Haas News
Table of Contents
A UC Berkeley Haas study in Nature, analyzing 1.4 million images and videos plus nine large language models, finds women are consistently portrayed as younger than men across online platforms—a distortion strongest in high-status jobs and amplified by AI in ways that could widen real labor-market gaps.
Even though U.S. Census data shows no systematic age differences between men and women in the workforce, that’s not what you’ll see if you search Google or YouTube or query an AI model like ChatGPT.
A new study in Nature of 1.4 million images and videos, plus nine large language models, shows striking, pervasive age and gender distortion across online images and text, where women are consistently presented as younger than men in nearly 3,500 occupational and social categories. This distortion is strongest in high-status, high-earning occupations, such as CEO or astronaut.
“These misrepresentations feed directly into the real world in ways that could be widening gaps in the labor market and skewing the ways we associate gender with authority and power,” says Assistant Professor Solène Delecourt, who co-authored the study with colleagues from Stanford and the University of Oxford.
Online algorithms were found to amplify this age-gender bias. In one experiment, researchers prompted ChatGPT (GPT-4o mini) to generate and evaluate nearly 40,000 resumes across 54 occupations using distinctively male and female names. ChatGPT assumed the women were younger and less experienced and rated older male applicants more highly for the same jobs.
The study reveals a deeply inaccurate picture of our world, one fed by the online information people increasingly consult to learn about society. When applied in real life, these AI tools are likely to reshape the world even more in line with the stereotypes inherent in their training. Delecourt calls this “a culturewide, statistical distortion of reality.”
In the case of resume screening—where AI is already widely used—AI biases are directly skewing its perceptions of who is and is not qualified for a given job.
“To fight pervasive cultural inequalities, the first step is to recognize how stereotypes are coded into our culture, our algorithms, and our own minds,” Delecourt says.
Yes. A UC Berkeley Haas study published in Nature analyzed 1.4 million images and videos and nine large language models and found women are consistently presented as younger than men across nearly 3,500 occupational and social categories. The distortion is strongest in high-status, high-earning occupations such as CEO or astronaut, even though U.S. Census data shows no systematic age difference between men and women in the workforce.
In one experiment, researchers prompted ChatGPT (GPT-4o mini) to generate and evaluate nearly 40,000 resumes across 54 occupations using distinctively male and female names. The model assumed the women were younger and less experienced, and rated older male applicants more highly for the same jobs—directly skewing perceptions of who is qualified in tasks like resume screening, where AI is already widely used.
Co-author Solène Delecourt calls it a culture-wide statistical distortion of reality. Because people increasingly rely on online information and AI tools to learn about society, these misrepresentations can feed back into the real world—widening labor-market gaps and reinforcing stereotypes that associate gender with authority and power.
Posted in:
Topics:
Traffic Signals
A UC Berkeley Haas and Stanford GSB study of more than 1,000 major companies finds that website traffic can forecast revenue and stock prices months before earnings—more accurately than traditional metrics alone. Wall Street, the researchers say, is leaving money on the table.
Talk Is Deep
A UC Berkeley Haas study forthcoming in Nature Communications finds that disagreement is less contentious and leads to better understanding when discussed by voice or video rather than in writing—even though most people instinctively prefer to hash out conflict over text.
