Learning in a Correlated World
GAICIL Insights
Every choice is also an experiment on the choices next to it.1 Learning in a Correlated World, by Arjada Bardhi and Steven Callander, surveys a growing literature on exactly that: how information from one choice changes beliefs about others when outcomes are correlated. The core modeling innovation represents the unknown mapping from choices to outcomes as the realized path of a stochastic process, most commonly Brownian motion. Knowledge, in this view, has geography.
The framework has been applied to four canonical problems where correlated learning is key yet understudied: search and experimentation, communication, innovation and market competition, and attribute problems. The authors review emerging empirical and experimental evidence and outline new theoretical and methodological directions. The approach captures informational spillovers across players and time, whether learning is effective or gets stuck, and a sense of distance and direction in choice. Some steps are short and safe; some are long and revealing.
The open frontier is institutional. The abstract models apply broadly and lack institutional context. How do markets, politics and organizations shape correlated learning, and how should they be designed to improve decisions? Embedding the framework in concrete institutions is the natural next step, with work under way.
Politics supplies the most counterintuitive example. Policy choice is structured by legislatures, bureaucracies and elections, where decision-making power is separated from authority and decisions are often made by groups. Work embedding the framework in a model of elections finds that party competition leads in the short run to policies less attractive to voters, and in the long run to better outcomes, as power switching induces more search that finds better policies. Competitive elections keep search going where it would otherwise get stuck. Turnover looks messy and works as exploration.
The parallel with today's technology is close. Artificial intelligence is correlated learning on a grand scale. Each experiment with a model architecture, training method or application reveals information about nearby alternatives. Firms and researchers decide how far from known approaches to venture, balancing the reliability of incremental steps against the potential of bold departures. The framework analyzes exactly these choices.
Getting stuck is the danger to watch. Decision-makers who explore only near what they know may converge on solutions that are good and far from the best. Institutions encouraging continued search, such as competition, keep learning moving. In AI, competition among firms and research groups may play this role, pushing exploration into new areas and preventing premature convergence on dominant approaches. A field where everyone tweaks the same recipe may be climbing a small hill with great precision.
Deployment is correlated learning too. Organizations adopting AI learn from each application about related ones; early successes or failures shape beliefs about adjacent uses. Understanding how such learning spills over across firms and sectors can inform policies that speed diffusion of productive technologies. One firm's pilot is every competitor's data point.
Regulators face the same structure. Deciding how to experiment with rules for a technology like AI, each choice reveals information about similar choices. Institutions allowing experimentation, comparison and adjustment, while preserving the ability to change course, suit environments where learning is correlated and outcomes uncertain. Rules frozen early stop the search before the map is drawn.
Evidence keeps the agenda grounded. Theory about correlated learning gains value when tested against behavior, and the review highlights emerging evidence and the methods needed to deepen it.
Pulled together, the structure of knowledge shapes how decisions should be made. When outcomes are correlated, each choice is an experiment, and the institutions governing exploration determine whether learning succeeds. The survey offers a unified framework for search, communication, innovation and attribute choice, a timely guide for a world reshaped by AI, where exploring the unknown is the central activity.
A hard question for AI strategy: the biggest risk may be a field that converges too fast, ahead of any single failed experiment. Competition and turnover, often treated as wasteful, may be the mechanisms that keep the search alive.
Introduction
We survey a growing literature on correlated learning—that is, how information from one choice changes beliefs about others when outcomes are correlated. The core modeling innovation is to represent the unknown mapping from choices to outcomes as the realized path of a stochastic process, most commonly the Brownian motion. We show how the framework has been applied to four canonical economic problems in which correlated learning is key but understudied: ( a ) search and experimentation, ( b ) communication, ( c ) innovation and market competition, and ( d ) attribute problems. We review emerging empirical and experimental evidence on correlated learning and outline new theoretical and methodological directions.
Takeaway
Ongoing work has advanced the framework in several directions, but much remains to be done. We emphasize opportunities that remain open.1
5.1. Applications There is broad scope to delve further into applications with the new approach. Often this requires reinterpreting institutions and practices through the lens of correlated learning. The most progress has come from questions the framework is built to address: informational spillovers across players and time, whether learning is effective or gets stuck, and a sense of distance and direction in choice. 5.1.1. Deepening. The abstract models of the previous sections have broad applicability but lack institutional context. How do markets, politics, and organizations shape correlated learning, and how should they be designed to improve decisions? Embedding the framework in concrete institutions is a natural next step, with work already under way. 5.1.1.1. Political institutions. In politics, policy choice is structured by institutions, whether within legislatures, by the bureaucracy, or in elections. This adds context to correlated learning in two ways: (a) Decision-making power (the policymaking) is separated from authority (the voters), and (b) decisions are made by a group (e.g., a legislature) rather than an individual (i.e., collective choice). Papers in this vein have focused on how these features affect decision-making in complex environments when learning is correlated. Callander (2011b) explores the first of the two questions above by embedding the workhorse framework into a model of elections. Competition between parties leads in the short run to policies that are less attractive to voters; in the long run, however, it leads to better outcomes for voters, as power switching induces more search that finds better policies.
In particular, competitive elections allow search to keep going where it would otherwise be stuck. Callander & Hummel (2014) address the same question by shifting attention to parties’ incentives: How might a party choose policy when it knows power will switch to the other party in the future? They show that this creates a purely informational intertemporal linkage: The incumbent preemptively experiments with policy to shape the informational environment of its successor. Callander & McCarty (2024) turn to the second of the two questions above in the context of legislative policymaking. They embed the workhorse Brownian framework into the canonical model of agenda setting.
Notes
- Arjada Bardhi & Steven Callander, Learning in a Correlated World, 18 Ann. Rev. Econ. 391 (2026), https://doi.org/10.1146/annurev-economics-051624-072515. Takeaway source: 5. What Next?, reproduced from the openly licensed version (CC BY-NC-ND 4.0). ↩
Cite as: Arjada Bardhi & Steven Callander, Learning in a Correlated World, 18 Ann. Rev. Econ. 391 (2026), https://doi.org/10.1146/annurev-economics-051624-072515
Header image: AVAST!! (4944177444).jpg by Ian Sane from Oregon, USA, CC BY 2.0, via Wikimedia Commons.
.jpg){kind=link}