AI picks up gender and age stereotypes found online

Embargoed until: Publicly released: 2025-10-09 02:00

International

PHOTO: Fotos on Unsplash

Online images tend to represent women as younger than men in the same job, meaning that AI trained on these images learns the same stereotypes. Researchers analysed ages and genders of people in almost 1.4 million online images and videos, and found this stereotype was especially clear in high status jobs like doctors and bankers, despite the US census showing no age differences between genders in these roles. When the team prompted ChatGPT to create and rate around 35,000 resumes, those it generated for women showed them as as slightly younger and less experienced, while it rated older men as better candidates. The researchers say future research should look at how algorithms amplify such "distortions of social reality", and how this can be corrected to prevent further inequalities.

Media release

From: Springer Nature

Social sciences: Distorted online age and gender representations (N&V)

An analysis of 1.4 million online images suggesting that women are represented as younger than men across occupations and social roles is published in Nature this week. These stereotypes may be further reinforced by mainstream algorithms, revealing new challenges in the fight against inequality.

Previous research has found that social stereotypes represented online can bias our real-world perceptions. With the increasing presence of large language models in the running of the online ecosystem, there are concerns that such biases could be amplified without critique by artificial intelligence.

Douglas Guilbeault and colleagues used a catalogue of nearly 1.4 million images from five popular online platforms (Google, Wikipedia, IMDb, Flickr and YouTube) to analyse the average ages of women and men representing different occupations. They found that women were represented as younger than men across occupations and social roles (particularly in jobs with higher status or earnings, such as doctors or bankers, despite there being no systematic differences in the real-world workforce of the United States according to census data. The authors then assessed the presence of this trend amongst large language models (algorithms that are trained on internet data). They prompted Chat GPT to create 40,000 resumes for 54 occupations using 16 unique female and male names. The results showed that ChatGPT presumed the female candidates to be on average 1.6 years younger than their male counterparts. When asked to rate these resumes, ChatGPT rated the older, male candidates as higher quality than the female applicants.

The results illustrate how stereotypes surrounding gender and age can be distorted and perpetuated by online media and large language models, with the potential to disadvantage individuals in those groups. As Ana Macanovic writes in an accompanying News & Views article, these results provide evidence that “biased perceptions of age and gender are not only picked up by AI models but also actively reproduced by them”.

Multimedia

Age and gender on social media

Attachments

Note: Not all attachments are visible to the general public. Research URLs will go live after the embargo ends.

Research Springer Nature, Web page URL will go live after embargo ends.

Journal/
conference: Nature

Research: Link to Paper 1 | Paper 2

Organisation/s: Stanford University, USA

Funder: This project was partially funded by grants from the Fisher Center for Business Analytics and the Center for Equity, Gender & Leadership, awarded to D.G. and S.D., as well as the Barbara and Gerson Bakar Fellowship awarded to D.G., all through the University of California, Berkeley.

Media Contact/s

Contact details are only visible to registered journalists.