bias amplification

**Bias amplification** is the **phenomenon where model outputs exaggerate existing dataset imbalances beyond the original distribution** - amplification can make subtle societal bias significantly more pronounced in generated content. **What Is Bias amplification?** - **Definition**: Increase in biased association strength from training data to model prediction behavior. - **Mechanism Drivers**: Likelihood maximization, majority-pattern preference, and decoding dynamics. - **Observed Effects**: Over-association of demographics with specific professions, traits, or sentiments. - **Measurement Need**: Compare conditional output distributions against source-data baselines. **Why Bias amplification Matters** - **Fairness Degradation**: Amplified stereotypes cause greater representational harm than raw data alone. - **Decision Risk**: Amplification can distort downstream model-assisted judgments. - **Public Impact**: Stronger biased patterns are more visible and damaging in user-facing systems. - **Mitigation Priority**: Requires explicit controls beyond naive data scaling. - **Governance Signal**: Amplification metrics reveal hidden alignment weaknesses. **How It Is Used in Practice** - **Distribution Audits**: Track protected-attribute associations across model versions. - **Training Controls**: Use regularization and balanced objectives to reduce amplification pressure. - **Inference Safeguards**: Apply calibrated decoding and post-generation fairness filters. Bias amplification is **a critical failure mode in fairness-sensitive AI deployment** - mitigating exaggeration effects is essential to prevent models from intensifying societal bias patterns.