Dimitrios Ziakkas, Ibrahim Sarikaya, Debra Henneberry (2026). Training Challenges in Human -AI Teaming in Aviation. In: Advances in Human Factors of Transportation. AHFE Open Access, vol. 224, pp. 36–44. https://doi.org/10.54941/ahfe1007831
Human–AI teaming is rapidly emerging as a defining paradigm in next-generation aviation operations, reshaping pilot roles, altering cockpit task distribution, and challenging established assumptions regarding expertise, decision-making, and training. As artificial intelligence systems evolve from deterministic support tools into adaptive, autonomous teammates capable of perception, prediction, and intent-driven action, the aviation training ecosystem faces a suite of unprecedented challenges. These challenges extend beyond purely technical skills and encompass deeper questions of trust calibration, cognitive adaptation, workload redistribution, ethical responsibility, and sustained human performance. This paper examines the central training challenges associated with preparing pilots, instructors, and organisational systems for effective human–AI teaming across current and expected future aviation environments.First, the paper analyses the shifting cognitive and operational landscape introduced by AI-enabled systems, including adaptive automation, predictive analytics, natural-language interfaces, and mixed-initiative control architectures. Whilst these technologies promise enhanced situational awareness, reduced workload, and strengthened predictive safety nets, they simultaneously introduce risks such as automation complacency, algorithmic over-reliance, erosion of manual competencies, and emergent forms of mode confusion. Training organisations must therefore rethink curriculum design to cultivate appropriate levels of trust in AI agents while strengthening pilots’ abilities to monitor, interrogate, and, where necessary, override AI behaviour during uncertainty or system drift. Traditional training paradigms based on linear automation logic are insufficient to address the probabilistic and at times opaque behaviour of modern AI systems.Second, the paper explores the pedagogical complexities inherent in developing joint human–AI decision-making skills. Effective teaming requires robust communication transparency, alignment of mental models, and the formation of shared situational awareness between human operators and algorithmic agents. Yet many AI systems operate as “opaque teammates,” offering outputs without interpretive depth or explainable reasoning. Training must therefore introduce strategies for evaluating machine-generated recommendations, identifying algorithmic bias, integrating AI insights with experiential human judgement, and managing discrepancies between human and AI interpretations. Scenario-based training, explainable AI (XAI) tools, and structured failure-mode exploration are presented as essential approaches for mitigating these challenges.Third, organisational, regulatory, and standardisation constraints are evaluated. The absence of harmonised human–AI competency frameworks, variability in AI system behaviour across aircraft types, and ambiguities regarding accountability pose obstacles for both initial and recurrent training. A critical need exists for evidence-based human factors methodologies that define the skills required for pilots operating in mixed-initiative or partially autonomous environments. Emerging competency-based training and assessment (CBTA/EBT) methodologies offer a promising foundation but require expansion to incorporate AI teaming competencies, error management strategies, and resilience-building mechanisms.The paper argues that training for human–AI teaming must remain fundamentally human-centric, preserving pilots’ adaptive expertise, situational awareness, and critical thinking while ensuring that AI systems remain compatible with human cognitive strengths and limitations. It concludes by proposing an integrated training model to support safe, resilient, and ethically aligned human–AI cooperation in future aviation operations.
Accident investigation has always been a challenge in disciplined research. Aviation has long been characterised by a delicate balance between human expertise and technological support. From early mechanical automation to contemporary flight management systems, each technological transition has required corresponding adaptations in training philosophy and practice. The current wave of artificial intelligence represents a qualitatively different shift. Artificial intelligence (AI)-enabled systems are no longer limited to executing predefined logic but increasingly exhibit adaptive, data-driven behaviour, learning from operational environments and influencing decision- making in real time. This transition challenges the traditional conception of Received February 6, 2026; Revised April 9, 2026; Accepted April 27, 2026; Available online July 20, 2026
Training Challenges in Human – Artificial Intelligence (AI) Teaming in Aviation 37 automation as a predictable tool and reframes it as an active participant in operational decision processes.
Within this context, human–AI teaming emerges not as a technological option but as an operational necessity. Pilots are expected to collaborate with systems that generate probabilistic recommendations, anticipate future states, and, in some cases, initiate actions autonomously. While these capabilities promise safety and efficiency gains, they also introduce new forms of vulnerability. Automation complacency, erosion of manual flying skills, inappropriate trust calibration, and ambiguity in responsibility allocation have already been observed in highly automated environments. Training systems, largely designed around deterministic automation and rule-based logic, are increasingly misaligned with the realities of AI-mediated operations.
This paper addresses these challenges through a human-centric lens. Rather than asking how pilots can adapt to AI, it asks how training systems can be redesigned so that AI remains compatible with human cognitive strengths, limitations, and ethical responsibilities. The analysis is grounded in contemporary regulatory and conceptual developments, including the European Union Aviation Safety Agency (EASA) Artificial Intelligence Roadmap 2.0, The International Civil Aviation Organization (ICAO)’s competency-based training philosophy, and emerging Safety-II perspectives that treat human variability as a source of resilience rather than error.
CONCEPTUAL FOUNDATIONS OF HUMAN–AI TEAMING Human–AI teaming in aviation represents a paradigmatic shift from tool- based automation toward socio-technical collaboration. Earlier generations of automation were designed around stability, predictability, and procedural containment. Pilots could reasonably anticipate system responses, rehearse failure modes, and anchor their situational awareness in deterministic logic. Artificial intelligence disrupts this equilibrium. Learning systems adapt to data, modify internal representations over time, and generate outputs that are probabilistic rather than rule-bound. As a result, pilots no longer interact with automation as a passive executor of intent, but as an active cognitive agent shaping perception, prioritisation, and decision framing.
From a human factors perspective, this transition has profound implications for cognition, authority, and responsibility. Teaming implies reciprocal influence. Humans adapt their strategies based on system behaviour, while AI systems infer human intent and operational context from data streams. This co-adaptation challenges classical models of situational awareness, which assumed a largely static relationship between operator and system. Instead, shared situational awareness must now be dynamically negotiated between human and machine, often under time pressure and uncertainty.
Trust calibration becomes a central construct in this environment. Trust is neither blind reliance nor sceptical rejection; it is a calibrated relationship grounded in understanding system capabilities, limitations, and uncertainty boundaries. Empirical studies in highly automated domains consistently demonstrate that miscalibrated trust leads either to automation complacency
or to under-utilisation of safety-enhancing functions. AI systems, by virtue of their opacity and adaptive behaviour, amplify this risk. Training systems must therefore move beyond teaching what the system does toward fostering understanding of why it behaves as it does, when its recommendations are reliable, and how to recognise degradation or bias.
The EASA Artificial Intelligence Roadmap 2.0 explicitly frames trustworthiness as a socio-technical property rather than a purely technical attribute. Transparency, human oversight, robustness, and ethical governance are presented as interdependent dimensions. Each dimension carries training implications. Transparency requires that pilots understand not only outputs but also confidence levels and contextual dependencies. Human oversight demands sustained monitoring skills and the confidence to intervene. Robustness requires exposure to degraded and unexpected modes during training. Ethical governance necessitates explicit engagement with responsibility, accountability, and moral agency in AI-supported decisions.
Human–AI teaming must therefore be understood as a competence- intensive activity. It draws on metacognitive skills, critical thinking, and adaptive expertise—capabilities traditionally associated with experienced aviators but insufficiently codified in training syllabi. These human capabilities are not residual skills to be preserved incidentally, but core competencies to be deliberately cultivated.
METHODOLOGY
The study adopts a qualitative, interpretivist methodology aligned with the applied nature of aviation training research. Saunders’ research onion provides the overarching structure, guiding philosophical stance, methodological choice, and data interpretation. An interpretivist philosophy is adopted to capture the lived experience of pilots, instructors, and training designers interacting with AI-enabled systems. The research approach is abductive, integrating theory and practice iteratively to refine understanding of training challenges.
At the methodological level, the ICAO ADDIE instructional design framework—Analysis, Design, Development, Implementation, and Evaluation (ADDIE)—is employed as both an analytical lens and a structuring device for proposed solutions. In the analysis phase, current training practices are examined against emerging AI capabilities and regulatory expectations. The design and development phases focus on identifying competencies, learning objectives, and pedagogical strategies suitable for human–AI teaming. Implementation considerations address organisational readiness, instructor qualification, and integration within Competency-Based Training and Assessment (CBTA)/Evidence-Based Training (EBT) frameworks. Finally, evaluation emphasises continuous feedback, performance data, and Safety-II learning mechanisms.
This dual-framework approach ensures methodological rigour while maintaining operational relevance (Table 1). It allows training challenges to be examined not as isolated deficiencies but as systemic misalignments between technology, people, and organisations.
Table 1: Research methodology overview - alignment of addie phases with human–AI teaming competencies.
Table available in the original PDF paper.
ADDIE Phase Focus Area Human–AI Training Outcome Analysis AI operational impact Identification of cognitive and ethical demands
Design Competency definition Observable human–AI interaction behaviours Development Scenario construction Exposure to uncertainty, drift, and conflict
FINDINGS Key challenges include cognitive adaptation to opaque systems, workload redistribution, trust miscalibration, and ethical responsibility. These challenges demand training systems that prioritise sensemaking, adaptability, and professional judgment over procedural compliance alone are presented in Table 2.
Table 2: Training challenges associated with AI-enabled aviation operations.
Table available in the original PDF paper.
Domain Challenge Training Implication Cognition Probabilistic and opaque system behaviour Sensemaking and uncertainty management skills
Skills Workload redistribution and automation surprise
Deliberate skill preservation and recovery training
Trust Over-reliance or under-utilisation of AI
Trust calibration through scenario- based exposure
Ethics Ambiguity of responsibility and moral agency
Explicit ethical reasoning and Just Culture integration Within the context of human–AI teaming, Functional Analysis and Risk Integration (FAIR) 3 offers a critical lens for understanding how risk emerges not from isolated technical failures but from dynamic interactions among human cognition, intelligent systems, and organisational constraints. FAIR 3 extends traditional risk models by emphasising variability, coupling, and context, making it particularly relevant for AI-enabled operations characterised by probabilistic outputs and adaptive behaviour. In training environments, FAIR 3 supports a shift away from outcome-based risk assessment toward an examination of how decisions are shaped under uncertainty, time pressure, and competing cues. When applied to human–AI collaboration, FAIR 3 encourages pilots and instructors to explore how algorithmic recommendations, human judgement, cultural influences, and organisational norms interact to produce both successful and degraded performance. Embedding FAIR 3 principles within CBTA and EBT frameworks strengthens Safety-II learning by legitimising discussion of near-misses, weak signals, and successful recoveries, rather than focusing exclusively on deviations or errors. In this sense, FAIR 3 functions as a cognitive bridge between advanced automation and human adaptability, enabling training
systems to surface latent risk pathways and support resilient performance in complex, AI-mediated aviation environments.
Cognitive Adaptation and Mental Models One of the most significant challenges lies in supporting pilots’ cognitive adaptation to AI systems whose internal logic may be non-transparent. Traditional automation training emphasised mode awareness and procedural compliance. AI systems require an additional layer of cognitive engagement: understanding confidence levels, data provenance, and contextual validity of recommendations. Without this understanding, pilots may either defer uncritically to AI outputs or dismiss them altogether.
Workload Redistribution and Skill Preservation AI promises workload reduction, yet empirical evidence suggests that automation often redistributes rather than eliminates workload. Monitoring adaptive systems, managing exceptions, and intervening during system uncertainty can increase cognitive load, particularly during non-normal situations. Training must therefore address not only routine operations but also degraded modes, ensuring that manual flying skills and problem-solving capabilities are preserved.
Trust, Ethics, and Responsibility Human–AI teaming raises profound ethical questions regarding accountability and moral agency. While AI may generate recommendations, responsibility for outcomes remains with human operators and organisations. Training must explicitly address this asymmetry, reinforcing ethical reasoning, Just Culture principles, and clear understanding of accountability boundaries. Without such grounding, pilots may experience moral disengagement or undue stress when interacting with autonomous systems.
Pedagogical Implications and Training Design Scenario-based training, instructor upskilling, and integration of explainable AI concepts are essential pedagogical responses. CBTA and EBT provide a robust foundation but require targeted extension to accommodate human– AI teaming competencies.
Organisational and Regulatory Considerations Training does not occur in isolation. Organisational culture, regulatory clarity, and standardisation significantly influence training effectiveness. The absence of harmonised human–AI competency frameworks across regulators complicates curriculum development and assessment. EASA’s roadmap and ICAO’s performance-based philosophy offer a pathway toward convergence but require sustained collaboration among regulators, operators, and training organisations.
Training Challenges in Human – Artificial Intelligence (AI) Teaming in Aviation 41 Safety-II and FAIR 3 perspectives further underscore the importance of learning from normal operations and weak signals. AI systems can support this learning, but only if organisations cultivate a Just Culture that encourages reporting, reflection, and continuous improvement.
DISCUSSION
The findings underscore the need for a systemic rethinking of aviation training. Human–AI teaming cannot be addressed through incremental updates; it requires a transformation of instructional design, assessment philosophy, and organisational learning mechanisms.
The analysis presented in this paper demonstrates that human–AI teaming is not a marginal extension of existing automation practices, but a fundamental reconfiguration of how work, authority, and responsibility are distributed in the cockpit and across the wider aviation system. Training systems that were designed around predictable automation and linear cause– effect relationships are increasingly misaligned with AI-enabled operations characterised by probabilistic reasoning, adaptive behaviour, and contextual sensitivity.
Confusion remains one of the most under-appreciated yet safety-critical challenges in AI-enabled flight operations. Traditionally, confusion in automated cockpits has been framed as mode confusion, arising from pilots misunderstanding the current state or logic of deterministic automation. In human–AI teaming environments, however, confusion assumes more complex and subtle forms. AI systems may generate recommendations that are contextually valid but cognitively opaque, probabilistic rather than deterministic, or temporally misaligned with pilots’ expectations. As a result, pilots may experience uncertainty not only about what the system is doing, but why it is behaving in a particular manner and whether its reasoning remains aligned with operational intent.
In high-workload or time-constrained situations, such ambiguity can erode situational awareness and impair decision-making, especially when multiple sources of information compete for attention. Unlike classical automation surprises, which are often discrete and identifiable, AI-related confusion may persist unnoticed until operational margins are significantly reduced.
From a training perspective, confusion cannot be mitigated solely through increased system knowledge or procedural reinforcement. Human–AI confusion is fundamentally a sensemaking problem, rooted in mismatched mental models, incomplete transparency, and uncertainty about system confidence and limitations. Training programmes must therefore move beyond teaching nominal system behaviour and explicitly address how confusion arises, how it can be detected early, and how it should be managed operationally. This includes developing pilots’ ability to verbalise uncertainty, articulate discrepancies between expectation and system output, and employ structured strategies for validating or challenging AI recommendations. Within CBTA and EBT frameworks, confusion should be treated as an observable and assessable behavioural condition rather than an individual cognitive failure. Effective human–AI teaming is demonstrated when pilots recognise emerging ambiguity, communicate it clearly, manage workload accordingly,
and take timely action to restore shared situational awareness. Scenario-based training that deliberately introduces ambiguous AI behaviour—without clear right or wrong answers—supports the development of these competencies. Debriefings grounded in Safety-II principles further enable crews to reflect on how confusion was managed successfully, reinforcing adaptive strategies rather than focusing exclusively on error avoidance.
FAIR 3 provides an additional lens for understanding confusion as an emergent risk condition rather than a discrete causal factor. From this perspective, confusion arises through the interaction of system design, human cognition, organisational expectations, and operational context. Training that incorporates FAIR 3 thinking encourages pilots and instructors to examine how confusion develops, how it propagates through the system, and how it can be contained before it escalates into unsafe outcomes. Addressing confusion explicitly within human–AI training is therefore not merely a matter of improving system familiarity, but of strengthening resilience, adaptive expertise, and ethical responsibility in increasingly complex aviation environments.
From a training perspective, this misalignment manifests as cognitive brittleness. Pilots may be procedurally compliant yet ill-prepared to interrogate AI outputs, recognise system drift, or resolve conflicts between human intuition and algorithmic recommendations. The findings therefore reinforce the need to reconceptualise competence as adaptive, reflective, and ethically grounded. These attributes cannot be assumed to emerge organically from experience; they must be deliberately cultivated through training design.
For Approved Training Organisations, the implications are substantial. Curriculum design must move beyond teaching system functionality toward developing sensemaking, uncertainty management, and judgment under ambiguity. Instructor roles must expand to include facilitation of reflective learning around AI behaviour, limitations, and ethical tensions. Assessment systems must be capable of capturing not only task success, but the quality of human–AI interaction, including appropriate challenge, monitoring, and intervention.
For operators, the introduction of AI intensifies the importance of organisational culture. Just Culture principles become more fragile in environments where algorithmic recommendations may be perceived as authoritative or blame-shifting. Training must therefore be embedded within broader organisational commitments to learning, transparency, and psychological safety.
For regulators, the findings highlight the urgency of harmonising human– AI competency frameworks and clarifying accountability boundaries. While EASA’s Artificial Intelligence Roadmap 2.0 provides strategic direction, its effectiveness depends on translation into concrete training and assessment requirements aligned with ICAO’s performance-based philosophy.
CONCLUSION
The emergence of artificial intelligence as an operational teammate represents a structural shift in aviation operations and training rather than
Training Challenges in Human – Artificial Intelligence (AI) Teaming in Aviation 43 an incremental extension of established automation paradigms. Human–AI teaming introduces new forms of cognitive demand, ethical responsibility, and organisational complexity that cannot be adequately addressed through traditional, procedure-centred training models. This paper has demonstrated that effective preparation for AI-enabled operations requires a reorientation of training systems toward adaptive expertise, trust calibration, sensemaking under uncertainty, and culturally informed decision-making. As AI systems increasingly shape perception, prioritisation, and risk interpretation, pilots’ ability to engage critically with algorithmic outputs is influenced not only by technical competence, but also by cultural norms related to authority, communication, and uncertainty tolerance. Ignoring these cultural dimensions risks misaligned trust, degraded teamwork, and fragile human–AI interaction in multinational operational environments.
To respond to these challenges, training for human–AI teaming must remain fundamentally human-centric, embedding AI-related competencies within competency-based training and assessment and evidence-based training frameworks. Integrating the ICAO’s ADDIE instructional model with Safety-II, FAIR 3, Cultural Intelligence, and Just Culture principles enables training organisations to cultivate resilience, ethical accountability, and sustained performance across diverse cultural and operational contexts. Cultural Intelligence functions as a resilience-enabling capability, ensuring that human variability becomes a source of adaptive strength rather than misinterpretation when interacting with intelligent systems. Rather than positioning artificial intelligence as a substitute for human judgement, future training systems must ensure that AI remains a cognitively and culturally compatible partner that enhances, rather than erodes, human adaptability. In doing so, aviation can leverage the benefits of artificial intelligence while preserving the human and cultural capacities that remain central to safety in complex, globally distributed operations.
ACKNOWLEDGMENT
The authors thank Coventry University, Purdue University and Turkish Airlines faculty members, HF Horiz ons analysts for their invaluable feedback, which contributed to this work.
REFERENCES
Ang, S., & Van Dyne, L. (2008). Handbook of cultural intelligence: Theory, measurement, and applications. Routledge. European Union Aviation Safety Agency (EASA). (2023). Artificial intelligence
roadmap 2.0: Human-centric approach to AI in aviation. Cologne, Germany: European Union Aviation Safety Agency. https://www.easa.europa.eu/en/domains/ research-and-innovation/artificial-intelligence Hollnagel, E. (2014). Safety-I and Safety-II: The past and future of safety management.
Ashgate. Hollnagel, E., Wears, R. L., & Braithwaite, J. (2015). From Safety-I to Safety-II: A white paper. The Resilient Health Care Net.
International Civil Aviation Organization (ICAO). (2013). Manual of evidencebased training (Doc 9995). Montreal, Canada: International Civil Aviation Organization. International Civil Aviation Organization (ICAO). (2020). Procedures for air
navigation services – Training (PANS-TRG) (Doc 9868, 3rd ed.). Montreal, Canada: International Civil Aviation Organization. International Air Transport Association (IATA). (2024). Competency assessment and evaluation for pilots, instructors and evaluators (3rd ed.). Montreal, Canada: International Air Transport Association. Reason, J. (1997). Managing the risks of organizational accidents. Ashgate. Reason, J. (2008). The human contribution: Unsafe acts, accidents and heroic
recoveries. Ashgate. Saunders, M., Lewis, P., & Thornhill, A. (2019). Research methods for business students (8th ed.). Pearson Education. Ziakkas, D., & Vink, L.-S. (Eds.). (2023). Implementation guide for artificial
intelligence in aviation: A human-centric approach for practitioners and organizations. HF Horizons. Ziakkas, D., Henneberry, D., & Lekea, I. (2025). Implementation of artificial intelligence in transport accident investigations. In Applied Human Factors and Ergonomics Conference Proceedings (Vol. 199, pp. 1095–1104). AHFE Open Access. https://doi.org/10.54941/ahfe1006922 Ziakkas, D., & Plioutsias, A. (2024). Artificial intelligence and human performance
in transportation. https://doi.org/10.1201/9781003480891 Ziakkas, D., Pechlivanis, K., & Plioutsias, A. (Eds.). (2026). Cultural intelligence in aviation: Bridging social attitudes, culture, and safety in a global environment. CRC Press.
LISTED IN THE AHFE DIGITAL BOOKSTORE CATALOGUE
Explore the complete digital volume
Advances in Human Factors of Transportation
Read this individual research paper free through the original publisher, or explore the full digital proceedings book.
ISBN 979-8-950676-00-0
Explore this ISBN in the AHFE Bookstore ↗Listed in the July 2026 bookstore catalogue; search by ISBN to confirm current availability · Digital book purchase optional