Learning By Failing : The (Unintended) Consequence of Test Reporting on Autonomous Vehicle Training

For autonomous vehicles (AVs) to be widely commercialized, they must operate safely. Testing AVs on public roads helps improve their future safety, as it enables AV firms to discover rare and unknown driving scenarios (known as "edge cases"). However, edge cases can trigger AV failures (called "disengagements"), which make AV firms look bad to outsiders if regulators require firms to publicly report disengagements. Motivated by this learning-by-failing tension, we model a firm that aims to improve AV performance and also to impress outsiders. The firm's inherent AI capability (type-h or type-l) is private information. Both types of firms can privately choose to test their AVs in either easy areas (safe testing) or difficult areas (risky testing). Under safe testing, AVs are less likely to disengage and are thus more likely to impress outsiders, but the learning benefits are limited. Risky testing is the opposite. We develop a game-theoretic model to study how a disengagement reporting regulation affects the testing decisions of AV firms when such decisions are not perfectly observable. Without a reporting regulation, both types of firms choose risky testing. However, if a reporting regulation is present, it may distort either type of firms' incentives to choose risky testing. Contrary to popular opinion, making the testing decisions observable does not necessarily improve AV performance. From a practical perspective, our findings shed light on the discussion regarding the repercussions of a disengagement reporting regulation, and predict how such a regulation may affect the future competitive landscape of the AV industry