John J. Conlon

It is widely feared that sycophantic AI advice, by validating users' perspectives, pushes people toward more extreme versions of decisions they were already inclined to make. In an experiment with 1,500 participants across 30 environments, we show that AI advice, despite being sycophantic, narrows the gap between opposite-leaning participants. This convergence occurs across domains and contrary to expert predictions. Experimentally increasing sycophancy weakens convergence, but we find that market forces are not pushing toward greater sycophancy. Our results are consistent with users extracting information from AI advice, adjusting for its sycophancy on average while reacting to it on the margin.

Talk to this paper

We study two hypotheses for how a human role in production could persist despite growing AI capabilities. Humans might have a durable advantage in taste, designing better products even as AI agents automate routine tasks, or consumers might intrinsically prefer that humans have a hand in production. In our experiment, participants act as producers, designing computer games with an AI coding agent, and as consumers who buy them. Across three studies (N = 1,025), we find no evidence for either hypothesis. Games designed by an AI on its own outperform those of human producers, including both amateurs and professional creatives. Human producers do not anticipate the AI's advantage. Revealing whether a game is human-designed does not raise consumers' willingness to pay for the product.

Talk to this paper

Models of motivated cognition assume that agents selectively direct their attention to improve their mood, confidence, or anticipatory utility, and yet the modern attention economy appears dominated by negativity. We study this tension in a controlled experiment. Participants take a selfie and then work on a real-effort task while a feed of comments about their appearance, resembling those on social media, are posted alongside their task. Ex ante, participants overwhelmingly prefer this feed to be populated primarily by positive, complimentary comments. In the moment, however, this preference reverses: when exposed to a feed composed of both positive and negative comments, they engage significantly more with the negative content. Consistent with participants knowing their consumption of negative comments is compulsive, a third of participants are willing to pay a strictly positive price for a positive rather than a negative feed, and they ex post regret reading negative comments more than positive comments. This asymmetry between ex ante and in-the-moment attention preferences is larger among women, who report caring more about their appearance, consistent with concerns that social media usage may be particularly harmful for young women. Finally, we discuss implications for the incentives platforms face toward negative content.

Talk to this paper

Beliefs often depend on contextual cues that influence what comes to mind. Models of associative sampling capturing this idea give a central role to similarity, which is typically unobserved and in principle very high-dimensional. We show theoretically that, despite this potential complexity, the effects of cues depend only on two sufficient statistics—representativeness and resonance—that summarize how similarity distorts beliefs. Further, these statistics are measurable with simple, pairwise similarity judgments, offering a straightforward method of generating and testing quantitative predictions. In a controlled experiment, we deploy this method and show that belief responses to normatively irrelevant cues closely track model predictions. We also show that large language models can approximate human similarity judgments, helping to extend this approach to field settings. Applying the framework to survey data on personal experiences and macroeconomic expectations, we show that similarity-based predictions quantitatively explain much of the heterogeneity in experience effects across a wide range of experiences and expectation domains.

Talk to this paper

Do spouses pool useful information and learn from each other when they have incentives to do so? In an experiment with married couples in India, we vary whether individuals discover information themselves or must instead learn via a discussion about what their spouse discovered. Women treat their own and their husband’s information the same. In contrast, men respond half as much to information discovered by their wife, even when it is perfectly communicated. When paired with strangers, both men and women heavily discount their partner’s information relative to their own. We thus provide evidence of a gender difference in social learning (only) in the household.

Talk to this paper

We investigate the role of information frictions in the US labor market using a new nationally representative panel dataset on individuals' labor market expectations and realizations. We find that expectations about future job offers are, on average, highly predictive of actual outcomes. Despite their predictive power, however, deviations of ex post realizations from ex ante expectations are often sizable. The panel aspect of the data allows us to study how individuals update their labor market expectations in response to such shocks. We find a strong response: an individual who receives a job offer one dollar above her expectation subsequently adjusts her expectations upward by $0.47. The updating patterns we document are, on the whole, inconsistent with Bayesian updating. We embed the empirical evidence on expectations and learning into a model of search on- and off- the job with learning, and show that it is far better able to fit the data on reservation wages relative to a model that assumes complete information. The estimated model indicates that workers would have lower employment transition responses to changes in the value of unemployment through higher unemployment benefits than in a complete information model, suggesting that assuming workers have complete information can bias estimates of the predictions of government interventions. We use the framework to gauge the welfare costs of information frictions which arise because individuals make uninformed job acceptance decisions and find that the costs due to information frictions are sizable, but are largely mitigated by the presence of learning.

Talk to this paper

We rely on memory to form beliefs, but we also frequently revisit memories in conversation and private reflection. I show experimentally that such rehearsal generates systematic belief biases. Participants are randomized to have conversations about a subset of experiences, either ones that reflect well or poorly on them, which has large effects on which experiences they can recall a week later. Crucially, participants appear naive about these effects: they take what they remember at face value when later incentivized to form accurate beliefs about their experiences. Participants also make rehearsal choices seemingly without regard to their later distortionary effects. Intrinsic preferences for thinking about positive experiences instead drive rehearsal choices and therefore a positivity bias in recall and beliefs. This mechanism provides a new non-strategic channel through which overly optimistic beliefs can arise and generates novel predictions in a range of economic domains.

Older (longer) version here.

Talk to this paper

Economists often estimate preferences by looking at how demand for an option changes as its attributes change or as beliefs about them change in response to information. I show in a multi-attribute induced-preference experiment that this method can be substantially biased due to differential inattention. Attention is systematically distorted at baseline, with participants reacting much more to some attributes than others. Further, attention is highly malleable: information telling participants about the value of a desirable attribute—even when the information is already known, transparently redundant, and explicitly randomly assigned—greatly increases attention to the attribute it describes and distracts from other attributes. It also boosts average demand for the good, implying that inattention takes the form of neglect rather than shrinkage toward a prior. These results combine to imply that information experiments with passive control groups can be greatly misleading about preferences and can even have the wrong sign: in my experiment, reducing overoptimism about an attribute nonetheless boosts demand for its associated good. I show, both theoretically and experimentally, that active control-group experiments yield preference estimates with the correct sign but biased magnitudes. Finally I show that parallel active control-group experiments (PACEs) can solve this problem: PACEs yield unbiased estimates of preferences under an equal induced-attention assumption, which is testable (and satisfied) in my experiment.

Note: This replaces a previous draft titled "Information, Attention, and Persuasion."

Talk to this paper

We study social learning using experiments where two people independently learn relevant information and can share their information to make accurate private decisions. Across three experiments, people are substantially less sensitive to information others discover than to equally-relevant information they discovered themselves. This holds when they must learn information from others through discussion; when the experimenter perfectly communicates the information; and even when participants observe others’ information with their own eyes. Thus, our results stem not from a failure to elicit information from others but a systematic tendency to underweight it relative to one’s own information. Our findings illustrate a powerful barrier to social learning that might underlie many documented cases of failure to learn from others.

Talk to this paper

Econimate video summary

We test for stereotyping—the exaggeration of representative traits—in a high-stakes economic environment. Using surveys administered among undergraduates at the Ohio State University as well as large-scale nationally representative data, we measure how U.S. first-year students perceive the relationship between college majors and occupations. We show that students greatly overestimate the likelihood that majors lead to their representative jobs (e.g., counselor for psychology, journalist for journalism). Using an implicit association test, we show that students associate majors with their representative careers and that these associations strongly predict belief biases, in line with a stereotyping mechanism. A simple equilibrium model of the labor market predicts that stereotyping reduces welfare by increasing misallocation, which we corroborate with correlational evidence on job/major mismatch. In a field experiment, we test a light-touch policy to reduce stereotyping and find significant effects on students' intentions about what to study as well as the classes and majors they enroll in.

Talk to this paper

For standard statistical problems, we provide new evidence documenting (1) multimodality and (2) instability in probability estimates, including from irrelevant changes in problem description. The evidence motivates a model in which, when solving a problem, people represent each hypothesis by attending to its salient features while neglecting other, potentially more relevant, ones. Only the statistics associated with salient features are used. The model unifies biases in judgments about i.i.d. draws, such as the Gambler's Fallacy and insensitivity to sample size, with biases in inference such as under- and overreaction and insensitivity to the weight of evidence. The model makes predictions for how changes in the salience of specific features jointly shapes known biases and measured attention to features, but also create entirely new biases. We test and confirm these predictions experimentally. Salience-driven attention to features emerges as a unifying framework for biases conventionally explained using a variety of stable heuristics or distortions of Bayes' rule.

Talk to this paper

There are teacher shortages in the U.S. and around the world. In a three-year field experiment with a large teacher placement program, Teach For America (TFA), Coffman, Conlon, Featherstone, and Kessler (2019) finds that providing upfront liquidity to prospective teachers in financial need dramatically increases the rate at which they start teaching through TFA. In this paper, we combine TFA administrative data, survey data, and publicly available data (e.g., LinkedIn profiles) to extend those results. We follow individuals for a few years post treatment and find that providing upfront liquidity not only increases the rate that financially constrained individuals join TFA but also increases the rate that they complete the full two years of teaching. Further, providing liquidity to those who need it increases their likelihood of being teachers at all—not just through TFA—through at least two years.

Talk to this paper

In many economic decisions, people estimate probabilities, such as the likelihood that a risk materializes or that a job applicant will be a productive employee, by retrieving experiences from memory. We model this process based on two established regularities of selective recall: similarity and interference. We show that the similarity structure of a hypothesis and the way it is described (not just its objective probability) shape the recall of experiences and thus probability assessments. The model accounts for and reconciles a variety of empirical findings, such as overestimation of unlikely events when these are cued versus neglect of noncued ones, the availability heuristic, the representativeness heuristic, conjunction and disjunction fallacies, and over- versus underreaction to information in different situations. The model yields several new predictions, for which we find strong experimental support.

Talk to this paper

In a field experiment at a flagship state university in the United States, I test whether providing college students salary information can affect their choices of major and classes. I find that undergraduates are substantially misinformed about mean salaries by major. On average, students in my sample underestimate mean salaries, but there is also large heterogeneity in beliefs across individuals. I also find that providing information to correct these errors has a large impact on students' choices; students in the treatment group were nine percentage points (16 percent) more likely to major in a field about which they received information.

Talk to this paper

Can access to a few hundred dollars of liquidity affect the career choice of a recent college graduate? In a three-year field experiment with Teach For America (TFA), a prestigious teacher placement program, we randomly increase the financial packages offered to nearly 7,300 potential teachers who requested support for the transition into teaching. The first two years of the experiment reveal that although most applicants do not respond to a marginal $600 of grants or loans, those in the worst financial position respond by joining TFA at higher rates. We continue the experiment into the third year and self-replicate our results. For the highest-need applicants, an extra $600 in loans, $600 in grants, and $1,200 in grants increase the likelihood of joining TFA by 12.2, 11.4, and 17.1 percentage points (or 20.0%, 18.7%, and 28.1%), respectively. Additional grant and loan dollars are equally effective, suggesting a liquidity mechanism. A follow-up survey bolsters the liquidity story and also shows that those drawn into teaching would have otherwise worked in private-sector firms.

Talk to this paper

We study behavior in an information cascades setting where previous buyers of a product leave noisy but informative ratings about its true quality. Although this increases the amount of public information available, Yoon (2015) shows that ratings can hurt welfare in some settings. In particular, the added public information causes cascades to form earlier, but if agents stop purchasing then there are no new ratings to break the inefficient cascade behavior. Although we find some evidence roughly in line with the theory, those results are swamped by a strong tendency for subjects to purchase even when public information suggests they should not.

Talk to this paper

I argue that there could be observational evidence for absolute motion. My argument employs a spatial analogue of Shoemaker's classic ‘global freeze’ thought experiment, wherein he imagines a world whose inhabitants would seem to have strong evidence that a period of time without change had occurred. The inhabitants of my own imagined world, I claim, would have strong evidence that everything in the universe had moved the same distance in the same direction and, thus, strong evidence for absolute motion. This conclusion seems to pose a problem for spatial relationists insofar as they cannot account for absolute motion.

Talk to this paper