DIALOG open day and workshops: exploring the core questions on human-AI collaboration in ATM

As artificial intelligence becomes increasingly capable of supporting complex decision-making, one of the key questions facing air traffic management (ATM) is not whether AI can assist controllers, but how humans and AI can work together safely, effectively and with clear accountability. The DIALOG project is exploring this challenge by investigating how AI-powered digital assistants can support air traffic controllers during demanding operational situations, while ensuring that humans remain firmly in control of safety-critical decisions.

At the heart of the research are fundamental questions about human-AI teaming: Which controller tasks can be supported or automated without compromising situational awareness? How can AI systems understand operational intent rather than simply process data? How can a system adapt its support to a controller's cognitive workload? And how can trust in automation be built without creating overreliance?

These questions were the focus of the DIALOG Open Day and workshops, held on 9 and 10 June 2026 in Rome, Italy. The event brought together the project consortium, led by SINTEF (Project Coordinator) and comprising the Direction des Services de la Navigation Aérienne (DSNA, French Air Navigation Service Provider, Technical Manager), Deep Blue, the French Aerospace Lab (ONERA), the Deutsche Flugsicherung GmbH (DFS, German Air Navigation Service Provider), and Radboud University, alongside external advisory board members and experts from the German Aerospace Center (DLR), Airudit, commercial airline pilots and the Single European Sky ATM Research Joint Undertaking (SESAR JU). The primary objective of this gathering was to present the project achievements and engage in collaborative workshops to evaluate the progress of human-AI collaboration in ATM.

Day 1: Putting the teamwork assistant to the test in a simulated sky

The first day of the event was heavily focused on knowledge sharing and dissemination, while still gathering relevant insights for the overall benefit and impact evaluation of the project form the stakeholder perspective, providing attendees with a deep dive into the broader research scope of human-AI collaboration in ATM. The consortium presented a live demonstration of the core capabilities of the intent-inferring service (IIS) and the teamwork assistant (TA), configured to simulate a high-complexity en-route environment based on a sector of the Marseille Air Control Centre (ACC). While a full Teamwork Assistant operationally relies on real-time ATCO workload assessments, this demonstration focused on showcasing how the system behaves under different simulated workload modes, providing a streamlined yet highly effective presentation for knowledge-sharing purposes.

DIALOG low-fidelity prototype showing tactical conflict detection (TCT) and task load monitoring.

The system operates continuously in the background using two core components: an intent-inferring service, which interprets pilot voice and Controller-Pilot Data Link Communications (CPDLC) requests, and a teammate awareness service, which monitors the controller's psychophysiological state. During the technical sessions, the partners explained how a dedicated survey was administered to evaluate the cross-national generalisability of the hierarchical task Analysis (HTA) used in the project. This analysis serves as the baseline for determining which tactical subtasks can be successfully automated, directly supporting the project's goal of targeting EASA Level 2B to achieve an effective and adaptive human-AI teaming.

A major point of interest was how the project monitored a controller's cognitive state seamlessly within an operational environment. The team opted for a non-intrusive suite of eye-tracking glasses and hand electrodes, ensuring maximum comfort for the operator. This combination of physiological data allowed the system's algorithm to successfully predict high and low workload states with a 74% accuracy rate during testing.

Day 2: What an unexpected simulation glitch revealed about human authority

The second day shifted the focus towards the operational validation results, reviewing qualitative and quantitative data gathered from 20 simulation runs involving 10 licensed air traffic control officers (ATCOs). While the self-report measures showed that controllers found the interface intuitive, describing the assistant as a supportive "second pair of eyes", the workshop discussions highlighted a fascinating human-factors lesson that occurred due to an unexpectedsoftware simulation error.

During one of the evaluation runs, a technical glitch caused the AI to suggest an incorrect flight Flight Level change advisory . Rather than disrupting the session, this moment provided the advisory board with concrete proof of the project's safety-by-design concept. The controller in the hot seat immediately identified the suboptimal suggestion, overrode the system and resolved the aircraft conflict manually.

This scenario confirmed that the participating controllers did not suffer from automation bias (the tendency to trust automated suggestions blindly). They maintained full situational awareness and treated the TA as an aid rather than an absolute authority, proving that preserving ultimate human command acts as the most effective final safety barrier in advanced aviation systems.

Addressing the core research questions of human-AI collaboration in ATM

The concluding workshops allowed the consortium and external experts to address several critical research questions that offer practical utility for the wider aviation community, moving past project rhetoric to focus on operational realities:

• Does AI automatically increase airspace capacity? The short answer is no. The DIALOG validation results demonstrated that the true value of the tool lies in mitigating efficiency degradation rather than expanding nominal capacity. During severe weather conditions, when sector capacity typically drops by up to 30%, the digital assistant acts as a cognitive buffer, protecting controllers from short-term memory overload.
• Is high word recognition enough for voice systems? It is not. While a 97% word-recognition rate sounds impressive, the critical metric for aviation safety is intent recognition. The system must be robust enough to accurately distinguish whether a transmission is a clearance, a request or a read-back, even when dealing with non-standard phraseology, diverse accents or overlapping communications.
• How fast should the interface update? Managing data refresh rates requires a careful balance. The DIALOG trials identified an optimal data refresh cycle of approximately 10 seconds to keep the interface reliable without creating cognitive clutter on the radar screen.

The flight path towards 2035

Currently sitting at technology readiness level 2 (TRL 2), the DIALOG prototype represents a foundational milestone rather than a finalised operational tool. The long-term roadmap targets potential operational deployment around the year 2035.

Reaching that goal will require extensive certification processes, deeper integration with actual flight plan processing systems and the development of specialised training programmes focused on human-automation interaction. The ultimate takeaway from the two days in Rome is that the future of flight safety relies on an incremental partnership where technology manages the data load, but the human operator always retains final authority.

More about DIALOG