Behavioural interventions land in social worlds where people copy one another, though not all in the same way. How can we use these insights to deliver effective policy? Giuseppe Alessandro Veltri and Alberto Acerbi develop an agent-based model that shows an identical seeding campaign can stall, drift or cascade depending on whether people follow the majority, the prestigious or the successful. They argue that policy design should pay more attention to how people learn from each other.
Behavioural public policy is having an uncomfortable conversation with itself. Meta-analyses suggest that nudges work, but modestly and unevenly. Comprehensive evidence from government nudge units shows that effects at scale are far smaller than the academic literature once promised. One response is to conclude that interventions are simply too weak. Another is to ask whether we evaluate them while ignoring the social machinery through which they spread. In a new open-access paper in Behavioural Public Policy, we take the second route.
Our starting point is simple: behaviour-change interventions are almost never a private transaction between a policymaker and an isolated individual. They land in communities where people watch, discuss and copy one another. Whether an intervention takes off can depend less on its content than on how the people around it learn from each other.
People copy each other, but not at random
Decades of research in cultural evolution show that people deploy social learning strategies: rules of thumb about when and whom to imitate. Four are canonical:
- Conformists disproportionately copy whatever the majority does: a behaviour must feel common before they join in.
- Prestige-biased learners copy high-status or respected individuals.
- Payoff-biased learners copy whoever seems to be doing well out of their choices.
- Random copiers simply imitate a randomly encountered peer.
A growing ‘fourth wave’ of behavioural science, known as cultural evolutionary behavioural science, argues that these transmission rules drive population-level change at least as much as individual biases do. Real populations are also mosaics of such learners: a neighbourhood dominated by conformists is a different policy environment from one where success is highly visible, even if every individual receives the same message.
One intervention, four different fates
To isolate what learning rules do, we built a deliberately minimal agent-based model: a simulated population of 400 individuals who repeatedly decide whether to adopt a behaviour by observing others. Every simulation compares two identical worlds: a control arm that starts with 5% of people already adopting, and a treatment arm where a campaign additionally ‘seeds’ adoption up to 25%. The seeding is the whole intervention; everything that happens afterwards comes from social learning. We then track the gap in adoption between the two arms over time, which we call the intervention’s ‘lift’.
Figure 1: The same 20-point head start on behaviour change adoption meets four very different fates
Notes: Adoption over time in treatment (green, seeded to 25%) and control (blue, 5% baseline) when everyone uses the same learning rule. Faint lines are individual simulation runs; solid and dashed lines are replication medians and means (400 agents, 150 steps, 100 paired replications). Source: Veltri & Acerbi (2026); model code openly available on OSF.
As Figure 1 shows, there are three notable trends:
- Among conformists, the intervention stalls: even 25% adoption sits below the majority threshold, so both arms collapse back towards zero and the campaign leaves no trace, echoing older threshold models of collective behaviour and more recent tipping-point experiments.
- Among random copiers, adoption drifts: there is no directional force, so the head start largely persists. Prestige-biased copying preserves a solid gap too, provided prestige actually tracks who benefits from the behaviour.
- Among payoff-biased learners, finally, the behaviour cascades, and it does so in both arms.
Real populations are mixtures, and mixtures matter
No real community is made only of conformists or only of payoff-chasers, so we also simulated mixed ecologies in which one rule dominates (70% of agents) and the other three share the rest. A natural intuition is that a minority of payoff-biased learners would drag every mixture towards full adoption.
The simulations tell a different story (Figure 2). Conformist-dominant mixtures still stall: even when followed for ten times longer, a 30% minority of other learners cannot push the population across the majority threshold. Payoff-dominant mixtures saturate in both arms, erasing the final gap. In between sit the interesting cases: prestige-dominant mixtures slowly climb to high adoption in both arms, leaving only a modest final lift (treatment 93%, control 90%), while random-dominant mixtures preserve the largest residual advantage (treatment 78%, control 69%), although control gradually catches up the longer one watches.
Figure 2: Behaviour change adoption in communities with mixed learning rules
Notes: Adoption over time in mixed 70/30 ecologies, where one learning rule dominates (70% of agents) and the remaining 30% is split among the other three. Treatment in green, control in blue; note the different time horizons per panel, chosen to make slow convergence visible. Source: Veltri & Acerbi (2026).
Look at the panel where the dominant learning rule is payoff-biased learners. The evident and sweeping rise in adoption from 0% to 100% demonstrates how narrow the window of time can be for learning to take place. Final lift peaks when roughly 20% of the population is payoff-biased, a share large enough to move the system without saturating it, and collapses again by 30%. Even at its peak the residual gap is small. In other words, the main effect of payoff-biased learners is to flip the whole system from stagnation to saturation rather than to preserve the intervention’s advantage.
When success erases its own evidence
The payoff panels in Figures 1 and 2 carry a sting. In payoff-driven ecologies the intervention clearly helps, in the sense that the treated population reaches high adoption much sooner. But because the behaviour is visibly rewarding, the control arm eventually catches up. Measure outcomes at the end, and the trial reports a null effect. The intervention changed the timing of diffusion rather than its eventual reach, and an evaluation that only looks at the endpoint cannot tell the difference.
Our robustness checks sharpen the point. When adopting carries a clearly negative payoff, nothing spreads, whatever the seeding. When payoff signals are strongly positive, everyone gets there anyway. In these setting, it is no surprise that conventional efforts to evaluate interventions suggest modest or short-lived treatment effects. The biggest durable treatment effects appear in between, in scenarios where payoffs are weak or noisy, and where conformist populations sit close to a tipping point. The measured ‘effect’ of an intervention, in other words, is largest where the social system leaves it room to matter.
Designing for the learning ecology
If the same intervention can stall, drift or cascade depending on how people learn, then the learning environment itself becomes a design parameter. Our simulations suggest conditional and context-tailored recommendations for policy designers:
- In conformist-heavy settings, diagnose local thresholds and either seed above them or change what people perceive the majority to be doing.
- Where prestige matters, recruit demonstrators who are credible in the eyes of the community, as in the well-known experiments that seeded information through socially central ‘gossips’.
- Where payoffs drive imitation, make good outcomes visible and attributable, and expect early acceleration rather than a permanent gap.
- In weakly structured settings, broad seeding beats narrow targeting.
Our findings speak to the i-frame versus s-frame debate: social learning ecologies are the meso-level machinery that connects individual choice architecture to system-level change. They also raise ethical questions of their own, since amplifying prestige or payoff cues means deciding whose status and whose outcomes get amplified, which calls for the kind of transparency emphasised in the boosting tradition.
The model is deliberately stylised, a proof of concept rather than a calibrated forecast. Its central lesson travels well, though: the diffusion potential of an intervention depends jointly on what we seed and on how people learn from one another. Before asking whether an intervention ‘works’, we should ask what kind of learners it is about to meet.
You can read the full article in Behavioural Public Policy.
Giuseppe Alessandro Veltri is Full Professor of Computational Social Science and Behavioural Data Science at the Department of Sociology and Social Research, University of Trento, and is affiliated with the Centre for Behavioural and Implementation Science, Yong Loo Lin School of Medicine, National University of Singapore. His research combines computational methods and behavioural science to study risk perception, social influence and public policy. His book Algorithms of Care: Behavioural Science Meets AI in Health is forthcoming with Cambridge University Press.
Alberto Acerbi is Associate Professor at the Department of Sociology and Social Research, University of Trento, and a member of the Centre for Computational Social Science and Human Dynamics. He works on cultural evolution at the interface of psychology, anthropology and sociology, and is the author of Cultural Evolution in the Digital Age (Oxford University Press) and Individual-Based Models of Cultural Evolution (Routledge).