The Dilemma of Uncertainty Estimation for General Purpose AI in the EU AI Act

TL;DR

The paper addresses uncertainty estimation as a measure for compliance and quality assurance of GPAI under the EU AI Act. It surveys uncertainty estimation methods (ensembles, MC-Dropout, and others), analyzes their computational costs, and discusses how systemic-risk classification based on compute thresholds like $10^{25}$ FLOPs interacts with practice. It argues that uncertainty estimates can enhance risk management, dataset governance, documentation, and human oversight, while acknowledging a key tension: additional computation could trigger stricter regulatory regimes. The work calls for standards development and careful policy design to balance technical feasibility with the Act's accountability goals.

Abstract

The AI act is the European Union-wide regulation of AI systems. It includes specific provisions for general-purpose AI models which however need to be further interpreted in terms of technical standards and state-of-art studies to ensure practical compliance solutions. This paper examines the AI act requirements for providers and deployers of general-purpose AI and further proposes uncertainty estimation as a suitable measure for legal compliance and quality assurance in training of such models. We argue that uncertainty estimation should be a required component for deploying models in the real world, and under the EU AI Act, it could fulfill several requirements for transparency, accuracy, and trustworthiness. However, generally using uncertainty estimation methods increases the amount of computation, producing a dilemma, as computation might go over the threshold ($10^{25}$ FLOPS) to classify the model as a systemic risk system which bears more regulatory burden.

Figures (1)

• Figure 1: Sample relationship between computation and uncertainty quality. Note how most methods plateau, and these are synthetic results representative of real-world