Against the wall of complexity: human cognition, political decision-making, and human-AI collaboration

In 2014, the report A World That Counts, commissioned by the United Nations Secretary-General, argued that better access to high-quality data would improve public decision-making to address major global challenges more effectively (Independent Expert Advisory Group on a Data Revolution for Sustainable Development [IEAG], 2014). Since then, data availability has increased substantially, yet the corresponding benefits for public decision-making remain mixed (Fischer et al., 2025). If the difficulties political systems face in addressing contemporary problems persist despite this abundance, data availability cannot be the whole explanation. Policy processes and outcomes also depend on political pressures, institutional rules and interactions among policy actors, limited administrative capacity, and uncertainty about policy problems and their solutions (Howlett et al., 2015). These influences increasingly operate within complex systems, which may amplify the demands each of them creates (Fernández-i-Marín et al., 2024). We focus here on the cognitive processes through which policymakers interpret information, form judgments and make decisions under these demands. We define policymakers as individuals exercising decision-making responsibilities within governance systems, from the local to the supranational level. Following Cairney and Weible (2017) in viewing policymaking as imperfect cognition operating in complex contexts, we contend that the interaction between cognitive limits and rising systemic complexity is an important source of difficulty in policymakers' judgments and decisions. To support this thesis, we draw on research in complex systems, cognitive psychology and decision science, before examining how systems thinking, simulations and human-AI collaboration can help mitigate these limitations. We conclude by arguing for humility about what can be predicted and controlled in complex policy systems (Jalonen, 2025).Contemporary societies have become more complex as digitalization and globalization have deepened interdependencies among social, economic, and political spheres (Benbya et al., 2020;OECD, 2019). In such systems, interactions among actors generate unpredictable dynamics, limiting the ability of policymakers to anticipate how these systems evolve (Geyer & Rihani, 2010). The problems governments face are similarly complex: they evolve over time, emerge from multiple interactions, and involve actors with divergent interests. Their ambiguity leaves room for multiple interpretations, such that evidence alone is insufficient to determine what should be done (Cairney, 2019). At the global level, new governance arrangements have emerged to address such problems, including multi-stakeholder partnerships, adding another layer of institutional complexity (Elsässer et al., 2025). Yet responses of this kind can prove counterproductive, raising the risk of wicked solutions that leave decision-makers with a system more complex than before (Peters et al., 2019). Within national systems, policies accumulate (Adam et al., 2019), often faster than public administrations expand their capacity to implement them (Fernández-i-Marín et al., 2024), while policy texts themselves add complexity through demanding language and connections to other legal instruments (Senninger, 2023). Complexity is therefore not only a feature of the environment policymakers face, but is also amplified by the responses they devise.Decades of research on complex problem solving and dynamic decision-making document persistent difficulties in understanding complex, dynamic environments (Dörner & Funke, 2017;Osman, 2010). Following Béchard et al. (2025), we use the wall of complexity as a descriptive label for the substantial deterioration in judgment and decision performance that can occur when decision-makers confront highly complex tasks. It does not denote a single or fixed threshold. Although capacity limits have been quantified for specific processes such as working memory and relational reasoning, the extent of this deterioration depends on task characteristics, individual differences, the strategies used, and the external support available. Among the many ways complexity has been described, we retain the three dimensions that bear most directly on cognitive demand: computational, relational, and dynamic complexity.Computational complexity arises when the number of possible ways to solve a problem grows so rapidly that exhaustive consideration becomes infeasible (Hélie & Pizlo, 2022). In politics, we see this in the multiplicity of actors, objectives, and policy instruments that can be combined. Because working memory holds only a few items at a time (Cowan, 2010), decision-makers can consider only a fraction of these possibilities.Relational complexity reflects how many interdependent variables must be integrated simultaneously to represent a problem (Halford et al., 1998). Performance drops substantially from three to four variables, and falls to chance level at five (Halford et al., 2005). Political decision-making routinely involves many more interacting elements, as governance networks connect numerous interdependent public, private, and nonprofit actors (Morçöl, 2023).Finally, dynamic complexity arises from feedback loops, time delays, and nonlinear relationships that shape how systems evolve (Sterman, 1994). Political decision-making reflects these dynamics, as choices made in one sphere of governance can alter decisions taken in another (Klijn & Snellen, 2009). Yet humans have particular difficulty understanding delayed feedback (Brehmer, 1992) and basic dynamic processes (Cronin et al., 2009).These forms of complexity impose demands that can exceed human cognitive capacities. How cognition adapts to them is therefore central to understanding political decision-making.Whether policymakers can escape the wall of complexity depends on a prior question: are they cognitively better equipped to withstand it than ordinary citizens (Béchard et al., 2025)? Although political science has often treated political elites as cognitively distinct from the public (Kertzer, 2022), experimental comparisons reveal few systematic differences in how elected officials and citizens process information and make judgments (Griffin, 2013;Vis & Stolwijk, 2022). This question also extends to the organizations in which policymakers work. Such organizations can increase their information-processing capacity by delegating routine matters to specialized units, allowing multiple issues to be handled in parallel. When a problem cannot be adequately addressed through established procedures, however, it requires central coordination and a sequential search for alternatives, making the limits of individual cognition more consequential (Jones, 2003).Organizations can therefore distribute and sometimes mitigate the demands of complexity, but they do not eliminate the human cognitive constraints associated with the wall. Political decision-making thus remains cognitively bounded (Cairney, 2019), requiring policymakers to simplify what they attend to, represent, and reason about.Policymakers face chronic information overload that exceeds what they can process (Walgrave & Dejaeghere, 2017). Since attention is a limited resource, they must allocate it selectively, focusing on a few problems at a time (Baumgartner et al., 2023). Part of this selection is delegated to staff, while the rest relies on heuristics that filter what merits processing (Walgrave & Dejaeghere, 2017). What fails this first filter never reaches the later stages of the decision-making process. Such filtering can also have representational consequences, making some interests more cognitively accessible than others and leaving decision-makers with an incomplete picture of those affected (Miler, 2009). Nor is what passes the filter necessarily represented accurately. Even when a problem is noticed, its severity may be underestimated. When deterioration occurs through small, routine changes, the reference point for judging the problem may gradually adjust, making its cumulative severity appear less urgent -a boiling frog dynamic that Epp and Thomas (2023) associate with slower policy responses.Mental models organize incoming information by selecting the variables and causal relationships deemed relevant, but tend to neglect accumulations, delays, feedback, and nonlinearities (Sterman, 1994). They further reduce complexity by grouping elements into larger units or processing them sequentially (Halford et al., 1998), so reasoning ultimately operates on a simplified representation rather than the full problem (Hélie & Pizlo, 2022).While simplification is useful for making complex tasks more tractable, it can also produce systematic errors in judgment (Tversky & Kahneman, 1974). Experimental evidence suggests that elected office does not insulate decision-makers from these errors. In a three-country study, elected representatives were strongly affected by gain-loss framing and status-quo presentation, and were more likely than citizens to continue investing in a failing program after substantial prior expenditure (Sheffer et al., 2018). Local councillors are similarly prone to conjunction errors and scope neglect (Stolwijk & Vis, 2021). Yet what matters for the quality of political decision-making is less the occurrence of these errors than the resistance of these errors to correction. Individuals readily detect distortions in the reasoning of others while judging their own conclusions to be sound (Pronin et al., 2002). Among politicians, additional evidence does not reduce, and can even increase, the influence of prior attitudes on how information is interpreted (Baekgaard et al., 2019). Requiring them to justify a decision has similar effect, strengthening motivated reasoning rather than tempering it (Christensen & Moynihan, 2024). Expertise and experience appear to offer little protection against such biases (Stolwijk & Vis, 2021). Despite reporting greater political knowledge, elected officials did not outperform the general population in a complex political simulation (Béchard et al., 2025).As the cognitive limits discussed above are structural, they cannot be eliminated, nor can decisionmakers overcome their effects on their own. The challenge is therefore to equip them to navigate complexity more effectively despite these constraints. We examine three ways of doing so: training decision-makers to reason about complex systems (Section 5.1), extending their capacities through human-AI collaboration (Section 5.2), and cultivating epistemic and regulatory humility about the limits of their judgment and of policy itself (Section 5.3).A first step is to develop systems thinking by learning to reason about the overall structure of a system or problem rather than its isolated elements (Fischer & Gonzalez, 2016). In our view, this should include an explicit understanding of key mechanisms of complexity, such as feedback, delays, nonlinear effects, and emergence. Complexity thinking provides an analytical lens for recognizing the evolving and partly unpredictable nature of these systems, and the resulting limits of prediction and control (Geyer & Rihani, 2010). The aim should be to foster what Funke (2022) calls complex systems literacy. Yet the dynamics of complex systems are often too counterintuitive to be inferred through unaided reasoning (Sterman, 2002). Merely thinking and talking about systems may therefore do little to reshape the mental models underlying decision-making, emphasizing the need for active modeling (Forrester, 2016). Although empirical evidence suggests that modelling used as a cognitive tool combined with explicit systems-thinking instruction can improve systems thinking (Hung, 2008), routinely building such models may be impractical for policymakers. We therefore see microworld simulations as a more accessible route to experiential training. These computerized environments incorporate key characteristics of complex systems while allowing participants to experiment with strategies and observe their consequences, without having to build the models themselves (Brehmer & Dörner, 1993). Microworlds have long been used to study complex and dynamic decision-making (Gonzalez et al., 2005), including in simulated governance environments, such as Ecopolicy (Lafond et al., 2012), and Democracy 3 (Béchard et al., 2025;Lafond et al., 2014).We see this as a promising foundation for extending their use to the training of policymakers. However, the effectiveness of such training depends on how the microworld is used rather than exposure alone (Sterman, 1994). Decision-making performance improves when training favours fewer, longer practice episodes over a greater number of short repetitions (Gonzalez, 2004), and when participants observe how an expert handles the task rather than reviewing their own decisions (Gonzalez, 2005). These findings, however, come from laboratory studies with university students, and whether skills acquired in microworlds transfer to real policy decisions remains to be established. And even if they do, training, however well designed, can change how a decision-maker approaches a system, but not eliminate the cognitive limits that constrain what they can process. That asymmetry is where tools designed to support decision-making become worth considering.Training can improve how decision-makers reason about complexity, but unaided cognition still has limits. A complementary approach is therefore to externalize part of the cognitive burden through decision-support technologies. Such tools can function as "cognitive prostheses," offloading part of the information-processing and representation demands onto external systems (Tremblay et al., 2017). AI considerably expands the possibilities for such cognitive support, not by providing decision-makers with more information, but by increasing their capacity to process and act upon it. It can support routine, information-intensive policy analysis (Wellstead et al., 2024), transform large volumes of raw data into decision-relevant information (Madanchian et al., 2026), and extend policy modelling by allowing decision-makers to test interventions before implementation (Margetts, 2022). Yet such cognitive support comes with a trade-off: the systems that help decision-makers navigate complexity can also shape how that complexity is perceived and judged. Algorithmic risk assessments, for example, can recast fundamental uncertainty as statistical risk (Hartmann & Wenzelburger, 2021), while AI-generated recommendations can become anchors for subsequent human judgments (Carter & Liu, 2025). More generally, cognitive vulnerabilities change with the degree of automation in the decision-making system: as human involvement decreases, risks shift from biases in human judgment toward excessive trust in the system and disengagement from the decision-making process (Van Rooy, 2024). These vulnerabilities mean that adding AI to human decision-making does not automatically lead to better decisions. Depending on how it is used and implemented, AI may mitigate some biases while reinforcing or creating others (Bertrand et al., 2022;Kliegr et al., 2021). AI may also relocate part of this complexity rather than reduce it, since machine-learning systems derive their outputs from high-dimensional computations that do not readily translate into humanly interpretable reasons (Burrell, 2016). Decision-makers who rely on systems they cannot fully understand can find it harder to answer for the decisions those systems inform, a difficulty that Grimmelikhuijsen and Meijer (2022) count among the threats to the legitimacy of algorithmic decision-making in government. We argue that these risks are not a reason to reject cognitive offloading, but to design human-AI collaboration more carefully. The challenge is therefore not to maximize AI involvement, but to design for complementarity between human and artificial cognition (Tremblay et al., 2026). Decision-support systems should make the goals, assumptions and uncertainties behind their recommendations intelligible to decision-makers, and allow these recommendations to be audited (Grisogono, 2020). This also requires AI systems specifically developed for public-policy contexts (Henderson et al., 2022), explicit allocations of human and algorithmic roles and degrees of autonomy (Simmler and Frischknecht, 2021), and anticipation of the organizational disruptions their implementation within governments creates (Tangi et al., 2025). Conceived in this way, human-AI collaboration could expand the ability of policymakers to navigate complexity beyond their natural capabilities, without replacing the judgment necessary for decision-making.There are several approaches to support human cognition; some are applied upstream, such as the development of systems thinking or the use of computer simulations, while others, such as AI, can enhance the decision-support tools used during the decision-making process itself. But these tools alone will not suffice, because training to deal with complexity involves not only acquiring new tools but also developing a certain form of epistemic humility. First and foremost, it involves recognizing that decision-making biases exist, that they have structural origins (Korteling et al., 2020), and that they influence our judgment even when we are aware of them (Pronin & Kugler, 2007). The next step is to be aware of the vulnerability of unaided human cognition in the face of complexity, and the need for support beyond individual effort alone. Jalonen's (2025) review recommends a humble approach, by recognizing that the properties of complex political-administrative systems make their evolution inherently difficult to predict and control. This mindset can be learned. Dunlop and Radaelli (2016) propose training practitioners in what they call regulatory humility, that is, questioning the limits of what can be regulated, recognizing that certain situations are partly beyond the control of public policy, and confronting their own illusions of control. Making policymakers aware of what, amid complexity, will always remain beyond their control is not an admission of powerlessness: it is the prerequisite for working around the wall rather than running into it.Political systems are becoming more complex. Under high complexity, the natural capacities of human cognition may no longer be sufficient for effective judgment and decision-making. Faced with this wall, policymakers adapt the way they process information by simplifying reality. Heuristics are a natural response, but they can also become a source of bias and errors in judgment. Suboptimal political decisions are therefore not necessarily a sign of individual incompetence but may also reflect cognitive mechanisms that reach their limits when confronted with high levels of complexity.We posit that the interaction between the universal limits of human cognition and the complexity of the contexts in which this cognition operates (Cairney & Weible, 2017) is an important, often overlooked, source of difficulty in political decision-making. We have proposed three complementary responses to this difficulty: developing systems thinking through training and simulations, extending decision-making capacity through human-AI collaboration, and cultivating the epistemic and regulatory humility needed to use both well. None of these responses removes the cognitive limits of policymakers, but together they can help them navigate complexity more effectively.