The Internet's Age and Gender Stereotypes: How Online Data Shapes AI Bias and Hiring

Nature investigates how online data encode age and gender stereotypes, revealing that women are often depicted as younger than men across images and text. By analyzing vast datasets from IMDb and Google Images, the study shows these biases permeate AI training, potentially affecting resume screening and hiring decisions. The video also highlights the gap between online stereotypes and real-world data, and calls for designing AI that better calibrates our understanding of society to promote equity.

Introduction

The Nature video examines how the Internet may distort our understanding of age and gender, and how this distortion feeds into artificial intelligence and real-world outcomes. The host interviews Douglas Guilbeault, lead author of the study, to unpack how online data reflect and reinforce stereotypes about who is older or younger, and who should look a certain way for various jobs.

Age Bias in Online Images

Using data from IMDb, the Internet Movie Database, and large image databases like Google Images, the researchers cross referenced 450,000 celebrity photos with real information about those people, including their real ages and self-identified genders, plus the timestamp when the image was taken. The results show that female celebrities were, on average, 6.4 years younger than male celebrities in their photos. In Google Images, the distribution of ages shows women peaking around 25 while men peak around 44, illustrating a strong online age bias for women and older men among influential figures.

Textual Bias Across Occupations

Beyond images, the team analyzed text data from the Internet, including language used to describe social categories like occupations and relationships. They looked at pronouns and descriptors commonly associated with youth and age, finding that categories tied to women were more often described with younger-age terms, while male-associated categories carried older-age descriptors. This pattern held across thousands of job-related terms, indicating a pervasive stereotype about what women and men in certain roles should look like.

Reality Check: Data vs. Reality

The researchers compared online stereotypes with U.S. Census data and life expectancy. Census data show little to no real age gap between male and female workers in many fields, and women live longer on average, meaning they are often older in real life. This discrepancy suggests online data and AI models trained on them may misrepresent the world, reinforcing myths rather than reflecting reality.

Implications for Artificial Intelligence

The study demonstrates that the textual data used to train large language models like ChatGPT inherit and propagate these stereotypes. In a separate experiment, the researchers showed that ChatGPT tends to rank resumes in ways that advantage older men, while younger women can receive comparatively higher scores. This self-fulfilling prophecy can influence hiring and promotion patterns, contributing to real-world wage gaps and glass-ceiling dynamics that are not fully explained by market factors alone.

Addressing the Bias

The video emphasizes the need to design AI systems that calibrate their models of society more accurately, reducing statistical biases and promoting equity. It calls for robust editorial standards, cross-media linking, and AI tools that help users explore content with a more trustworthy representation of demographics and roles, rather than perpetuating stereotypes.

Conclusion

Age and gender stereotypes are deeply entrenched in online data and AI training material. The Nature findings show these biases are not just about perception; they have tangible consequences for hiring and wages. The video advocates for intentional design choices in AI to counter stereotypes and support a more equitable and accurate understanding of the social world.