Gaël Le Mens

Latest News

2024-03-16: New publication in PNAS! (another one!!) Frequent winners explain apparent skewness preferences in experience-based decisions.  (with Sebastian Olschewski and Mikhail Spektor). Proceedings of the National Academy of Sciences of the United States of America (PNAS). Data in description-based decisions, investment decisions and choice theories (Prospect Theory) point at a preference for right-skewed options (small prob. of high $). In contrast, experience-based decisions point at an apparent preference for left-skewed distributions. We reconcile this puzzle by examining the effect of frequent winners on experience-based choice. Statistically, left-skewed reward distributions have the higher value in most direct comparisons of reward samples against a right-skewed distribution. We show in experience-based decision experiments that: 

1. People choose left-skewed over matched right-skewed options (while the left-skewed options are the frequent winners),

2. People choose frequently winning options,

3. When making the right-skewed options the frequent winners, people choose the right-skewed options over the left-skewed ones,

4. When both options have the higher value in exactly the same number of reward samples, people choose the right-skewed options.

We reconcile experience-based decisions with the mainstream literature w.r.t. skewness preferences: Pairwise experience-based decisions stress the direct sample comparison and the choices for the frequent winners overshadow skewness preferences in this task.  In the paper, we also present a reinforcement learning model with tallying and skewness preferences to account for all observed behavior, and we discuss the the adaptivity of frequent winner as well as skewness preferences. 

2023 11 30: New publication in PNAS! Uncovering the semantics of concepts using GPT-4  (with Balázs Kovács,  Michael Hannan, & Guillem Pros). We use GPT-4 to create “typicality measures” that quantitatively assess how closely text documents align with a specific concept or category. Unlike previous methods that required extensive training on large text datasets, the GPT-4-based measures achieve state-of-the-art correlation with human judgments without such training. Because training data is not needed, this dramatically reduces the data requirements for obtaining high performing model-based typicality measures. Our analysis spans two domains: judging the typicality of books in literary genres and the typicality of tweets in the Democratic and Republican parties. Our results demonstrate that modern Large Language Models (LLMs) can be used for text analysis in the social sciences beyond simple classification or labelling. The European Research Council (ERC), the Spanish Ministry of Science and Innovation and the ICREA are gratefully acknowledged for funding this research.

2023 11 29: New working paper available on arXiv.org! Scaling Political Texts with ChatGPT (with Aina Gallego). We introduce an innovative methodology for positioning political texts in continuous policy spaces using GPT-4. We validate our approach in two ways. Firstly, using a standard dataset of party manifestos, our analysis reveals that the policy positions derived via GPT-4 exhibit a higher correlation with expert assessments than the positions produced by crowd workers and by other automated text analysis approaches based on word frequency. Secondly, by analyzing tweets from members of the US Congress, we show that the policy positions inferred by GPT-4 show a remarkable correlation (exceeding 90\%) with crowd-sourced estimates. These results demonstrate that GPT-4 is an efficient and valid method to scale text documents in latent political space. The European Research Council (ERC), the Spanish Ministry of Science and Innovation and the ICREA are gratefully acknowledged for funding this research.

2023-11-16: What if there was a 'dislike' button on Twitter? In a paper just published in PLOS One, with my former doctoral students Elizaveta Konovalova and Nikolas Schöll, we show that adding a 'dislike' button could make opinions expressed on social media less extreme. This is because, this would expose social media users to negative quantitative feedback that they do not usually get (because most social media platforms do not have a one-button negative feedback function such as a 'dislike' button). Using simulations of a learning model, we compare two feedback environments that differ in terms of the availability of negative reaction counts. We find that expressed opinions are generally more extreme when negative reaction counts are not available than when they are. We rely on analyses of Twitter data and several online experiments to provide empirical support for key model assumptions and test model predictions. Our findings suggest that a simple design change might limit, under certain conditions, the expression of extreme opinions on social media. The European Research Council (ERC), Fundación BBVA, the Spanish Ministry of Science and Innovation and the ICREA are gratefully acknowledged for funding this research.

2023-09: Cambridge University Press has published Sampling in Judgment and Decision Making, a new book edited by Klaus Fiedler, Peter Juslin and Jerker Denrell. Three chapters feature my work with Jerker Denrell (The Hot Stove Effect), Elizaveta Konovalova (Opinion Homogenization and Polarization: Three Sampling Models) and Balázs Kovács, Judith Avrahami & Yaakov Kareev (The Collective Hot Stove Effect). The European Research Council (ERC), Fundación BBVA and the Spanish Ministry of Science and Innovation are gratefully acknowledged for funding this research.

2023-03-23: How do politicians learn from the feedback they receive from citizens on social media? To learn more about this, check out my new paper in the American Journal of Political Science, with Nikolas Bahati Schöll and Aina Gallego. We used the BERT language model to identify the topic of 1.2M tweets by all elected Spanish politicians (local and national assemblies) over the 2016-19 election cycle and estimated a learning model on this naturally occurring data (think 'bandit problem in the wild'). We find that female politicians are more strongly rewarded by other Twitter users for writing about gender issues than male politicians and that this could contribute to explaining why they write more about this topic. Thanks to European Research Council (ERC), Fundación BBVA and the Ministry of Science, Technology and Innovation of Spain for funding this research and to Guillem Pros Rius for excellent research assistance in putting together the replication material.

2023-03-13: Our Sociological Science method piece on how to use ML classifiers to compute the typicality of texts documents in concepts has now been published. See here for the final version. A follow up that adapts the methods to the use of openAI GPT3 (the language model that powers chatGPT) is coming up soon!

Current Funding

My research focuses on learning by individuals and organizations. Several of my papers explain how individuals might develop and maintain inaccurate beliefs because they rely on the biased samples of information they obtain from their experiences. The information used for learning may be biased because more information is available about alternatives that decision makers believe to be good and thus sample again, whereas less information is available about alternatives that decision makers believe to be poor and thus avoid. In addition, the information to which individuals have access also depends on the opportunities the social context provides them with for learning about alternatives they might otherwise avoid. 

In another line of research, I have developed models of the influence of concepts categories on inference and valuation, which lead to a book on the topic (Concepts and Categories: Foundations for Sociological and Cultural Analysis). 

My current work combines these two streams to analyze how producers of social media content learn from feedback and how learning processes lead to the reinforcement of existing stereotypes. I test the predictions of my models using a variety of methods, such as the analysis of text data using large language models learning (BERT and openAI text embeddings) and a combination of online and laboratory experiments

My research has been published in top scientific journals such as Psychological Review (x3), the Proceedings of the National Academy of Science of the USA (PNAS), Psychological Science (x2), Journal of Personality and Social Psychology, Cognition, Psychonomic Bulletin and Review, Organization Science (x2), Management Science and Administrative Science Quarterly, Sociological Science, American Journal of Political Science. Popular accounts have appeared in the New York Times, the Times (London), WSJ.com, FT.com, USA Today, ABCNews.com, Focus and other in-print and online periodicals. I have taught graduate courses at INSEAD, London Business School, the University of Lugano, the Barcelona Graduate School of Economics and the Barcelona School of Management and have given guest lectures in doctoral courses at Stanford GSB and MIT Sloan School of Management. 

Electronic copies of most of my papers are provided here. I am sharing these for personal and teaching uses only. In no way should these electronic copies be made available on an electronic repository.