Research map / Large language models / AI and society
Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds theme-specific image-caption datasets (India-focused) filtered from Wikipedia-based WIT and trains a transformer text-to-image model (DALL-E-style) on them to show it captures Indian cultural context.
Problem
Generative AI models trained on internet-scale data exhibit Anglo-centric and other social biases, underrepresenting diversity and culturally specific content.
Approach
Construct moderately sized theme-specific image-caption datasets by keyword-filtering Wikipedia-derived WIT pairs (e.g., captions or pages mentioning "India", in English), then train a text-to-image model on them so outputs reflect the target culture.
Key results
- IID contains over 250k image-caption pairs focused on Indian context, with a second dataset IHD focused on Indian architecture.
- Qualitative comparison: for the prompt 'beach', the IID-trained model produced brown-sand beaches typical of India, whereas DALL-E produced white-sand beaches.
- For 'village', the IID-trained model depicted green fields and plain geography, whereas DALL-E produced Eurocentric villages with hut-shaped houses and mountains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the text-to-image model on the India-focused IID dataset | Transformer-based text-to-image model · Indian-context image generation | Indian-context representation | Helped |
| Use the IID-trained model instead of DALL-E | Text-to-image generation · Beach and village prompts | Indian-context representation | Helped |
Surprising findings
- Generated images from the model are blurry, which the author attributes to the computational demands of generative training.
Limitations
- Evaluation is purely qualitative with a few prompts; no quantitative bias metrics or human study.
- Generated images are blurry due to limited compute and training resources.
- Only a single theme (India) and one prompt-based comparison against DALL-E are shown.
Experimental setup
Datasets: WIT (Wikipedia-based Image Text dataset)
Baselines: DALL-E
Builds on: Ramesh et al. text-to-image transformer (DALL-E) Wikipedia-based Image Text dataset (WIT)
Closest papers by meaning
- Would Deep Generative Models Amplify Bias in Future Models? (CVPR 2024 · similarity 0.90)
- More of the Same: Persistent Representational Harms Under Increased Representation (NeurIPS 2025 · similarity 0.88)
- AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data (CVPR 2025 · similarity 0.87)
- SocialStigmaQA: A Benchmark to Uncover Stigma Amplification in Generative Language Models (AAAI 2024 · similarity 0.87)
- Intersectional Unfairness Discovery (ICML 2024 · similarity 0.87)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.