Online, women are shown as younger than men, even though real working women are not younger than working men. A study in Nature found that gap across images, videos, text and AI tools, and showed it can shift what people believe and whom an AI prefers to hire.
The authors are Douglas Guilbeault of the Graduate School of Business at Stanford University, Solène Delecourt of the Haas School of Business at the University of California, Berkeley, and Bhargav Srinivasa Desikan of the Autonomy Institute in London and the Oxford Internet Institute. The paper was published in Nature on October 8, 2025.
Their starting point was US census data, which show no correlation between the share of women in an occupation and its median age, and no clear difference in the age distribution of working women and men since 2009.
A gap that shows up on every platform
The team analysed nearly 1.4 million images and videos from Google, Wikipedia, IMDb, Flickr and YouTube, covering 3,495 social categories such as doctor or banker. Women were coded as younger than men in every dataset, whether age was judged by human coders, by algorithms or taken from verified birth dates.
With true ages of celebrities, the gap was large. Women were on average 6.5 years younger than men in IMDb images, 3.27 years younger on Wikipedia and 5.35 years younger in Google Images. The most common age for women was in their 20s, while for men it was 40 on IMDb and 50 on Google.
In a YouTube dataset of celebrity videos, 33% of women were classified as young compared with 20% of men. The gap was widest for higher-status and better-paid occupations, where Google Images most often showed men as older than women.
Text told the same story. In GPT-2 Large and other language models trained on internet text, words associated with men were also strongly associated with older age, with a correlation of 0.87 across categories.
Searching and AI tools make it worse
In a pre-registered experiment, 459 people from a nationally representative US sample searched Google Images for occupations. Those who uploaded a picture of a woman estimated the average age in that job 5.46 years lower than those who uploaded a man. Compared with a control group that did not search images of the jobs, seeing a woman lowered estimates by 1.75 years. Jobs seen as more male were also linked to an older ideal hiring age.
The researchers then had ChatGPT (GPT-4o mini) generate nearly 40,000 resumes for 54 occupations in June 2024. For female names it made applicants 1.6 years younger, with graduation dates 1.3 years more recent and 0.92 fewer years of experience. When asked to score the resumes, ChatGPT rated older applicants higher, and the boost from age was larger for men.
The authors acknowledge that the large observational analyses cannot pin down what causes the bias, and that census data with both gender and age are not available at the level of individual occupations, so comparisons were made by industry.
More than 400 million people use ChatGPT weekly, the authors note, which is why they see AI screening of resumes as a direct route for this distortion to reach real hiring decisions.
Study Details:
- Title: Age and gender distortion in online media and large language models
- Authors: Douglas Guilbeault, Solène Delecourt, Bhargav Srinivasa Desikan
- Journal: Nature
- Publication Date: October 8, 2025
- DOI: 10.1038/s41586-025-09581-z