SCCS-UAH

Navarro-Meneses, F.J., Pablo-Marti, F. Reimagining human agency in AI-driven futures: a co-evolutionary scenario framework from aviation. Eur J Futures Res 13, 16 (2025). https://doi.org/10.1186/s40309-025-00260-w

 

Aviation is often treated as a special case: high reliability, extreme safety constraints, deep regulation, and a culture of procedural rigor. But that “specialness” is exactly why it is analytically valuable. In our recent paper, we argue that aviation is a uniquely revealing testbed for understanding how human roles evolve when AI becomes deeply embedded in socio-technical systems—precisely because the sector cannot simply “move fast and break things.” Instead, it must co-evolve: technology, institutions, and workforce practices have to adapt together, or the system produces hidden fragilities.

The core move of the paper is to reject deterministic automation narratives (“AI replaces humans”) and replace them with a co-evolutionary scenario framework grounded in evolutionary economics and socio-technical transitions: AI integration is a recursive process shaped by technological innovation, institutional selection/legitimation, and workforce adaptation/retention. In other words, the future of human agency is not a technological outcome; it is a governance- and design-dependent trajectory.

 

1) The analytical pivot: from “automation” to “institutionally mediated co-evolution”

The paper uses a variation–selection–retention (VSR) logic to model socio-technical change: innovations emerge (variation), pass through institutional filters (selection), and become stabilized in routines, training, and professional roles (retention). Crucially, these are not linear stages but recursive feedback loops.

Why does this matter for AI? Because AI is not merely a tool deployed into stable organizations. It changes what counts as expertise, shifts accountability boundaries, and creates new dependencies (data ownership, proprietary platforms, certification regimes). If institutions adapt in synchrony with technological diffusion, AI can augment capability while preserving accountability and trust. If institutions lag, AI can create “silent” erosion: hollowed skills, ambiguous liability, brittle crisis response, and legitimacy loss—even if routine performance looks fine.

This is a complex-systems claim: the key risks are second-order and emergent, not necessarily visible in first-order KPIs.

2) Why aviation clarifies the general problem

Most AI adoption research is either (i) micro-level human factors (interfaces, workload, trust experiments) or (ii) macro narratives (jobs apocalypse / productivity boom). The missing middle is the meso-level: how AI scales through organizations and regulatory architectures, and how that reshapes roles and legitimacy over time. The paper argues aviation exposes this middle layer with unusual clarity because certification, incident investigation, recurrent training, and shared standards are intrinsic to the sector.

For SCCS, the broader implication is methodological: if we want transferable insight about AI in high-stakes domains (healthcare, energy, logistics), aviation provides historically comparable transitions and a dense institutional environment where governance variables are observable—not latent.

3) Historical transitions as empirical “anchors” (and why they’re more than anecdotes)

A major strength of the framework is the use of three historical cases as anchors for scenario logic, not as illustrative stories: (1) glass cockpit digitization, (2) predictive maintenance, and (3) partially autonomous cargo drones.

Each case shows a distinct pattern of co-evolution:

3.1 Glass cockpits: skill reconfiguration and “automation complacency” as a systemic property

The transition didn’t simply add digital displays; it shifted the pilot identity toward hybrid system management, creating new training regimes and institutionalized crew resource management (CRM) to stabilize safety and coordination. Importantly, trust and skill retention became governance issues, not just individual cognition problems.

Research angle: This is a canonical example of skill degradation under high automation emerging as a system-level risk—suggesting we should model skills as stocks that depreciate when not used, and recover only under deliberate retraining investments.

3.2 Predictive maintenance: data power, platform dependence, and hybrid roles

Here, the co-evolutionary dynamics include strategic data ownership (OEM vs airline), certification lag, and the creation of hybrid roles (“fleet health managers”) combining mechanical expertise with analytics. The institutional layer (standards, manuals, certification) is the mechanism through which predictive AI becomes operational reality.

Research angle: This is a rich setting for political economy of AI: platform concentration, after-market control, and bargaining power shifts are not side effects—they are trajectory-shaping forces.

3.3 Cargo drones: liminal regulation and contested retention

Autonomous cargo drones highlight institutional “liminality”: experimentation via temporary authorizations and fragmented arrangements, slower retention, and underdeveloped training/accreditation. The case illustrates that innovation can outrun category systems (licensing, command responsibility), producing governance gaps.

Research angle: This is where scenario planning meets regulatory theory: we can study how institutions create “sandboxes” vs “gray zones,” and how those choices shape safety and labor outcomes.

4) The scenario matrix is deceptively simple: the real content is institutional dynamics

The paper’s 2×2 uses two axes: Degree of AI Integration and Institutional Adaptability. This yields four plausible futures: Strategic Co-evolution, Human-Centric Continuity, Latent Obsolescence, and Human Displacement (Figure on the scenario matrix).

The conceptual punchline is that institutional adaptability is the decisive variable. High AI integration is not inherently good or bad; it becomes beneficial or dangerous depending on whether institutions can redesign roles, certification, training, and accountability in step with deployment.

4.1 Strategic Co-evolution (high AI, high institutional adaptability): “AI is everywhere, but humans still own accountability”

This is the “desirable and plausible” scenario: proactive coordination between regulators, unions, and training institutions; AI copilots and AI-supported ATC with robust override competence; role redesign ahead of deployment; transparent decision systems.

The subtlety here: the scenario is not “humans remain central” in a nostalgic sense; rather, human agency is recoded into supervision, ethical judgment, anomaly handling, and accountability stewardship.

4.2 Human-Centric Continuity (low AI, high adaptability): “intentional conservatism”

Institutions adapt well, but deliberately constrain AI’s operational authority due to trust, culture, and risk aversion. The system preserves professional identity and visible human oversight but risks competitive pressure and efficiency gaps if others automate faster.

This scenario is important because it breaks the assumption that “good governance implies maximal AI adoption.” Governance can also justify selective non-adoption.

4.3 Latent Obsolescence (low AI, low adaptability): “stagnation with sophisticated tools on the shelf”

A key insight: obsolescence can be latent—workers keep jobs but lose relevance because institutions cannot integrate tools coherently. You get piecemeal bolt-ons, inconsistent interfaces, and maintenance alerts that cannot be acted upon.

This is a classic complex-systems failure mode: capability exists locally but coordination failure prevents global performance gains.

4.4 Human Displacement (high AI, low adaptability): “fast automation, brittle resilience”

Competitive pressures drive rapid deployment without training, safeguards, or clear liability. Humans become “accountable supervisors without control,” a recipe for legitimacy crises and brittle disruption handling.

For SCCS, this scenario is analytically fertile: it foregrounds the control–accountability mismatch as a structural risk variable.

5) Delphi validation: not just “expert opinion,” but a structured measurement of plausibility and levers

The paper uses a three-round Delphi with 16 senior aviation stakeholders (20+ years experience) to validate and refine scenarios. Strategic Co-evolution dominates plausibility across rounds; Human-Centric Continuity is the stable runner-up; the other two are treated as risk scenarios with diagnostic value.

Two points matter for research design:

1. Delphi as a mechanism for surfacing contested assumptions, not merely for consensus. The paper explicitly treats scenarios as “permeable”—systems can move between trajectories via institutional interventions or shocks.
2. Institutional adaptability emerges empirically as the leverage variable, consistent with the theory: when governance keeps pace, futures cluster toward the top row; when it lags, risk trajectories become more likely.

This is an empirically actionable finding: adaptability can be operationalized and measured (see research agenda below).

6) Second-order effects: where complex systems thinking becomes indispensable

A particularly SCCS-relevant contribution is the explicit discussion of second-order effects: trust erosion, professional identity hollowing, market concentration via proprietary AI platforms, liability/insurance uncertainty, and geopolitical divergence enabling regulatory arbitrage.

These effects are structurally similar across domains:

• Trust behaves like a slow variable: it accumulates through transparent accountability regimes and collapses after salient failures.
• Skills are path-dependent: once hollowed, they cannot be reconstituted quickly during crises.
• Concentration changes system topology: fewer suppliers mean correlated failure modes and bargaining power shifts.
• Liability shapes incentives: unclear responsibility encourages blame shifting and underinvestment in safety assurance.
• Geopolitics introduces coupled regimes with different safety equilibria.

For complex systems researchers, these are the variables that should sit in the model—not as narrative garnish, but as state variables interacting with adoption dynamics.

Research lines for SCCS: from scenario logic to testable, computationally grounded programs

The paper itself proposes several directions—cross-sector transfer, hybrid modelling (ABM/system dynamics), workplace-level research, comparative governance, legal/ethical infrastructures, and “horizon technologies” like AGI. Building on that, here is a more SCCS-style agenda with explicit empirical and modelling hooks:

A) Measuring “Institutional Adaptability” as a latent construct with observable indicators

Goal: turn the key axis into an index measurable across sectors/countries.

• Indicators: certification cycle time for new AI functions; existence and scope of regulatory sandboxes; training curriculum revision frequency; union–management AI governance bodies; post-incident transparency standards (auditability/traceability requirements).
• Methods: Bayesian latent variable models; comparative case coding; event-history of regulatory updates across FAA/EASA/ICAO contexts (paper flags comparative governance as future work).

B) Hybrid modelling: scenario “stress testing” with ABM + system dynamics

Goal: move beyond qualitative narratives while preserving uncertainty structure.

• ABM layer: heterogeneous actors (regulators, airlines, OEMs, unions, training academies, workforce cohorts), each with incentives, constraints, and learning.
• System dynamics layer: slow variables (trust, skill stocks, institutional capacity), reinforcing loops (automation → fewer manual interventions → skill decay → higher fragility), and balancing loops (incidents → regulation tightening).
• Output: phase diagrams over the 2×2 space; tipping regions where small shocks push the system from Strategic Co-evolution to Latent Obsolescence, etc.

C) Data ownership and platform power as a network-structure problem

Goal: formalize the “OEM/platform concentration” pathway.

• Model the aviation AI ecosystem as a multiplex network: data flows, certification dependencies, maintenance service contracts, software update channels.
• Study concentration metrics (centrality, single points of failure), correlated risk propagation, and bargaining power effects on safety investments.

D) Accountability engineering: designing auditable AI functions and liability-compatible logs

Goal: connect socio-technical governance with technical design requirements.

• Research questions: What minimal “evidence schema” (logs, explainability artifacts, traceability) makes post-event accountability feasible without exposing proprietary IP?
• Empirical hook: compare emerging regulatory roadmaps and incident investigation practices; the paper explicitly recommends shifting from certifying artefacts to certifying AI functions with auditability/traceability requirements.

E) Skill dynamics and “role hollowing” as measurable risk

Goal: quantify hollowing and design countermeasures.

• Use training records and simulator curricula as proxy data; develop a “skill retention” metric tied to frequency of manual practice and exposure to AI failure modes.
• Intervention experiments: AI-in-the-loop recurrent training vs conventional training; measure response quality under rare disruption scenarios (cyberattack, ash cloud, cascading delays), explicitly invoked in the Strategic Co-evolution narrative.

F) Scenario permeability: modelling transitions between futures

Goal: treat scenarios as attractors, not endpoints (as emphasized by Delphi participants).

• Identify triggers: major incidents, regulatory harmonization shocks, labor strikes, OEM platform lock-in events, geopolitical divergence.
• Use Markov switching / regime-change models informed by expert elicitation + historical analogues.

References

• Geels, F.W. (2006). “Co-evolutionary and multi-level dynamics in transitions: the transformation of aviation systems and the shift from propeller to turbojet (1930–1970).” Technovation, 26(9), 999–1016.
• Felt, U., Wynne, B., Callon, M., et al. (2007). Taking European knowledge society seriously. Luxembourg: DG for Research (EUR 22700).
• EASA (2023). Artificial Intelligence Roadmap 2.0: Human-centric approach to AI in aviation.
• Downer, J. (2010). “Trust and technology: the social foundations of aviation regulation.” British Journal of Sociology, 61(1), 83–106.

Closing thought: the future of agency is a design variable

The strongest claim in the paper is also the most actionable: the “human role in AI-driven futures” is not determined by model capability alone. It is determined by how quickly and intelligently institutions can redesign accountability, training, and legitimacy under deep uncertainty. Aviation teaches this lesson early because it has to. Our opportunity—at SCCS and in allied complex-systems communities—is to generalize it rigorously: measure institutional adaptability, model co-evolutionary feedbacks, and turn scenario narratives into stress-tested, policy-relevant computational artifacts.