Adaptive Design Assurance Levels for AI-Based Avionics Systems A Framework for Certification and Continuous Assurance

Abstract

Artificial intelligence (AI) is increasingly proposed for perception, prediction, decision support, and automation in civil aviation. However, machine-learning components are data-dependent, statistically characterized, and vulnerable to uncertainty and distribution shift, while conventional Design Assurance Levels (DALs) define development rigor rather than real-time system health. This paper proposes an Adaptive AI Assurance Framework that preserves the baseline DAL assigned through aircraft-level safety assessment and adapts only the operational authority of the AI component. The framework combines AI-specific development evidence, a frozen deployed model, independent runtime monitoring, non-compensatory safety constraints, a deterministic authority manager, and a verified fallback function. A normalized assurance margin supports transitions among nominal, degraded, and fallback states, but is explicitly not interpreted as a probability of safety, an equivalent DAL, or a standalone means of compliance. The approach is mapped to DO-178C, ARP4754B, ARP4761A, EASA and FAA AI-assurance initiatives, and emerging EUROCAE/SAE standardization. A simulation protocol for an AI-assisted runway-alignment and landing-guidance function illustrates how uncertainty, sensor quality, operational-design-domain compliance, distribution shift, and cross-channel disagreement can govern AI authority. The paper argues that certification-oriented use of aviation AI should be based on a fixed development-assurance baseline combined with conditional operational authorization, verified protection mechanisms, and controlled post-deployment change assurance.