Prueba de sesgos y estereotipos en los generadores de imágenes de IA

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. Why does a prompt that does not mention gender or ethnicity still produce a biased image?
   • The model randomly assigns attributes with equal probability
   • The model falls back on the most typical picture it learned for the word, which reflects skewed data
   • Users secretly add hidden terms
   • Image generators cannot draw people accurately
2. What kind of data do models like Stable Diffusion learn from?
   • Hand-drawn images commissioned for training
   • Government census photographs
   • Large collections of image-caption pairs scraped from the web
   • Only stock photos with verified demographics
3. What is a common method researchers use to measure bias in image generators?
   • Asking the model to rate its own fairness
   • Counting how many words are in each prompt
   • Checking image file sizes
   • Generating many images for neutral prompts and comparing who appears
4. Why did the Stable Bias project use clustering instead of directly labeling race?
   • Labeling someone's race from an image is itself unreliable and fraught
   • Clustering is required by law
   • Race labels made the images too large
   • The images contained no people
5. What did a 2023 Bloomberg analysis of Stable Diffusion images find?
   • Outputs were perfectly balanced across all jobs
   • Only landscapes showed bias
   • High-paying jobs skewed toward lighter-skinned men
   • The model refused to draw people at work
6. What happened with Google's Gemini image generation in February 2024?
   • It was paused after producing historically inaccurate diverse depictions
   • It was praised for perfect historical accuracy
   • It was banned in all countries
   • It stopped generating landscapes
7. What is the main weakness of simple behind-the-scenes prompt rewriting as a fix?
   • It is too expensive to run
   • It can be context-blind and alter prompts that specify a period or identity
   • It makes images black and white
   • It deletes the training data
8. Which statement about the source of bias is accurate according to the guide?
   • Bias comes only from training data
   • Bias comes only from users' prompts
   • Bias is fully solved by larger models
   • Data, filtering, captioning, model design and post-training choices all shape bias