporting gaps) and a denominator (how many opportunities for harm existed).
This project adapts epidemiological methods to estimate AI harm rates for incident types relevant to German policy priorities — such as hiring algorithms under anti-discrimination law, clinical decision support in the German healthcare system, or content moderation on platforms operating under the DSA.
The work has two components:
1. Exposure estimation: How many opportunities for harm existed? Depending on data availability, we use direct measurement (e.g., DSA transparency reports), category-based assignment, decomposition into bounded components, or structured expert judgment.
2. Harm estimation: How many incidents actually occurred? Reported counts underestimate true incidence. We apply methods from disease surveillance — such as capture-recapture across overlapping sources and multiplication factors from analogous domains — to adjust for underreporting.
Mentees will:
- Select incident types relevant to German AI governance;
- Estimate both exposure and harm counts, with documented assumptions;
- Calculate crude harm rates and assess trends;
- Produce standardised output cards suitable for policy communication;
Expected outputs include worked examples for German-relevant incident types, a practical methodology guide, and a gap analysis identifying where better data collection would most improve monitoring capacity.
Theory of Change
Bad frameworks produce bad decisions. The question of machine moral status will increasingly affect AI development and governance. Currently, most people reasoning about it lack adequate conceptual tools. This matters for catastrophic risk in several ways.
Under-reaction: if AI systems develop welfare-relevant internal states and we lack frameworks to recognize this, we may create systems with misaligned interests while dismissing their signals as "mere computation." A system that experiences something like suffering under certain conditions, and whose operators dismiss this, is a system with reason to deceive.
Over-reaction: anthropomorphizing systems that lack morally relevant properties wastes attention and resources, and may constrain beneficial AI development without corresponding benefit.
Poor discourse: without shared conceptual foundations, public debate about AI consciousness polarizes between dismissive and credulous positions. Neither serves good governance.
The primer addresses these by training researchers and practitioners to reason carefully across multiple frameworks, recognize what each assumes, and navigate uncertainty without false confidence. The German focus (incorporating European philosophical traditions, piloting with German-speaking users) builds SAIGE's national infrastructure while contributing to the broader field.
Conceptual clarity is infrastructure. This project builds it.
Desired Mentee Background
Any or all, it's more about skills and resourcefulness than a field of study.
Desired Mentee Level of Education
Masters and above.
Other Mentee Requirements
* Comfort with quantitative reasoning (e.g., interpreting rates, proportions, uncertainty)
* Familiarity with basic statistics (confidence intervals, bias, sampling, regression at a conceptual level)
* Ability to read technical papers and policy documents
* Comfort working under ambiguity and with incomplete data
* Ability to produce clear, policy-facing summaries (not just technical notes)
* Pragmatic, impact-oriented mindset