Certifying the Uncertifiable: Aviation Safety Regulation for Learning Artificial Intelligence

Artificial intelligence is challenging one of aviation’s most fundamental certification assumptions: that an approved system will behave tomorrow exactly as it did when it was certified. This paper examines the evolution of airborne automation into machine-learning-based systems and explores the regulatory, technical, operational, military, legal, and human-factors implications of certifying artificial intelligence whose behavior may be derived from data or modified through continued learning. Particular attention is given to emerging assurance concepts such as operational design domains, learning assurance, explainability, bounded adaptation, runtime monitoring, and lifecycle-based certification. Current developments involving Airbus, the Federal Aviation Administration, the European Union Aviation Safety Agency, the United States military, and DARPA demonstrate that increasingly autonomous and adaptive systems are already moving from theory into operational experimentation. The paper also considers the civil and legal ramifications of distributing decision-making authority among pilots, operators, manufacturers, software developers, and AI systems whose internal reasoning may not be immediately transparent. It concludes that the future of aviation certification may depend less on proving that every possible AI behavior is predetermined and more on demonstrating that adaptive behavior cannot escape a rigorously defined safety envelope. In this emerging framework, the central challenge is not whether aircraft can learn, but whether aviation can permit learning without surrendering accountability, predictability, and control.