In a research interview, having participants generate AI images on the spot to express their ideas—supposedly more intuitive than describing them—sounds clever on paper. But after actually running a round, you realize the pitfalls. First, participants can’t write prompts, so the images end up reflecting the model’s biases, not what’s in their head.
You think you’re capturing their needs, but really you’re just getting SD’s default aesthetic. Second, generation has lag and randomness, so people tweak their intentions to fit the model, and the interview gets hijacked by the tool.
If I have to use it, I usually let the participant describe verbally first, then I help adjust the prompts on the side, treating image generation as a confirmation tool, not a collection tool. It works as a validation method, but you gotta be super careful using it as primary data source.