name substitution

**Name substitution** is the **fairness evaluation and augmentation technique that replaces personal names to probe demographic sensitivity in model behavior** - it helps detect bias tied to ethnicity, gender, or cultural identity signals. **What Is Name substitution?** - **Definition**: Paired-text transformation where only personal names are changed while context remains constant. - **Evaluation Purpose**: Measure whether outputs differ due to demographic proxy cues from names. - **Augmentation Use**: Build more demographically balanced training examples. - **Method Constraint**: Substitutions must preserve semantics and pragmatic plausibility. **Why Name substitution Matters** - **Bias Auditing**: Exposes unequal model treatment associated with identity-coded names. - **Fairness Improvement**: Supports targeted data interventions where name-linked bias is observed. - **Causal Clarity**: Paired tests isolate demographic signal effects from content differences. - **Risk Reduction**: Helps prevent discriminatory behavior in user-facing applications. - **Benchmark Alignment**: Useful for evaluating progress on fairness metrics over model versions. **How It Is Used in Practice** - **Name Sets**: Use curated balanced name lists with documented demographic coverage. - **Paired Scoring**: Compare probabilities, classifications, and generated sentiment across substitutions. - **Mitigation Feedback**: Feed detected disparities into retraining and policy refinement. Name substitution is **a practical fairness-testing instrument in LLM evaluation** - controlled identity-proxy swaps provide actionable evidence for detecting and correcting demographic bias patterns.