Salako & Muhammad: Coarse AV data yields dangerously optimistic safety claims

Kizito Salako and Rabiu Tsoho Muhammad's new paper, "The Impact of Operational-Data Fidelity when Assessing Safety-Critical Autonomous-Vehicle Software," tackles a core problem in AV safety validation: what happens when the historical data used to certify reliability omits important details about past failures. The authors extend conservative Bayesian inference (CBI)—a statistical technique that deliberately places cautious priors on failure rates—to model scenarios where operational data is only available at a coarse level of granularity. Their results, published on arXiv with 16 pages and 12 figures, reveal a troubling trend: even when assessors apply conservative methods, low-fidelity data can yield reliability claims that are significantly more optimistic than the underlying evidence supports.

The study's key contribution is a principled framework for checking the robustness of reliability claims against data insufficiency. By simulating AV operational histories at different fidelity levels and applying their extended CBI approach, the researchers demonstrate that missing detail about failure precursors or event sequences can hide emergent risks. They provide the first conservative estimates of how much data fidelity degrades assessment accuracy, showing that naive attempts to 'use low-fidelity data conservatively' can backfire. For AV developers and regulators, this means that simply collecting more miles of driving data isn't enough—the data must capture the operational context needed to distinguish benign events from latent safety-critical failures. The work reinforces earlier findings on statistical model fidelity but sharpens the message: data resolution is as important as data volume for trustworthy AV safety case.