Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Disparate Impact Test
Statistical method evaluating whether an algorithm produces disproportionately adverse outcomes for certain protected demographic groups.
Parity Analysis
Quantitative assessment measuring whether positive prediction rates are equivalent across different demographic groups in the results of an AI model.
Bias Metric
Quantitative indicator measuring the level of discrimination or inequity present in the predictions of an artificial intelligence system.
Ethical Cross-Validation
Iterative evaluation technique testing a model's robustness against bias across different data subsets representative of diverse populations.
AI Risk Mapping
Systematic process of identifying, assessing, and documenting potential ethical risks associated with the deployment of an AI system.
Model Transparency
Ability to make the internal mechanisms, training data, and decision processes of an artificial intelligence algorithm understandable.
Model Explainability
Set of techniques enabling the interpretation and explanation of individual predictions of an AI model in a human-comprehensible manner.
Continuous Bias Monitoring
Permanent monitoring system detecting the emergence or amplification of bias in an AI model deployed in production.
Ethical Impact Assessment
Comprehensive analysis of the potential consequences of an AI system on human rights, equity and inclusion before and during its deployment.
Algorithmic Governance
Organizational framework defining responsibilities, processes and controls to ensure the ethical development and use of AI systems.
AI Ethics Certification
Formal validation process attesting that an artificial intelligence system complies with established ethical standards and fairness criteria.
System Auditability
Ability of an AI system to be examined thoroughly and independently to verify its compliance with ethical and regulatory requirements.
Bias Sensitivity Analysis
Systematic evaluation of the variation in performance and biases of a model in response to changes in input data or hyperparameters.
Fairness Score
Composite indicator quantifying the overall fairness level of an algorithm by aggregating several bias and discrimination metrics.
Calibration Test
Statistical verification ensuring that the probabilities predicted by a model correspond to the actual frequencies observed in different demographic subgroups.