Viés em IA
Definição
Viés em IA refere-se a erros sistemáticos ou resultados injustos (por ex. across demographics) arising from data, model projeto, or deployment. Mitigation includes data audits, fairness metrics, and debiasing methods.
É a core concern in AI ethics and AI safety. Evaluation metrics for fairness (por ex. demographic parity, equalized odds) are used in audits and before deploying in regulated domains. Explainable AI can help identify when and why bias appears.
Como funciona
O viés pode entrar por dados de treinamento enviesados (sub-representação, viés de rótulo), variáveis proxy (por ex. CEP para raça), ou loops de feedback (modelo outputs influence future data). Detection uses fairness metrics (por ex. demographic parity, equalized odds, calibration by group) on evaluation sets stratified by protected attributes. Mitigation includes: data (reweighting, resampling, collecting more representative data); training (fairness constraints, adversarial debiasing); and post-processing (thresholds or rules per group). Trade-offs exist between fairness metrics and accuracy; legal and domain norms define which metrics and thresholds to use. Audits should be run before deployment and monitored in production.
Casos de uso
Bias work applies when model decisãos affect people in regulated or sensitive domains (hiring, lending, scoring, content).
- Auditing hiring, lending, or scoring systems for discriminatory impact
- Fairness checks before deploying models in regulated domains
- Explainability and remediation when bias is detected