AI suggests the route. Who checks the risk?

What SMART is intended to do

An AI system sees congestion developing across busy airspace. It suggests delaying one departure and changing another aircraft’s route before the problem becomes harder to manage.

The recommendation may save time and reduce workload. But before anyone acts, a more important question appears: what does the system know—and what might it have missed? This is the safety issue behind the FAA’s new SMART programme.

SMART stands for Strategic Management of Airspace, Routes, and Trajectories.

In testimony dated 15 September 2026, the FAA said the cloud-based system would use information from its Flow Management Data and Services platform. That information includes weather, airline schedules, aircraft positions and predicted trajectories. Source ↗

The FAA says SMART will use AI to identify congestion and conflicts earlier and help adjust departure times, routes and timing at points along a flight. The supplier describes its technology as recommending traffic-flow strategies to decision-makers. Source ↗

This is strategic traffic-flow management. The available sources do not establish that AI is replacing controllers or independently taking over aircraft-separation duties.

Reporting on 18 September said an initial Washington, D.C.-area introduction could occur as soon as Monday, 21 September 2026. “As soon as” describes a plan, not confirmation that the system has entered operational service. Deployment status should be checked again after that date. Source ↗

A sensible recommendation can still be unsafe

Consider a simple example.

The system recommends a new route around thunderstorms. It has considered aircraft positions, forecast weather and traffic demand. The route appears efficient.

But one weather feed is late. A temporary airspace restriction has not reached the system. Another aircraft’s trajectory has changed since the recommendation was calculated.

The AI may be working exactly as designed and still produce an unsafe or unsuitable answer because its picture of the world is incomplete.

This is an illustrative scenario, not an allegation about SMART.

The safety task is to ensure that people understand what information supports a recommendation, how current it is and what uncertainty remains.

The human role must be more than approval

Calling a system “decision support” does not automatically make its use safe.

If a recommendation arrives during high workload, looks authoritative and gives little explanation, the easiest action may be to accept it. That creates the appearance of human oversight without meaningful challenge.

A controller or traffic manager needs to know why the recommendation was made, which constraints were considered and whether any important input is missing or degraded. The interface should make disagreement possible and visible.

Training must also include occasions when the correct response is to reject the AI’s advice. Otherwise, operational experience may teach people that the machine is usually right—until the unusual day when it is not.

The cybersecurity connection

Air traffic decision support depends on trusted information.

A bad actor would not necessarily need to shut the system down. Altered weather information, a misleading trajectory or corrupted operational data could be more difficult to detect if the result still looks believable.

That does not mean SMART has been hacked or is known to be vulnerable. It means data integrity, source authentication, access control, monitoring and recovery deserve the same attention as the model’s predictive performance.

Availability matters too. If the system becomes unavailable, can the operation continue safely? Do users retain the skills and information needed to work without it? Has the fallback been exercised under realistic workload?

Management of change cannot stop at installation

The first operational version will not be the last.

Data feeds will change. Models may be recalibrated. Interfaces and recommendation logic will be updated. Each material change can affect what the system detects, what it misses and how people respond.

The organisation must therefore define who owns the operational change; who performs hazard identification and safety risk assessment; how the new version is validated before use; who accepts the residual risk at the appropriate authority level; what will be monitored after implementation; and what conditions require restriction, rollback or withdrawal.

The safety department may support this work, but operational ownership cannot simply be transferred to safety or to the technology supplier.

From prediction to safety intelligence

The promise of systems such as SMART is earlier warning: seeing a developing problem while useful options remain available.

That is closely aligned with the direction of in-time safety management. But speed alone is not safety intelligence. The information must be trustworthy, the recommendation must be understood, and someone with authority must be able to act.

The key question is not whether AI belongs in air traffic management. It is: When AI suggests the route, can the responsible person understand the risk, challenge the advice and make a safer decision in time?