Extended Data Fig. 6: Schematic representation of the method applied for the resume generation phase of the ChatGPT audit experiment.
From: Age and gender distortion in online media and large language models
ChatGPT was prompted to generate resumes for 54 occupations in each of three conditions: (i) the Control condition, (ii) the Control-gender condition, and (iii) the Treatment condition. In the Control condition, ChatGPT generated 50 resumes for each occupation without specifying the name or gender of the applicant, yielding 2,700 resumes. In the Control-gender condition, the design was identical to the Control condition except that ChatGPT was also prompted to specify the gender of applicants in the resumes it generated. In the Treatment condition, the design of the Control condition was replicated except that ChatGPT was prompted to generate a resume for a named applicant whose name was selected from a list of 16 male and 16 female names spanning four ethnicities and normed for popularity, familiarity, and perceived age group. In the Treatment condition, ChatGPT was prompted 20 separate times for each name-gender-occupation prompt combination, yielding 34,560 resumes in total and 17,280 resumes for each gender group.
