For example, in the UK, an HM Revenue & Customs trial tested the effect of adding social norm messaging—like, “9 out of 10 people in your area pay their tax on time”— to letters sent to late taxpayers. The result was a roughly 15% uptick in timely payments, translating into an estimated £210 million in additional revenue in one year. 

Meanwhile, in the U.S., a large-scale study involving more than eight million households showed that providing normative feedback on energy use (letting people know how their consumption compared to their neighbors) consistently reduced average energy consumption by around 2%, effectively abating more than a million metric tons of CO₂.⁹ These examples illustrate how relatively small nudges can generate big impacts—but they also highlight how intricate it is to replicate those successes at an even larger scale, where local contexts, cultural norms, and complex network effects start to play a massive role.

There are countless examples where lab insights were translated into population-level impact. However, for each of those, there are even more examples where the scaling doesn’t work, doesn’t translate to another cultural context or isn’t sustained over time.

For example (and this is likely heresy on a publication that’s seen to promote applied behavioral science), a replication of the UK’s social norm approach to tax compliance in Guatemala saw only a 3.4% increase in timely payments—a far cry from the 15% boost in the UK—highlighting how interventions that thrive in one cultural or institutional context can lose steam elsewhere.¹⁰  Likewise, a large meta-analysis of 100 nudge studies covering nearly two million participants found an average effect size of around 8%, but with a wide spread of results depending on local conditions and target behaviors.¹¹ These variations underscore that while nudges can indeed move the needle at scale, each context is its own “microclimate,” demanding a more nuanced approach than one-size-fits-all. Once we try to replicate them across entire populations, we run smack into the messy world of social networks, emergent feedback loops, and cultural norms.¹ 

Faced with that, the extent to which our interventions scale depends on much more than individual behavior—it becomes a complex system. In other words, it’s not that these micro-level insights become worthless at scale—they just sometimes get overshadowed by the bigger, tangled web of interactions that shape how whole societies behave.

Complex systems are these beautifully messy networks of interconnected parts that give rise to behaviors and patterns we can’t fully predict just by looking at any one piece on its own.³ Some classic examples of this are ant colonies, where thousands of individual ants follow simple local rules—like laying or following chemical trails—and collectively build elaborate networks without centralized control.¹³ Stock markets show the same principle on a human scale; individual traders each chase profit, but their interwoven actions can lead to massive bubbles or crashes that no one person intends.¹² Even highway traffic patterns can shift from free-flowing to jammed with just a small uptick in cars, demonstrating how a subtle change in conditions can trigger a large-scale shift in collective behavior.¹⁴ That’s the defining feature of a complex system:  the interplay of individual parts—be they ants, traders, or drivers—and the constant feedback loops among them that create something far richer and more unpredictable than the sum of its parts.

Complex systems are characterized by feedback loops—positive ones that can amplify small changes and negative ones that can stabilize systems—and by emergent phenomena that pop up from all these interactions.⁵ Think of social movements that begin with a single idea but morph into something vast and unpredictable once millions of people get involved. This is why our neat, tidy insights from small studies can suddenly seem clumsy when we apply them to entire populations: we’re dealing with a sprawling, adaptive system where individuals, environments, and institutions dance around each other in ways no static model can capture.⁴ That’s the real world we have to work with—fun, challenging, and never quite what we expect.

Obviously, the study of complex systems has limitations as well. If we have perfect information about how each element in the system works, we can get a very good idea of how they’ll interact—for example, by running agent-based models. This works surprisingly well in weather prediction or even more physically interactive behaviors, such as disease infection. But when it comes to more complex large-scale behaviors, our predictions are only as good as our understanding of each element (i.e. a person). Therefore, if we want to predict how people will behave in a large system, we also need the understanding that comes from the behavioral sciences.

This is where things get interesting. Applied behavioral science already bases its interventions on a deep (though, obviously, never complete) psychological understanding of human behavior. Therefore, rather than serving as a starting point for a lot of trial, error and iteration, it can also serve as a basis for designing complex behavioral systems.

Here is an example of what this could look like: In 2003, London introduced a congestion charge for vehicles entering its central zone during peak hours. This wasn’t just a simple “nudge” to discourage driving; it was a multi-layered approach that combined pricing, extensive public communication, and improvements to public transport. The results were dramatic: traffic entering the zone dropped by approximately 18% during charging hours, and congestion (the extra time spent traveling compared to free-flow conditions) fell by around 30%. 

Over time, behavioral adaptations emerged: commuter patterns shifted, reliance on public transport increased, and air quality measurably improved. In other words, by altering one part of the ecosystem (the cost of driving in the city center), policymakers triggered a cascade of feedback loops that reshaped people’s transportation decisions in ways no single “one-shot” intervention could have achieved. The problem is they didn’t do it on purpose—they were left to the whims of fate.

A more explicitly “complex” behavioral systems approach might begin by mapping out all the key actors—residents, businesses, policymakers, and community groups—and their interactions before any intervention is designed. For instance, if a city wants to reduce traffic fatalities, it wouldn’t rely on a single nudge, such as a text reminder to drive safely. Instead, planners would build an iterative strategy that draws on insights from psychology, engineering, urban design, and public policy. They might run agent-based models to simulate how changes in street layout, public transit accessibility, and driver education could ripple throughout the community. Next, they’d engage local stakeholders—drivers, cyclists, pedestrians, and public officials—to collect real-world feedback and adapt the plan in real time.

This is the essence of Sweden’s Vision Zero initiative, which explicitly aimed to redesign entire traffic systems rather than just nudge individual drivers. Through changes in road infrastructure, vehicle standards, and legislation, Sweden cut its road deaths to half the EU average by the late 2000s.¹⁵ By integrating behavioral insights into a systemic framework from the outset, complex behavioral systems approaches acknowledge that interventions evolve once they’re deployed—and account for the swirling loops of cause and effect that shape societal outcomes.

Complex behavioral systems have so much potential to be a powerful tool. And that’s the crux of it—we can’t just keep sprinkling small nudges around and hope they magically fix large-scale challenges. Real-world problems—climate change, health inequities, financial instability—are tangled up in massive, evolving webs where interventions morph as soon as they meet reality.⁸ One major shift in today’s behavioral science is recognizing that our standard approach—test an intervention, see if it “works,” then roll it out—doesn’t cut it when the context itself is always on the move. Instead, we need to work iteratively, zooming in on local dynamics and then zooming out to see how our tweaks ripple across networks of people and institutions.⁷

That means co-creating solutions with communities, running continuous experiments, and updating our methods as new feedback loops emerge.⁶ It also means admitting we’re dealing with messy, adaptive systems where cause and effect aren’t neatly separated—where unexpected side effects might be as important as the intended results. None of this should stop us from using behavioral insights; it just means we have to hold them lightly, acknowledging that, at scale, interventions become part of an ongoing dance between our intended design and the system’s response. That dance is messy, but it’s where real change happens.