Feminism in AI-generated visuals: codebook + dataset

Dataset of 2,400 AI-generated visuals + codebook used to code feminism in these visuals 

We designed eight prompts, divided into four types:

Type 1: underspecified prompts (i.e. ‘feminism,’ ‘feminisms,’ ‘feminists’), in order to avoid influencing the results and let the AI models reveal their representations. Two prompts were in the plural to implicitly stimulate more diversity than with the prompt in the singular.

Type 2: ‘féminisme’ (in French) and ‘feminismo’ (in Spanish) to uncover potential language-based differences.

Type 3: two more specific prompts (i.e ‘feminism diversity’ and ‘feminism intersectionality’) to invite the AI models to specifically design visuals that reveal some inclusiveness.

Type 4: ‘photorealistic feminism,’ a popular visual genre among AI-generated visuals, to uncover the potential impact of the style on the representations provided.

We generated 100 visuals per prompt for each of the three selected AI models - i.e DALL·E 3, Midjourney and Stable Diffusion in the summer 2024. In total, we thus generated 2,400 visuals.

Images were generated using Stable Diffusion XL (via stablediffusionweb.com, using default guidance scale of 7 and 1:1 aspect ratio), DALL·E 3 (accessed through ChatGPT 4.0 using temporary chat sessions to minimize personalization effects), and Midjourney v5.2 (via Discord interface).