Andrew J Stier
7 min
While implicit biases have generally declined in the United States over the last two decades, the mechanisms governing the speed and persistence of these changes remain poorly understood. This paper investigates how urban environments—specifically city population size—shape the dynamics of implicit bias change and the effectiveness of interventions designed to reduce them.
The author develops a mathematical theory of implicit bias change grounded in urban scaling theory and expertise-based learning. The model posits that implicit bias levels are influenced by intergroup exposure and that the rate of change is constrained by the social and cultural environment of a city. To test this, the author analyzes 14 million implicit association test (IAT) responses from U.S. residents over 14 years and conducts an experimental study with 1,150 participants to measure how quickly the effects of a counterstereotypic intervention fade across cities of different sizes.
The study reveals that implicit biases change more slowly in larger cities compared to smaller ones. This counterintuitive result stems from the fact that larger cities facilitate higher rates of social interaction and exposure to cultural information, which creates a more stable, reinforced social context. Consequently, while one-off interventions can temporarily reduce bias, these effects dissipate significantly faster in larger cities. The author demonstrates that this pattern is consistent across various types of implicit biases and aligns with a model where individuals have limited cognitive capacity to process the high volume of cultural information present in large urban centers.
These findings challenge the assumption that interventions will have uniform effects regardless of the social context. By demonstrating that the urban environment acts as a structural moderator of cognitive change, this research suggests that future bias-reduction efforts must be tailored to the specific social and cultural density of the target population. It provides a quantitative foundation for understanding how structural factors, such as city size, interact with individual psychology to influence societal-level shifts in attitudes.
Implicit biases remain a contributor to discrimination despite widespread endorsement of equality. Although there has been a large-scale reduction in sexuality, race, and skin tone biases in the past two decades, there is a current shortage of quantitative theories that describe how and why biases change. Here, I present a theory of implicit bias change, which focuses on changes in exposure due to cultural information and how the speed of these changes varies geographically. This model makes counterintuitive predictions when compared with previous findings, including 1) that biases change more slowly in larger cities and 2) that one-off interventions are less effective in larger cities. I confirm these predictions using 14 million tests of implicit biases from U.S. residents over 14 y and experimental data from 1,150 individuals. Documenting the typical timescales and patterns that govern implicit bias change is an important step toward designing interventions that can facilitate further reductions in biases and prevent biases from increasing.
Alex: That's where a careful referee would push back, and the paper is genuinely constrained here. The author uses population size as a proxy for interaction rate, which is theoretically grounded in the scaling literature, but doesn't fully disentangle it from correlated urban characteristics. The model controls for baseline bias levels across cities, and the scaling relationship between population and interaction rate is treated as established prior work rather than something estimated fresh in this dataset. So the causal interpretation — that it's specifically interaction density driving faster decay, rather than some other dimension of urbanicity — is supported by the theoretical framework more than by direct identification.
Sam: So the mechanism is plausible and the pattern is in the data, but the causal chain from population size to interaction rate to decay speed isn't directly observed — it's inferred.
Alex: That's a fair characterization. What the data does show clearly is the population-decay correlation at scale. The mechanistic story — finite cognitive capacity, dominant ambient signals effectively resetting the intervention effect — is the author's theoretical account of why that pattern exists. It's coherent, and it fits the scaling literature, but it's not independently validated within this paper.
Sam: What about heterogeneity across bias types? The IAT covers everything from racial attitudes to gender stereotypes. Does decay rate differ depending on what's being measured?
Alex: The paper examines this, and it's worth flagging as a secondary finding rather than a primary one. Some bias domains show more persistent effects than others, which the author attributes to differences in how deeply embedded those attitudes are in the ambient cultural signal — essentially, how much the surrounding environment is continuously reinforcing the bias being targeted. But sample sizes within individual bias categories are uneven, so those comparisons should be read with caution.
Sam: What are the practical implications? Because if intervention half-life is fundamentally shorter in dense urban environments, that has real consequences for how organizations design these programs.
Alex: The paper argues for what it calls "dosage-matched" interventions — repeated, high-frequency exposures calibrated to the information entropy of the local environment. If the cultural noise floor is higher in a large city, you need a correspondingly stronger or more persistent signal to achieve the same durable effect. A single workshop that might produce lasting change in a smaller community needs to be reinforced more aggressively in a metropolitan context.
Sam: Which is a more demanding prescription than most current practice. Most organizational interventions are still structured as discrete events.
Alex: And that's the practical tension the paper surfaces. The scaling framework predicts that discrete, one-off interventions will have systematically shorter effects in precisely the environments — large cities — where most major institutions are located. There's also a methodological implication that the paper doesn't fully develop but which follows naturally from the framework: if you're running a multi-site trial and not controlling for city size, you may be systematically misreading which programs work. Effect sizes measured six weeks post-intervention in a large metro could be substantially attenuated relative to what the same program produces in a smaller setting, which means cross-site comparisons of program effectiveness are potentially confounded by population size in ways that are rarely accounted for.
Sam: That's a meaningful point even before you get to the intervention design question. Population size as an unmodeled source of variance in multi-site trials — that's worth taking seriously on its own.
Alex: Right. And what's missing to close the loop is a design that can directly observe the interaction-rate mechanism rather than inferring it from population size — something like variation in interaction density within cities, or natural experiments that shift local social density independently of other urban characteristics. The IAT dataset is large enough that the population-decay relationship is unlikely to be a statistical artifact. The urban scaling framework gives you a principled reason to expect it. But the identification gap is real, and future work needs to close it.
Sam: So the honest summary: a robust empirical pattern, a theoretically grounded mechanism, and a causal identification problem that remains open.
Alex: That's about right. It's a paper that sits at an unusual intersection — social attitude research and urban science don't often speak to each other directly — and the framework it proposes is worth stress-testing with better-identified designs. Thanks for listening to ResearchPod.