· Xiaojing Yang · Explainability & Responsible AI · 1 min read
Bias Evaluation Is More Than Prompting
A responsible-AI note on controlled evaluation design, prompts, labels, metrics, human validation, and model-auditing evidence.
中文导读
偏见评估不能只靠“我问了模型几个问题”。如果要让结论可信，需要控制变量、设计 prompt 模板、定义标签、检查 annotation quality，并把自动指标和人工证据联系起来。
Working outline
- Why casual prompting is not evaluation
- Case design and controlled variables
- Prompt templates as experimental controls
- Label definitions and human validation
- Metrics and statistical reliability
- What counts as evidence in model auditing
Reference materials to digest
- Google People + AI Guidebook
- Google Responsible AI practices
- Distill: Building Blocks of Interpretability
My angle
This post should become a bridge between foundation topics like statistics and research topics like VLM bias, model auditing, and data attribution.