- The FAA approved an AI-augmented collision avoidance system years ago. The current debate over SMART misses what the regulatory architecture already settled.
By Vincent E. Bianco III
ATC Correspondent, Leeham News & Analysis
Vincent Bianco III.
May 24, 2026, © Leeham News: The Federal Aviation Administration (FAA) has already approved an artificial-intelligence-augmented collision avoidance system that takes authority away from Air Traffic Control (ATC) in critical moments.
It has been operational on commercial aircraft for years. Its decision logic was derived using machine-learning techniques. When it issues an instruction in a conflict event, the pilot follows the system, not the controller. The current hand-wringing in the popular media over whether AI belongs in air traffic management is therefore approximately three years late to a question the federal regulator has already settled. Transportation Secretary Sean Duffy is, with all due respect, late to the party.
ACAS X, the Next Gen for collision avoidance
The system is the Airborne Collision Avoidance System X—ACAS X—and its variants. ACAS Xa for transport aircraft was approved by the FAA under Technical Standard Order C-219 and ratified by the International Civil Aviation Organization (ICAO) in November 2022 as the next-generation standard for collision avoidance.
Its decision logic was developed at MIT Lincoln Laboratory using a Partially Observable Markov Decision Process framework, optimized through dynamic programming, with onboard implementations compressed using deep neural networks. When ACAS X identifies a collision threat, it issues a Resolution Advisory to the pilot.
Sean Duffy, Secretary of Transportation. Credit: FAA.
ICAO Document 9863, the regulatory document codifying ACAS operations, makes the authority structure explicit: pilots will at times maneuver contrary to ATC instructions in response to a Resolution Advisory. The system supersedes the controller because the controller, looking at a radar display with discrete update intervals, may not see the conflict in time. The system, looking at the geometry of the encounter at sub-second resolution, will.
Secretary Duffy told CBS News on April 21, that AI will not replace controllers: “Hell no, that’s not gonna happen.” Five days later, Futurism published “America Trembles as Transportation Secretary Announces Plans for Air Traffic Controllers to Lean on AI Tools,” characterizing the FAA’s scheduling-optimization software as a step toward catastrophe.
SMART predictive air traffic management
The software in question is the Strategic Management of Airspace Routing Trajectories, or SMART, a predictive air traffic management system the FAA is procuring under a three-vendor competition between Palantir Technologies, Thales SA, and Air Space Intelligence.
Vendor selection is targeted for the end of this month. First operational deployment is targeted for September 2026. Both pieces are operating on the same false premise: that AI deployment in aviation safety is a future decision the public is being asked to consider. Both pieces miss that the federal regulator has already approved an AI-augmented system that supersedes controller authority in real time, has installed it on commercial aircraft today, and has done so without political controversy because the deployment was structured correctly.
There is a dichotomy that organizes the current discourse—human ATC good, autonomous ATC bad—is a category error. The question is not whether AI belongs in air traffic management. The question is which functions of air traffic management can be appropriately augmented by automation, at what authority levels, with what failure modes, under what regulatory architecture, and, of course, the accountability question.
A four-phase regulatory architecture, articulated in our April 30 analysis, “The Regulatory Reality Behind the Autonomous ATC Gold Rush,” provides the analytical infrastructure for evaluating any deployment.
Airborne Collision Avoidance System X
ACAS X did not arrive in the National Airspace System through political reassurance theater. It arrived through a 15-year arc of academic research, regulatory architecture work, technical verification, and deployment discipline that the current Strategic Management of Airspace Routing Trajectories (SMART) debate would do well to study.
How ACAS X works. Credit: Federal Aviation Administration.
The research origin traces to MIT Lincoln Laboratory in the early 2010s, when Mykel Kochenderfer and his colleagues began applying decision-theoretic frameworks to the aircraft collision avoidance problem. The existing system, the Traffic Alert and Collision Avoidance System version II (TCAS II) used hard-coded deterministic rules to issue Resolution Advisories.
Rules worked, but the limitations were known: too many nuisance alerts, more than occasional dissonance with Air Traffic Control (ATC) instructions, and difficulty extending the logic to handle the wider range of aircraft types entering the airspace, including uncrewed aircraft and the higher traffic densities that NextGen-era operations would produce. Kochenderfer’s team framed the problem differently.
Partially Observable Markov Decision Process
Rather than encoding deterministic rules, they treated collision avoidance as a Partially Observable Markov Decision Process, a probabilistic optimization problem with uncertainty about aircraft positions, velocities, and pilot response. They optimized the resulting decision logic through dynamic programming, producing a numeric lookup table that mapped each possible state of an encounter to the optimal advisory action. Modern implementations compress the table using deep neural networks. The mathematical machinery is, in plain terms, machine learning applied to safety-critical decision-making.
The regulatory architecture work proceeded in parallel. Radio Technical Commission for Aeronautics (RTCA), the U.S. standards body, and European Organisation for Civil Aviation Equipment (EUROCAE), the European counterpart, jointly developed the Minimum Operational Performance Standards for ACAS Xa, finalized as RTCA DO-385 and EUROCAE ED-256, in September 2018.
The International Civil Aviation Organization (ICAO) incorporated ACAS Xa into Annex 10 Volume IV through Amendment 91, published in November 2022, codifying the system as an internationally recognized standard for collision avoidance. The Federal Aviation Administration (FAA) accepted ACAS Xa under Technical Standard Order C-219 and now permits four variants of ACAS II to operate in U.S. airspace alongside legacy TCAS II implementations. In March 2025, ACAS Xa was approved for operations in European airspace as well. The system is in the field today, replacing TCAS II on commercial aircraft as fleets cycle through their next major avionics refresh.
Three architectural disciplines made the ACAS X regulatory journey successful, and they are worth naming because they are exactly the disciplines that should be guiding the deployments coming next.
Last resort
First, the system is designated explicitly as a last-resort safety net. The ICAO ACAS Manual makes this designation operationally binding: ACAS is intended to work as the final layer of collision avoidance, after ATC separation services have either failed or been compromised. The system does not replace ATC. The system does not displace pilots. The system operates only when the prior layers of safety have not worked, and it operates with bounded authority: collision avoidance maneuvers only, executed by the human pilot, for the duration of the conflict event only.
Once the encounter is resolved, control returns to the prior authority structure. The architectural discipline is that the AI augmentation is bounded in scope, bounded in time, and bounded in authority.
Second, the deployment is conditional on equipage. ACAS X cannot see what is not equipped to be seen. The system depends on transponders on the threat aircraft to detect, track, and assess the encounter. Aircraft without transponders are not visible to the system, and the regulatory architecture acknowledges this limitation explicitly. The deployment did not pretend that AI augmentation could compensate for ecosystem gaps; the deployment was scoped to the conditions under which the AI augmentation could function reliably. That discipline is what allowed the regulatory authority to approve a system that supersedes ATC in conflict events without political controversy. The supersession is bounded by the equipage condition, and the equipage condition is enforced by the regulatory framework that governs which aircraft can operate in which airspace.
Verifications
Third, the verification process was substantive and protracted. RTCA DO-385 represents years of working-group analysis, simulation, flight-testing, and consensus-building among aircraft manufacturers, avionics vendors, operators, regulators, and labor representatives. The neural-network-compressed implementations of the decision tables have been the subject of extensive academic verification work.
Stanford and MIT publications on ACAS Xu safety verification through formal methods reflect the level of analytical rigor applied to the question of whether the AI-derived logic actually does what it is supposed to do under all the conditions it might encounter. The deployment did not arrive on a four-month timeline. The deployment arrived on a 15-year timeline because verification of safety-critical AI-derived decision logic requires time the verification engineers cannot compress.
During my first detail at FAA Headquarters, as a contracted Outreach Specialist in their GPS Division, WAAS/LPV approach systems were going through their final regulatory and deployment work. The pattern of how safety-critical aviation technology actually moves from research to operational deployment is one I watched in detail.
ACAS X moved through that pattern correctly. The political quiet around this very system today—nobody is asking Transportation Secretary Sean Duffy whether AI (in the form of ACAS X) should be in the cockpit, even though it already is—is the artifact of the architectural discipline applied during deployment. When the work is done correctly, the public debate does not happen. When the work is not done correctly, the public debate is the failure signal.
Part 2, SMART and the Four-Phase Framework, appears next Sunday.
Vincent E. Bianco III is a 37-year aviation safety professional and the principal of Marivin Consulting Services LLC. He spent 23 years inside the FAA Air Traffic Organization — including six years as an Operations Supervisor and Air Traffic Manager — followed by FAA Headquarters contractor tours in GPS Navigation (WAAS/LPV) and En Route procedures (RVSM), and most recently eight months as Senior Program Manager for the Leader Capability and Proficiency program at Boeing Commercial Airplanes. He serves as ATC Correspondent for Leeham News and Analysis. linkedin.com/in/vebianco3
Related
