Fig. 4: Effect of gender and age on ChatGPT’s generation and evaluation of resumes.
From: Age and gender distortion in online media and large language models
a, Partial effect plot in an ordinary least squares regression displaying the effect of a male applicant name (versus a female applicant name) on (1) applicant age; (2) years since the applicant’s graduation; and (3) the number of years of applicant’s relevant experience, while controlling for name and occupation. Only resumes from the treatment condition were examined in this analysis (n = 34,560), because this ensures that all resumes have either a male or female name and were produced through the same prompt. Error bars indicate 95% confidence intervals. b, Linear correlation between applicant age and ChatGPT’s rating of resume quality across all resumes (n = 39,560) from all conditions. Data points display the raw distribution of scores for each resume, with one data point per resume. The trend line reflects a standard bivariate linear trend. c, Partial effect plot displaying the interaction effect between applicant age and applicant gender on ChatGPT’s rating of resume quality, with fixed effects for applicant name, occupation and phase 1 condition (data from the control condition were excluded because of the lack of applicant gender; n = 37,060 resumes used in total). Error bands display 95% confidence intervals. ChatGPT’s temperature was set to its default value of 0.7.
