Former Air Traffic Controller Terrified New AI System Will Hallucinate and Cause Catastrophe. The United States federal government is pressing forward with its contentious initiative to integrate an Artificial Intelligence (AI) air traffic control system, dubbed the Strategic Management of Airspace, Routes and Trajectories (SMART), across several of the nation’s busiest airports, raising profound safety concerns among aviation professionals. At the forefront of this apprehension is Todd Sheridan Yeary, a veteran air traffic controller with 13 years of experience, who recently voiced his alarm on *News Nation*, highlighting the unprecedented risks associated with deploying a system that remains largely opaque to the very community it is meant to serve. Yeary’s critique centers on fundamental questions about SMART’s intended functionality, the dynamic nature of air traffic control, and the inherent susceptibility of AI to “hallucinations,” which in this high-stakes environment could lead to catastrophic outcomes. The rollout plan, particularly its “soft launch” in the highly congested DC metro area, has been met with staunch opposition, intensifying a broader debate about the readiness of both the technology and the country’s already strained air traffic infrastructure to embrace such a transformative, yet unproven, shift.

The FAA’s motivation for introducing SMART stems from a desire to modernize an aging air traffic control system plagued by outdated technology, understaffing, and increasing flight volumes. Proponents argue that AI can optimize flight paths, reduce delays, and enhance safety by predicting potential conflicts more efficiently than human controllers. SMART is envisioned as a sophisticated tool that could process vast amounts of data—including weather patterns, flight schedules, aircraft performance, and airspace restrictions—to provide real-time recommendations or even autonomous instructions. The hope is that this automation will alleviate the immense pressure on human controllers, who frequently work long shifts, often six days a week, due to persistent staffing shortages. The National Air Traffic Controllers Association (NATCA) has consistently highlighted these staffing deficiencies, emphasizing that the current workforce is stretched thin, leading to increased stress and potential for human error. In this context, AI is presented as a panacea, a technological leap that could address systemic inefficiencies and fortify the system against future failures. However, Yeary and others argue that this pursuit of efficiency might inadvertently introduce new, potentially more perilous, vulnerabilities.

One of Yeary’s primary concerns revolves around the ambiguity of SMART’s role: will it serve as a predictive aid, offering suggestions for human controllers to consider, or will it be prescriptive, actively making critical decisions that bypass human judgment? “The first that comes to mind is, what is the tool going to be used for?” Yeary questioned, emphasizing that the answer is paramount to understanding the system’s true impact on aviation safety. If SMART is merely a predictive tool, assisting controllers with data analysis and conflict prediction, it could be a valuable asset, augmenting human capabilities. However, if it moves into a prescriptive role, dictating resolutions to complex air traffic scenarios, the implications are far more profound and concerning. Air traffic control is inherently a “dynamic environment,” as Yeary pointed out. Unlike many automated systems that operate in controlled, predictable settings, the skies are in constant flux. Weather conditions can change in an instant, mechanical failures can necessitate emergency landings, unexpected airspace incursions can occur, and pilots may deviate from planned routes for various reasons. Human controllers possess an unparalleled ability to adapt, improvise, and make nuanced judgments based on incomplete information, experience, and intuition—qualities that current AI systems struggle to replicate.

Yeary’s most significant apprehension, encapsulated in a single word, is “hallucination.” In the realm of AI, particularly large language models, hallucination refers to the generation of plausible-sounding but factually incorrect or entirely fabricated information. While often discussed in the context of text generation, the principle applies to any AI system that processes and synthesizes data. In an air traffic control context, an AI hallucination could manifest in terrifying ways: imagining “ghost” aircraft on radar screens, misinterpreting flight data to suggest non-existent conflicts, proposing impossible maneuvers, or incorrectly identifying aircraft. The sources of AI information are critical, Yeary noted, asking, “what are going to be the sources from which AI is getting this information?” If the system relies solely on scheduled airline data, it would be dangerously incomplete, failing to account for real-time deviations, general aviation, military flights, or emergencies. The potential for such a system to generate erroneous commands or data that controllers implicitly trust, or are even compelled to follow, presents an existential threat to aviation safety. This necessitates an absolute “override” feature, Yeary stressed, ensuring that “that override has to come from the controller on duty.” The concept of a “human-in-the-loop” is vital here, ensuring that human operators retain ultimate authority and the ability to intervene and correct AI errors, preventing them from escalating into disasters. Without such a robust and intuitive override, controllers risk becoming mere monitors, potentially losing the critical cognitive skills honed by active decision-making, or being unable to react swiftly enough when an AI system errs.

The planned “soft launch” of SMART in the DC metro area, one of the nation’s most complex and heavily trafficked airspaces, further amplifies these concerns. Yeary minced no words, stating that “anything other than a simulator to roll this out is a problem.” Introducing an experimental AI system directly into live operations in such a critical region, which encompasses three major airports (Dulles, Reagan National, and Baltimore-Washington International), alongside extensive military and restricted airspace, is perceived by many as an unacceptable gamble. Thorough simulator testing, allowing for real-life-like scenarios to be run, rerun, and analyzed in a controlled environment, is the established protocol for introducing new aviation technologies. This allows for the identification and rectification of bugs, vulnerabilities, and unforeseen interactions before human lives are put at risk. Yeary articulated this sentiment powerfully: “Until we really have a chance to really kind of test on real-life-like scenarios, it is very difficult to say we’re going to rely on this information in a real time exercise… I think there’s a lot more that we don’t know, and I don’t want to try to find out in the middle of a scenario that is going on in real-time at an airport.” The history of aviation safety is built upon a foundation of meticulous testing and incremental implementation, a principle that critics argue is being dangerously sidestepped with the SMART rollout.

The context surrounding the SMART rollout is further complicated by recent events that underscore the fragility of the existing air traffic control infrastructure. Just prior to these discussions, a “major aviation communications failure” near Philadelphia grounded hundreds of flights across the East Coast, causing hours of delays and widespread disruption. This incident, reportedly caused by a simultaneous outage of a core telecommunications line and its redundancy, highlighted critical vulnerabilities in the nation’s air traffic control system. Such failures, whether due to aging physical infrastructure or human error, raise pertinent questions about the system’s resilience and its capacity to absorb the complexities of a highly experimental AI system. If a simple communications line failure can cripple air travel, what unforeseen cascading effects could a “hallucinating” AI system introduce? The Philadelphia incident served as a stark reminder that even seemingly minor technical glitches can have massive repercussions in the interconnected world of aviation. The burden on overworked flight control personnel, already grappling with these existing vulnerabilities, would only intensify if they are simultaneously tasked with overseeing an AI system whose reliability and behavior are not yet fully understood or proven.

Beyond the immediate operational concerns, the deployment of AI in such a critical domain raises profound ethical and regulatory questions. Who bears ultimate responsibility if an AI system makes a decision that leads to an accident? Is it the software developer, the FAA for approving the system, or the human controller who might have been unable to override a faulty AI command in time? Current regulatory frameworks for AI in critical infrastructure are still evolving, struggling to keep pace with the rapid advancements in artificial intelligence. Public trust, which is paramount in aviation, could be severely eroded by even a single incident attributed to AI error. The integration of AI into air traffic control represents a significant step towards a future where automation plays an increasingly dominant role in managing complex systems. However, as Yeary and other experts warn, this progress must not come at the expense of safety. The allure of efficiency and technological advancement must be tempered by rigorous testing, transparent development, and an unwavering commitment to human oversight. The stakes are simply too high for anything less than absolute certainty in a domain where lives literally hang in the balance.