AI 이미지 생성기의 편견과 고정관념

Text-to-image models such as Stable Diffusion, Midjourney and DALL-E learn the link between words and pictures from huge datasets of image-caption pairs, largely scraped from the web; Stable Diffusion, for example, was trained on subsets of LAION-5B. The web over-represents some countries, languages, body types and occupations, and captions reflect the assumptions of whoever wrote them. The model then learns the most typical picture for a word and tends to amplify it, because generation favors high-probability, 'average' outputs. A prompt that says nothing about gender or ethnicity still produces a default, and that default is often the stereotype. Researchers measure this in a few ways. A common method is to generate hundreds of images for neutral prompts (occupations, adjectives like 'attractive' or 'poor', nationalities) and label perceived gender, skin tone or age, sometimes compared against labor statistics. Hugging Face researchers' 'Stable Bias' project (2023) used clustering of model outputs rather than assigning labels directly, partly because labeling someone's race from a picture is itself fraught. A 2023 Bloomberg analysis of thousands of Stable Diffusion images found that high-paying jobs skewed toward lighter-skinned men and low-paying jobs toward darker skin tones. Fixes are harder than they look. In 2022 OpenAI said it applied a technique so DALL-E 2 produced more diverse people when prompts did not specify, widely reported to involve adding terms to prompts behind the scenes. In February 2024 Google paused Gemini's generation of images of people after it produced historically inaccurate images, such as racially diverse depictions of groups that were historically homogeneous. The lesson is that bias is not a single dial: a system must be representative where the prompt is open and accurate where the prompt is specific. A common misconception is that bias only comes from the training data; model design, filtering, captioning and post-training choices all shape the result too.