AI air traffic control, as it exists today, is decision support. Models predict routes, delays, congestion and arrival spacing, and they propose actions. A human decides whether to act on them. The FAA, the US Department of Transportation and IFATCA, the controllers’ international federation, all state this position. The harder questions are about aviation AI engineering: what data feeds these tools, how you test them before they reach an operations room, how you notice when they start to degrade, and who answers for a recommendation that turns out to be wrong.
Does AI replace human air traffic controllers?
No. US transport leaders and IFATCA explicitly reject AI as a replacement for controllers. US Transportation Secretary Sean Duffy said “AI is a tool, but we do not replace humans in how we manage the airspace”. The FAA’s own announcement of its SMART platform states that SMART “cannot be used to replace the critical role air traffic controllers play or take over control of an aircraft”, and adds that local leadership can accept or pass on its scheduling suggestions. When NPR reported on the launch, FAA and DOT leaders again called the system a tool for reducing congestion and conflicts before they happen, and Riley, quoted by NPR, said human controllers will still be in charge of keeping aircraft separated.
The European regulator frames AI by levels of human involvement. EASA’s AI Roadmap defines three general levels: human assistance (Level 1), human-AI teaming (Level 2) and advanced automation (Level 3). Its first two concept papers cover safety-related AI for end-user support, meaning pilots and ATCOs, and the roadmap reports that most industry players envisage Level 1 assistance scenarios first. The roadmap places conflict detection and resolution by “virtual co-controllers” at Level 2, around 2035, and says plainly that this date is a prediction rather than a target. Automation with human supervision (Level 3A), and eventually without it (Level 3B), is expected gradually between 2035 and 2050.
Researchers use a similar decision-support framing. The authors of a UK en route Digital Twin study define an “AI agent” as any AI-based decision-support technology for tactical ATC, whatever its intended level of autonomy. They also say certified operational deployment of such agents remains a significant long-term challenge. A review of human-AI collaboration describes the trajectory: as automation increases, the operator’s role moves from active control to supervision and later to strategic oversight.
Which AI tools are already in operational use in air traffic management?
The deployed uses sit mostly in flow management and planning, not in moment-to-moment separation. EASA reports that AI-enabled assistants are already in operation supporting ATCOs and flow management positions, for example by improving 4D trajectory predictions so that hotspots are detected earlier. The table below separates tools that are in operational use from research results. That distinction matters more than any accuracy figure attached to either.
Where AI is in use in ATM, and what it does
| Tool or study | Organisation | Task | Status | Reported result |
|---|---|---|---|---|
| Traffic Prediction Improvements (TPI) | EUROCONTROL | Predicts each flight’s most likely route using a deep neural network trained on historical trajectories | Integrated into the operational flow and capacity management system | For flights the network had not seen, 65% of predicted routes stayed within 6 NM of the flown trajectory, against 10% for the existing prediction |
| FADE (Forecast of ATFM Delay Evolution) | EUROCONTROL | Predicts delay trend and predicted delay for flights under ATFM regulations | Deployed in the Network Manager’s operational systems | Thirteen airlines tested the prototype before deployment |
| COAST | EUROCONTROL | Calibrates time-based separation on final approach using ML trained on local traffic and MET data | Available to ANSPs for their separation delivery tools | Runway throughput up to 6% higher at major airports |
| SMART | FAA / USDOT | Merges data streams to predict congestion and recommend schedule changes | Rollout started at three airports serving the Washington, D.C. region | Recommendations go to facilities for local consideration |
| Traffic management initiative prediction | NASA / FAA (SciTech 2024) | Predicts whether a ground stop or ground delay programme will be needed at Newark | Research | LSTM F1 scores of 0.97 (no initiative), 0.83 (GDP) and 0.68 (ground stop) |
| Long-range traffic forecasting | EUROCONTROL STATFOR | ML forecasts on seven North Atlantic flows | Methods being re-engineered into forecasting | Median absolute error cut by 8.5% to 71% against the STATFOR baseline |
TPI shows a clear integration pattern. EUROCONTROL presents it as producing “what-if” trajectories that add to existing capabilities rather than replacing them. If the new logic causes problems, the flow management position can switch back to the original trajectories. Because the core flight data processing system stays unchanged, the novel technique can run inside a system ranked safety-related. Our view is that this fallback is the most transferable lesson in the table.
On the US side, Duffy described SMART as looking up to 45 days ahead and moving flights a few minutes earlier or later to avoid predicted delays. The USDOT fact sheet says it sorts out landing times hours in advance, so the system sets the order of arrivals instead of leaving controllers to decide it in real time. For a wider view of how these tools affect safety outcomes, our piece on how AI can improve aviation safety covers the territory beyond flow management.
The data an AI air traffic control tool combines
The tools above combine several data feeds. An ICAO paper describes controller decision support that combines weather information, flight plans and surveillance data to produce crossing and risk forecasts. SMART brings together weather patterns, flight paths, arrival and departure statistics and controller staffing metrics on one platform, and FAA leaders say it considers all planned airline flights along with weather and wind. TPI was trained on flight and airspace data, including radar-observed trajectories, for more than 328,600 flights between December 2015 and December 2017.
The NASA/FAA study shows how its inputs were selected. The authors trained on three years of weather, demand, capacity and initiative history from Newark. They tried convective forecasts and terminal aerodrome forecasts, and found METARs alone sufficient for current-hour predictions. A route-choice study using EUROCONTROL data used route length, navigation charges and the rate of regulated flights (as a proxy for congestion) as explanatory variables. Its authors say wind, airport configuration and take-off delay could improve the model.
TPI’s own documentation points out that its offline accuracy assumes correct knowledge of the predictors, while in live operations military area activation schedules change and flight plans are refiled.
What models are used for conflict prediction in air traffic control?
AI-based conflict detection and resolution is still being investigated, including with reinforcement learning. EASA notes that SESAR 3 projects, industry and ANSPs are investigating reinforcement learning for conflict detection and resolution. Here are the approaches named in the sources:
- Multi-agent reinforcement learning. A deep ensemble MARL framework suggests real-time speed adjustments to help a controller avoid congestion and near-misses. It was trained and evaluated on EUROCONTROL’s open-source ATM simulator.
- Off-policy RL with attention networks. For advanced air mobility corridors, a decentralised Discrete Soft Actor-Critic agent uses attention to handle a variable number of nearby aircraft.
- Rules, search and optimisation agents alongside RL. Five mature agents developed on the UK Digital Twin cover this whole range, up to black-box methods such as reinforcement learning.
- Graph neural networks for workload. A GNN prototyped on the same Digital Twin predicts near-term controller task demand in a sector and flags which aircraft are relevant to each other.
- Sequence models for flow constraints. For upstream problems such as predicting initiatives, LSTMs beat Random Forest and XGBoost because they learn how controllers revise programmes over time.
SESAR projects provided evidence of technical and operational feasibility for AI conflict detection and resolution at the controller working position. This includes configurations where the system implements resolutions while the human monitors and can intervene at any time. Feasibility is not certification, which remains a significant long-term challenge. The authors of the AAM separation study propose, as a near-term route, using their algorithm as a recommender that gives the controller a suggested manoeuvre under human supervisory control.
Large language models deserve a separate warning. LLMs acting as controllers reproduce the register of ATC speech well, yet fail to issue the right instruction “roughly as often as not”. The best configuration reached a ROUGE-L of 0.31 against a human controller, which the authors call far below the reliability safety-critical use needs. The authors also point to the specific danger: a system that sounds right while acting wrong is an operational hazard, because fluency is the cue a listener uses to decide how closely to check. They attribute the gap to missing operational state (traffic, trajectories, airspace geometry, weather) that no prompt can supply. Our reading is that an LLM belongs outside the control loop, in training or transcription support, until it is grounded in that live state.
How does AI handle drones and unmanned aircraft near airports?
We found no source describing a deployed AI system managing drones near airports. What the sources do show is the regulatory and research framing. The EU AI Act’s requirements are expected to flow into future aviation rules for airworthiness, ATM/ANS and unmanned aircraft. EASA uses drones to illustrate where its highest automation levels divide: the difference between Levels 3A and 3B could be whether human oversight is lost, for example when the command and control link to a drone fails.
In research, separation for dense advanced air mobility traffic is being studied with decentralised reinforcement learning that uses speed and vertical manoeuvres. Those authors say real-world deployment would need further steps, including latency, weather and live traffic models built from sources such as the FAA’s System-Wide Information Management system. They suggest following the multi-organisation approval route that ACAS X went through with RTCA, EUROCONTROL and the FAA. A drone-specific airport deployment would need the same evidence as any other system.
Testing an ATM model before it reaches live operations
Before any ATM model goes live, it needs controllers in the loop, realistic uncertainty and validation on data it has never seen. The Digital Twin authors argue that formal methods and automated scenario generation are not enough on their own, because they cannot capture the situational judgement trained controllers make. They conclude that human-in-the-loop assessment by qualified controllers is essential. They also argue that deterministic simulation gives an incomplete basis for evaluation, since real operations carry uncertainty from schedule deviations, convective weather and unexpected events.
That study shows what the evidence looks like in practice. Its probabilistic trajectory predictor achieved mean absolute errors 27% and 44% lower than the BADA baseline for descending B738 airspeed and rate of climb or descent. Over two rounds, a rules-based agent improved from “Unsatisfactory” in all four competencies to “Satisfactory” in three. A fast-time twin of the whole ATC task can screen procedures and routing strategies before anyone commits to expensive human-in-the-loop trials.
Pre-live checklist for an ATM decision-support model
- Held-out period, not held-out rows. Test on a later period in which route options and airlines may have changed, as the route-choice study did with its separate testing year. That study’s validation error of around 10% is the kind of number worth reporting honestly.
- Consistency across the network. EUROCONTROL’s forecasters found that methods which worked on small examples improved predictions in only 51% of States and made the rest worse, with no way to tell in advance which would be which. Their rule is that without a proven, consistent improvement they cannot adopt a method.
- Outlier screening. Even their long-established ARIMA system produces occasional absurd forecasts (“monsters”), and humans still clean these up.
- Controller validation. IFATCA says controllers’ role shall include setting user requirements, defining training needs, taking part in risk assessment and validating the system. It also says AI introduction shall be tested for safety, security, robustness and reliability, with backup procedures for disruption.
- Breadth of verification. For learning-based separation, the AAM study names adaptive stress testing, large-scale Monte Carlo evaluation, real-world testing and formal methods.
- User trial before deployment. FADE was deployed in operations after airlines gave positive feedback that its predictions helped them plan.
An aggregate improvement can hide regions or flows where the model is worse, and the operations room will see those regions.
Drift in an AI traffic-flow prediction model
Air traffic behaviour changes over time, so models need regular updates. Researchers analysing flight delay prediction describe this changing behaviour as concept drift. They found that the right drift-handling strategy depends on the scale (one airport or the whole system) and on the model used. EUROCONTROL says COAST needs regular model updates to cover the full fleet, new aircraft types and operational changes. TPI keeps up with change through iterative training on newly recorded data.
One traffic-flow forecasting study compares each client’s recent data distribution against a KL-divergence threshold, so only clients whose traffic has shifted request a model update. Loosening the threshold lowered participation and raised RMSE, and the authors chose a value that balanced accuracy against compute. EASA treats real-time learning in operations as incompatible with current certification and recommends a strong limitation for safety-related systems. Its continuous safety management process adds operational-phase requirements that may trigger retraining, and its assurance building block includes data recording to monitor safety continuously. In practice that means monitoring runs in operations while retraining happens outside them.
Who owns the decision when an AI recommendation is wrong?
The ATCO keeps the decision, and accountability for every party has to be clearly delineated. IFATCA’s policy states that AI and ML systems “should only be implemented as decision support systems and shall not replace the decision of the ATCO”. It adds that controllers must be able to manage their area safely if the tools fail. It also says a controller should be free not to use an AI tool while liability remains unresolved. Its policy is that controller responsibility should decrease in proportion to the level of automation and their ability to intervene. IFATCA notes that current ATM liability models do not yet address the interaction between automated systems and humans.
ICAO adds the human-factors side. Automation can create over-reliance and lapses of attention in emergencies, and training in human-machine interaction is needed to counter this. ICAO also calls for regulatory frameworks that define responsibilities in human-machine interaction. IFATCA asks that refresher training include how to recognise and validate AI outputs and how to correct their mistakes. SESAR research flags the de-skilling risk for controllers who move into a monitoring role.
Transparency alone did not raise controllers’ acceptance in SESAR research. EASA notes that an ML model with fixed weights can still give unpredictable outputs for new inputs, because the output depends on how the input relates to the training data. In the SESAR project Artimation, controllers under pressure did not find time to review AI explanations and preferred a simple display of the output. Experienced controllers were confident enough to reject AI solutions that did not match their experience. The projects concluded that trust is built in certification and training, not through explanations in the tactical phase. For the wider problem of opaque models, see our piece on AI’s black box.
Our view is that the escalation path is part of the system and needs specifying and testing like any other component. The flow manager who rejects a SMART-style recommendation, the engineer who sees a drift alert and the supervisor who decides whether to fall back to legacy trajectories each need a defined trigger and a defined fallback. TPI’s switch back to unmodified trajectories is a working example of such a fallback.
