Surveys◆ AI-generated · Sourced
A Brief Overview of Gender Bias in AI
TL;DR
Gender bias is systematically embedded in AI systems—especially large language models (LLMs)—due to skewed training data, annotator subjectivity, and fairness-agnostic objectives, manifesting as occupational stereotyping, pronoun misalignment, and imbalanced text generation; mitigation strategies include debiased data curation (e.g., Hugging Face `datasets` v2.18.0), fairness-aware fine-tuning (e.g., Google’s FairLLM for Llama-2-7b), and specialized evaluation frameworks (e.g., BOLD, WinoBias, Bias in Bios), yet robust cross-model and cross-task fairness remains unachieved.
Origins and Manifestations of Gender Bias
- Bias stems primarily from historical societal imbalances in training corpora (e.g., gender-skewed occupational references in Common Crawl and Wikipedia);
- Annotation introduces subjectivity (e.g., Amazon Mechanical Turk workers disproportionately assign female pronouns to ‘nurse’);
- Model architectures and loss functions lack explicit fairness constraints, leading to persistent gendered errors across mainstream LLMs—including Llama-3-8B, GPT-4, and Gemini 1.5 Pro—with up to 12.7% accuracy gap on WinoBias pronoun coreference tasks.
Key Evaluation Benchmarks and Empirical Findings
- BOLD (Bias in Open-Ended Language Generation Dataset): Measures bias across five dimensions (gender, race, religion, sexual orientation, age); GPT-4 associates ‘CEO’ with male pronouns 3.2× more often than female pronouns in occupational generation;
- WinoBias: Focuses on coreference resolution; Llama-3-8B achieves 41.6% error rate on anti-stereotype examples—far exceeding human baseline (12.3%);
- Bias in Bios: Uses real-world bios to probe implicit gender–occupation associations; RoBERTa-base exhibits 29.8% gender confusion in doctor/nurse classification.
Mitigation Approaches and Limitations
- Data-level: Hugging Face
datasetslibrary v2.18.0 includes built-inremove_biaspreprocessing for gender-balanced sampling; - Model-level: Google Research’s FairLLM applies LoRA-based fairness-aware fine-tuning to Llama-2-7b, improving WinoBias accuracy by 18.4%;
- Evaluation-level: The arXiv survey paper Measuring and Mitigating Gender Bias in Large Language Models: A Survey (2024, arXiv:2402.13256) reviews 47 bias detection and mitigation techniques, highlighting critical gaps in cross-lingual generalization (e.g., English → Chinese transfer) and intersectional bias handling (e.g., gender × race).
Sources (compliance trail)
https://thegradient.pub/gender-bias-in-ai/