with Will Jennings, Lunzheng Li, and Zacharias Maniadis, European Journal of Political Economy (Accepted).
with Emre Aytimur and Robert Schwager, Canadian Journal of Economics,
vol. 49, no. 3, pp. 997-1015.
with Jamal Ouenniche and Mohammad Rajabi, Journal of Optimization Theory and Applications, vol. 169, pp. 314–343.
Using the methodology pioneered in previous research, we demonstrate that ChatGPT (GPT-3.5) and GPT-4 are not well-trained on economic topics. We provide some concrete examples of how ChatGPT fails to comply with standard economic principles. We warn the general public against using it for economic advice for the time being.
In a set of preregistered experiments with general population, participants receive recommended answers to an economics questionnaire by two computerized advisors. One advisor is of high-accuracy (`the Expert') and recommends the answers produced by academic consensus. The other advisor is of low accuracy (`the Populist'), and recommends the modal answers of participants from a pilot study. Participants do not know who the Expert is, and have to judge this from the recommendations. We examine which advisor participants identify as the Expert via revealed preference, i.e. participants select an advisor to answer the questionnaire on their behalf. We call this the 'identify the Expert' problem. Our participants overwhelmingly choose the Populist, even when fully informed about the modus operandi of the two advisors. Bayesian models fail to explain these choices, even in the degenerate case where participants should be able to identify the Expert with 100% accuracy. Overconfidence in one's ability only partly accounts for advisor choice, while ego-involvement cannot explain behavior.
How do people choose economic advice? Previous literature suggests that advisers are prone to pandering, offering advice to the public that is wrong, but conforms to commonly held beliefs. We develop a set of validated questions on economic policy to examine the persuasiveness of expert versus populist advice. Populists, in our context, conform to commonly held beliefs, even when wrong. Two computerized advisers suggest answers to each question, and experimental participants are incentivized to choose the most accurate adviser. Do participants choose the high-accuracy adviser ('the Expert'), or the low-accuracy one ('the Charlatan'), whose answers are designed to be similar to the modal participant's? Our participants overwhelmingly choose the Charlatan. Revealing the Charlatan's exact modus operandi to debias behavior scarcely helps. Sequential feedback on the correct answer improves choices only slowly and partially. Bayesian choice models fail to explain behavior; a naive choice model akin to reinforcement learning with high inertia fits behavior well.
I develop a simple, static general equilibrium model with two classes of individuals, workers and entrepreneurs, and two goods. One good is in fixed supply and the other is a standard, producible and consumable commodity. Entrepreneurs set prices, wages and job positions, and competition in the labor market is modelled à la Bertrand. Even though the model does not feature any type of price rigidity, asymmetric information or labor-market friction, its pure symmetric Nash equilibria produce markedly different results from the canonical competitive equilibrium models: (i) A positive output gap and unemployment may emerge in equilibrium. (ii) Income and wealth inequality matter for the determination of equilibrium prices and employment. (iii) An increase in income/wealth inequality or of productivity may reduce employment and increase the output gap. As a result, Say's Law may not hold in the economy and minimum-wage policies may have desirable effects in terms of employment and output.
We provide a simple model that relates the search intensity of households for products to the price distribution and the wage. Households decide how much time to spend on work and on search for finding better deals in a market where firms charge different prices. Thus, the equilibrium price distribution and the wage depend on the endogenous search intensity and labor supply. Moreover, we show that positive technological shocks, which reduce price posting costs for firms, lead to an increase in labor supply, wages and to the average price, while they lead to a decrease in search effort. These results are consistent with recent empirical findings.
"Science versus Populism: Why and When Do Charlatans Beat Experts?"
"ChatGPT and Economic Advice: A Cautious Warning and a Few Pitfalls"
"How Does Political Uncertainty Affect the Optimal Degree of Policy Divergence?"