Overview

• Analyzing about 1.4 million images and videos plus nine language models, researchers found women are depicted as younger than men in thousands of roles, with the distortion most pronounced in high-status jobs.
• These portrayals contradict U.S. Census data showing men and women in the workforce are roughly the same age, indicating a systematic mismatch between online content and reality.
• A randomized experiment showed that viewing occupation-related images shifts beliefs about typical worker ages, lowering estimates after exposure to women and raising them after exposure to men, with knock-on effects on preferred hiring ages.
• In tests generating and scoring roughly 40,000 resumes, ChatGPT made female candidates 1.6 years younger on average, described them as less experienced, and then ranked those resumes lower than men’s.
• The authors warn of feedback loops as AI learns from biased online data, while OpenAI told Sky News it has studied systemic bias in ChatGPT and found fewer than 1% of responses reflected harmful stereotypes.