Faces, Firms & The Future of Economics: MLESI 2025 Highlights

In a room packed with economists and researchers, a mugshot appeared on the screen. The question wasn’t about guilt or innocence—it was what an algorithm perceived in the photo that humans could not. This was one of the many engaging conversations at this year’s Machine Learning in Economics Summer Institute (MLESI), where the future of economics was reimagined through data, algorithms, and debates that challenge our understanding of the world—with some of the brightest students and researchers paving the way.

As the Institute celebrated its 5th year, it once again established itself as an environment provoking dialogue about data-driven discovery. Organizers from top institutions put together a week consisting of more than 25 speaker presentations, a two-day research conference, and convened over 40 graduate students representing 15 universities across the world to push boundaries at the intersection of AI and economics. 

How Machines Learn, and What They Miss

The phrase ‘There ain’t no such thing as a free lunch’ is a familiar saying with a rich history. Its 19th century origins trace back to when bar and saloon owners would offer “free” food to patrons who bought alcoholic drinks. The food was never truly free though; the appearance of a "free lunch" was a way to incentivize buying drinks, a more profitable purchase.

To MIT Assistant Professor of Economics and conference co-organizer Ashesh Rambachan, the idea directly applies to the field of Machine Learning. The same way “free lunch” is not actually a shortcut to saving money, a single algorithm is not a shortcut to solving every problem. Each algorithm is tailored to a specific set of input and tasks, meaning that a learning algorithm perfectly attuned to execute one task is inherently not well suited for completely unrelated tasks. 

But tailoring an algorithm’s inductive bias—the assumptions it uses to generalize training data to new data—to the structure of a specific problem can reveal shared characteristics among tasks. The concept is like opening a toolbox: while a hammer is perfect for hammering a nail in the wall, it is useless when trying to turn a screw. Although literature about AI often makes sweeping guarantees, Rambachan reminded students at MLESI that in prediction policy, choosing the right model is essential. For data scientists and economists, a careful consideration and understanding of which algorithm to use is just as important as the algorithm itself for success. 

Taking Things at Face Value, Literally

As the discussion turned to real-world examples of prediction policy in action, the focus shifted to the courtroom. MIT Professor and Chicago Booth Distinguished Service Fellow Sendhil Mullainathan presented his work using mugshot data, demonstrating the capabilities of AI’s predictive power. He showed how algorithms can detect the nuanced facial features that humans—often unconsciously—use to form judgements. The algorithm went further still, accurately predicting a judge’s decision to detain or release a defendant and doing so at surprisingly large magnitudes. 

This research had implications for how personal bias might creep into our perception of faces, with real-consequences when applied in high-stakes decisions. However, the scope goes beyond just faces and the justice system. The scientific process advances not just when theories are confirmed, but when anomalies are found in patterns within high-dimensional data (e.g. cell phones, satellites, online behavior, news headlines, corporate filings, high-frequency time series, etc.). In Mullainathan’s case, the data was facial features, and it uncovered behavioral patterns that humans cannot easily detect. He showed students how algorithms can act as “hypothesis generators,” and offer new insights that push economists to rethink theories and refine models. 

At MLESI, Mullainathan pulled threads from his paper, Machine Learning as a Tool for Hypothesis Generation, co-authored by Edwin A. and Betty L. Bergman Distinguished Service Professor Jens Ludwig. The two aimed to automate the process of scientific advancement. They designed an algorithm and fed it a dataset of human faces and theory of human perception, then asked it to generate anomalies and highlight holes in that theory. 

The algorithm functions almost like an argument in court: the anomaly generator produces potential inconsistencies, and the defendant presents plausible explanations, causing the algorithm to refine itself as it goes. The research acted as a tangible example to the graduate students that AI can do much more than make predictions; it is a powerful new tool for pattern recognition and discovery. 

Markets Speak and Machines Listen

Of course, the institute was not entirely about mugshots and courtrooms, this was an economics institute after all! Students found themselves deep in the world of financial markets. The final few speaker sessions provided a fresh perspective on how economic data is analyzed.

Assistant Professor of Finance and Applied AI and Biehler Junior Faculty Fellow Suproteem Sarkar—a recent addition to the Booth community—showed how stock values are not just driven by balance sheets but by how markets talk about firms. By turning financial news into embedding vectors, he revealed how sentiment and perception can be quantified and measured in ways that firms might eventually base decisions on.  

Associate Professor of Marketing Giovanni Compiani followed suit, extending the idea of harnessing unstructured data to consumer products. He emphasized that the images, titles and reviews behind a product reveal consumer perception, and how those perceptions shape demand. In his paper Demand Estimation with Text and Image Data, Compiani established a way to estimate demand by analyzing consumers’ value of hard-to-quantify attributes, like visual design or functional benefits. 

While challenges remain—LLM’s aren’t built to understand words in the right economic context, for example—the message to the graduate students in the room was clear: unstructured data, once considered noise provides a powerful new lens through to capture and analyze markets’ hidden drivers of value and how those values, and the markets themselves, evolve.

The summer institute made it exceedingly clear that machine learning is not just a tool for prediction. It is actively reshaping how economists think about theory and new types of data. By harnessing AI in the right ways, it becomes a major tool in research and discovery, encouraging economic growth.

The week wrapped up with some big questions still on the table: How can machine learning's ability to reveal what humans miss be harnessed without losing sight of fairness, accountability, and human judgment? And how can we push the limits of AI and economics to inspire informed research and real-world impact? What is clear is that the graduate students of today are already chasing those answers. Hosting some of the brightest among them at the Machine Learning in Economics Summer Institute at Chicago Booth was both a privilege and a glimpse into the future of the field.