On a Tuesday afternoon at Chicago O’Hare TRACON — the radar facility that sequences traffic into one of the world’s busiest airport complexes — a controller named Maria is managing roughly 20 aircraft simultaneously. A regional jet is running fast on final approach. A cargo flight has been issued a speed reduction that it hasn’t fully complied with. A business jet pilot has just declared minimum fuel. And a weather cell is drifting toward the outer marker, threatening to collapse the entire arrival sequence.
Maria doesn’t consult a manual. She doesn’t wait for a computer to render a recommendation. She makes four radio calls in ninety seconds, reshuffles her mental picture of the airspace, and everybody lands safely. Nobody writes a news story about it.
This scene, multiplied tens of thousands of times a day across the United States alone, is the quiet miracle of air traffic control — and the reason why, despite decades of investment, billions of dollars in modernization programs, and the most dramatic advances in computing and artificial intelligence in human history, the system still fundamentally depends on people like Maria.
What Controllers Actually Do (It’s Not What You Think)
The popular image of an air traffic controller — someone watching blips on a radar screen and telling pilots which way to turn — obscures the genuine cognitive complexity of the job. Controllers are managing four-dimensional puzzles in real time: three dimensions of space plus time, across a constantly shifting canvas of weather, aircraft performance, pilot skill, equipment status, and regulatory constraint.
The job is divided into distinct operational domains. En route controllers, working at one of the FAA’s 22 Air Route Traffic Control Centers (ARTCCs) nationwide — 20 of them in the continental United States, plus Anchorage and Honolulu — handle aircraft cruising at altitude — typically above 18,000 feet. They work large sectors, sometimes covering tens of thousands of square miles, and their primary tool is radar with data-tagged returns showing callsign, altitude, and speed. Terminal controllers at TRACON facilities manage the messy transition zone between cruising altitude and the airport surface, typically within about 50 miles of the airport and below 17,000 feet. Tower controllers manage the runway environment itself: clearances, sequencing, and the precise ballet of aircraft landing and departing on parallel or crossing runways.
Each domain requires a different cognitive style. En route control rewards systematic, long-horizon thinking — a good en route controller is building traffic pictures 20 minutes into the future. Terminal control rewards rapid pattern recognition and fine motor-verbal coordination; a skilled TRACON controller can be issuing a heading, mentally computing closure rates, and listening to a pilot readback simultaneously. Tower control is intensely visual and time-critical, more like a chess player operating under blitz conditions.
The FAA reports that U.S. controllers handle roughly 45,000 flights per day in normal operations, a figure that climbs toward 50,000 on peak summer days. The system’s safety record is extraordinary: the last fatal midair collision involving an airliner under positive radar control in U.S. airspace occurred decades ago. That record is not an accident. It is the product of a system engineered with extraordinary redundancy — and at the center of that redundancy is the human controller.
The Automation That Already Exists — and Its Limits
It would be wrong to suggest that air traffic control is a technology-free zone. Automation has been layered into ATC for decades, and its contributions are real and significant.
The most fundamental tool is radar itself — specifically, Secondary Surveillance Radar (SSR), which interrogates aircraft transponders and returns precise position, altitude, and identification data. Modern systems display this data through platforms like the FAA’s STARS (Standard Terminal Automation Replacement System) at TRACON facilities and ERAM (En Route Automation Modernization) at ARTCCs. These systems do far more than show blips: they calculate projected flight paths, flag potential conflicts, and generate alerts.
The Traffic Alert and Collision Avoidance System (TCAS), now called ACAS (Airborne Collision Avoidance System) in its international iterations, operates entirely independently of ground control. When two aircraft equipped with TCAS get too close, the system issues Resolution Advisories directly to pilots, telling them to climb or descend. Critically, TCAS can override a controller’s instruction — and pilots are trained to follow TCAS over ATC in the event of a conflict. This is the most successful autonomous safety layer in aviation, and it works because it has an extremely narrow, well-defined function.
The FAA’s NextGen program, announced in 2007 and still rolling out, introduced Automatic Dependent Surveillance-Broadcast (ADS-B), which uses GPS to provide more precise position data than radar. ADS-B has improved surveillance fidelity significantly, particularly in areas of marginal radar coverage. Performance-Based Navigation (PBN) procedures allow properly equipped aircraft to fly precise, curved arrival paths — the Optimized Profile Descents you sometimes see at major airports — that reduce fuel burn and noise while improving sequencing efficiency.
But notice what each of these technologies has in common: they are tools that enhance controller awareness and reduce workload at the margins. None of them sequence traffic. None of them negotiate with pilots about weather deviations. None of them decide, in real time, that the inbound flow needs to be stopped because a wind shift is about to flip the active runway. That remains the controller’s job.
The FAA’s own research has found that highly automated decision-support tools can improve efficiency in nominal conditions but often degrade performance during irregular operations — exactly the situations where good judgment matters most. A system optimized for the average case can become a liability in the exceptional case. And in aviation, the exceptional case is the one that kills people.
Why AI Has Struggled to Crack the Code
Given the advances in machine learning over the past decade, it seems reasonable to ask: why can’t AI do this? Deep learning systems have defeated world champions at chess and Go, mastered protein folding, and generated human-indistinguishable text. Surely sequencing aircraft is a tractable problem?
The answer reveals something important about the difference between narrow AI success stories and the messy reality of safety-critical domains.
Chess, despite its apparent complexity, has a closed rule set, a stable environment, perfect information, and discrete turns. ATC has none of these properties. The state space is continuous and four-dimensional. Information is often incomplete or ambiguous — a pilot might give an imprecise position report, or a weather system might be moving faster than the forecasters said. The “rules” are actually a layered stack of regulations, letters of agreement between facilities, local procedures, and informal operational norms accumulated over decades. And the system has no defined endpoint: it just runs, continuously, for as long as aircraft are flying.
More fundamentally, ATC involves continuous negotiation between human agents. A controller issues a clearance, a pilot reads it back (sometimes with modifications or questions), and the controller must evaluate the pilot’s competence, the aircraft’s performance capabilities, and whether the requested deviation from the clearance is acceptable. This is a social and inferential process as much as a computational one. Current large language models can approximate parts of this kind of exchange, but approximation is not acceptable in a system where a single failure can kill hundreds of people.
Aviation safety researcher Sidney Dekker, whose work on human factors has influenced operational thinking at airlines and ATC facilities globally, has long argued that safety in complex systems is not the absence of failures but the continuous presence of adaptive human expertise. “People are not the problem to control,” Dekker has written. “They are the solution to harness.” This perspective has gained significant traction within the aviation community, even as pressure to automate for efficiency and cost reasons has grown.
There is also the certification problem. Any automation introduced into the national airspace system must be certified by the FAA, a process that requires demonstrating safety performance across a comprehensive range of operational scenarios — including rare, high-consequence events that may never have occurred in practice. Training a machine learning system to handle novel, corner-case situations is precisely the thing modern AI is worst at. The FAA’s DO-178C standard for software in airborne systems and its ground-system equivalents were not written with neural networks in mind, and the agency has struggled to define an acceptable certification pathway for probabilistic AI systems.
The Staffing Crisis the Public Doesn’t Know About
While debates about automation unfold in research papers and standards committees, the U.S. ATC system is confronting a more immediate and concrete problem: it doesn’t have enough controllers.
The FAA’s Office of Inspector General reported in 2023 that the agency was more than 3,000 controllers short of its own staffing targets. The problem has multiple roots. The FAA Academy, located in Oklahoma City, is the only pipeline for new controllers, and it was disrupted severely during the COVID-19 pandemic. Certification is slow — it takes, on average, two to three years for a newly hired controller to become fully certified at a facility, and washout rates at complex facilities can be significant. Meanwhile, a large cohort of experienced controllers hired in the early Reagan era — many of them filling slots left by the mass firing of striking PATCO controllers in 1981 — have been retiring steadily.
The consequences are visible to travelers even if the cause isn’t. Ground delays, miles-in-trail restrictions, and ground stops that frustrate passengers and cost airlines money are often the direct result of facilities operating below staffing minimums. When Newark Liberty International Airport experienced significant and widely reported delays in 2025, FAA officials acknowledged that controller shortages at the Philadelphia TRACON, which handles Newark arrivals, were a contributing factor.
This staffing crisis has renewed calls for automation — not as a safety argument but as a workforce argument. If you can’t hire controllers fast enough, perhaps you can give the ones you have better tools that allow them to handle more traffic. This is a more reasonable argument than wholesale replacement, and it’s the direction most serious aviation technology research is actually moving.
Where Automation Is Making Genuine Progress
The most promising near-term automation in ATC is not in the control function itself but in the decision-support and information-management layers that surround it.
NASA’s research division has been developing tools under the umbrella of the Airspace Technology Demonstration programs that allow controllers to work larger sectors with lower workload by automating routine conflict detection and sequencing calculations. The idea is not that a computer will issue clearances, but that it will flag conflicts earlier and present pre-calculated resolution options, freeing the controller’s cognitive bandwidth for the irregular events where human judgment is irreplaceable.
Surface management systems — which track aircraft and vehicles on the airport movement area and optimize taxi routing and departure sequencing — have shown genuine efficiency gains at major airports. These systems operate in a more constrained, better-defined environment than the en route or terminal domain, which makes them more tractable for automation.
The Remote Tower concept, which has been piloted in Scandinavia and tested in the United States, is worth watching carefully. Remote Towers use high-definition cameras, audio systems, and data displays to allow a controller sitting at a central facility to provide tower services to a small airport from a distance. This is not full automation — there’s still a controller — but it changes the economics of providing ATC services to low-traffic airports where staffing a physical tower is expensive. Leesburg Executive Airport in Virginia ran the country’s first operational remote tower, judged operationally viable by the FAA in 2021, but the agency ended that pilot in 2023 after tightening certification requirements the vendor chose not to pursue. The FAA has since been testing a new remote tower system at Atlantic City International Airport, with formal design approval still pending.
Advanced Air Mobility (AAM) — the emerging category of electric vertical takeoff and landing aircraft (eVTOLs) that companies like Joby Aviation, Archer, and Wisk are developing — is forcing a harder reckoning with automation. Urban air mobility operations, if they ever scale to the volumes their proponents envision, simply cannot be managed by human controllers using today’s methods. The traffic density would overwhelm any conceivable staffed system. The FAA’s Urban Air Mobility concept of operations envisions a highly automated Urban Air Mobility Traffic Management (UTM) system for low-altitude operations, essentially a separate air traffic system running in parallel with the existing one. How that system will interact with traditional ATC, and how responsibility will be allocated when things go wrong, remain genuinely open questions.
The Road Ahead: Augmentation, Not Replacement
The serious, informed consensus in aviation — among controllers, researchers, safety experts, and thoughtful technologists — is not that AI will replace human air traffic controllers in any near-term horizon. It is that the humans doing this work need dramatically better tools, and that the current generation of AI and automation, applied carefully to well-defined subproblems, can provide some of them.
The more important near-term challenge is not automation but sustainability. Fixing the controller pipeline, building more resilient staffing at complex facilities, and retaining experienced controllers whose institutional knowledge is difficult to quantify and impossible to quickly replace — these are the pressing problems. The FAA’s recent reauthorization legislation included provisions addressing the Academy’s capacity and the pay structure for new hires, but workforce rebuilding takes years, not months.
The deeper lesson of air traffic control is one that resonates well beyond aviation. The systems we build to keep people safe in complex, dynamic, consequence-laden environments are not simply engineering problems. They are sociotechnical systems, in which the performance of the whole depends on the relationship between tools and the humans who wield them. Automation that ignores this — that treats human controllers as a problem to be automated away rather than a capability to be augmented — tends to produce systems that are efficient in normal conditions and brittle in the crises that matter.
Maria at O’Hare will retire eventually. The system needs to be ready for that. But replacing her judgment with an algorithm, in the sky above one of the world’s most complex airports, on a stormy afternoon in the middle of summer travel season, remains — for now, and for the foreseeable future — a problem that the machines haven’t solved.
For aviation professionals and serious enthusiasts who want to go deeper, the FAA’s Air Traffic Organization publishes comprehensive performance reports at its official website, and the MITRE Corporation’s Center for Advanced Aviation System Development (CAASD) releases accessible research summaries on NextGen and automation research.