In a landscape increasingly dominated by artificial intelligence (AI), concerns about its safety and reliability loom large. Michael O’Connor, a former Air Force flight test engineer and current U.S. Space Force Fellow, steps forward with an insightful perspective, drawing crucial parallels between the world of aviation and the challenges of AI. His article, “The Right Stuff for AI: Hard-won Safety Lessons from the world of Flight Testing,” delves into three pivotal lessons gleaned from the rigorous history of flight testing.
O’Connor meticulously outlines strategies that bridge the gap between aviation’s well-established safety protocols and the burgeoning realm of AI, emphasizing the pressing need for quantified standards, realistic testing environments, and an innovative approach termed “envelope expansion.” Through this astute comparison, he proposes a roadmap, rooted in aviation’s hard-earned wisdom, to navigate the complexities and uncertainties of AI’s development and deployment safely.
- Establish Design Standards for Humans and Technology: Drawing parallels with aviation standards, O’Connor emphasizes the need for quantified recommendations and performance standards for AI systems. These standards should consider not only the technology itself but also the human factor, including biases and blind spots that may affect the AI’s training data and real-world applications.
- Test Systems in Realistic Conditions: O’Connor suggests that AI safety testing should focus on realistic tasks rather than ideal conditions. Similar to the systems testing in aviation, where day-to-day evaluation ensures the safety and intended functionality of aircraft, testing AI systems should address interactions between AI systems, from training data to deployment, and consider potential methods of abuse.
- Systematically Expand the “Envelope” of Safe Performance: O’Connor introduces the concept of “envelope expansion,” a process used in aviation to push the limits of aircraft performance gradually. He suggests applying this approach to AI by understanding the complex and multi-dimensional nature of an AI’s performance envelope. This involves testing the system’s behavior in various conditions, acknowledging the unpredictability and non-deterministic nature of modern neural network AI.
The op-ed concludes by emphasising the importance of learning from the military flight test world’s history, applying hard-earned knowledge, and adapting the mindset to benefit the safe development and deployment of AI systems in the future. O’Connor acknowledges the challenges of AI complexity and the evolving nature of both AI and its operational environment, urging regulators and developers to adopt a testing-centric approach and consider the human factor in designing standards for AI.
Michael O’Connor is a U.S. Space Force Fellow at the Center for Security and Emerging Technology (CSET), Georgetown University. Prior to joining CSET, he was a space program test lead at Los Angeles Air Force Base, Calif. He previously served as an evaluator flight test engineer supporting remotely piloted aircraft testing at the Air Force Test Center at Edwards Air Force Base, Calif. The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of the Department of the Air Force, the Department of Defense, or the U.S. Government.
Photo credit: An Army experimental test pilot checks out cockpit electronics. (Army Redstone Test Center photo by Collin Magonigal)
