Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Explainability (XAI)
Set of techniques and methods that make AI model decisions understandable to humans, essential for regulatory compliance and user trust.
Interpretability
Ability of an AI model to present its internal mechanisms in an understandable way, distinguishing intrinsic interpretability (transparent models) from post-hoc interpretability.
Right to Explanation GDPR
Legal obligation for organizations to provide a clear and meaningful explanation of automated decisions affecting individuals, in accordance with Article 22 of the GDPR.
Algorithmic audit
Systematic process of evaluating AI algorithms to verify their compliance with legal, ethical and technical requirements, including bias testing and documentation.
Black box
AI model whose internal mechanisms are opaque and difficult to interpret, posing major challenges for regulatory audit and transparency.
SHAP (SHapley Additive exPlanations)
Explanation method based on game theory quantifying the impact of each feature on the prediction, offering theoretical guarantees for regulatory audit.
LIME (Local Interpretable Model-agnostic Explanations)
Local interpretation technique explaining individual predictions by approximating the complex model with a locally interpretable simple model.
Feature Importance
Quantitative measure of the relative influence of each input variable on model predictions, essential for documenting decision factors in audit.
Counterfactuals
Explanations showing what minimal modifications to input characteristics would change the model's decision, helping to understand and challenge automated decisions.
Model documentation
Structured and comprehensive recording of the characteristics, performance, limitations, and decision-making processes of an AI model, required for regulatory compliance.
Decision traceability
Ability to track and document the entire decision-making process of an AI system, from input data to final output, essential for legal audit.
AI impact assessment
Systematic evaluation of the potential risks of an AI system on fundamental rights and society, mandatory under European AI regulation.
Decision justification
Obligation to provide clear and specific reasons supporting each automated decision, allowing individuals to understand and challenge the results.
Fairness metrics
Quantitative indicators measuring potential biases and discrimination in algorithmic decisions, essential for compliance with anti-discrimination regulations.
Robustness tests
Systematic evaluations of the stability of a model's predictions in the face of variations in input data, ensuring the reliability required for regulatory audit.
Global sensitivity analysis
Method assessing the overall impact of each variable on the model's predictions across all data, providing an overview for regulatory audit.
Algorithm Register
Centralized database listing all AI algorithms used by an organization, with their characteristics and risk levels, required for regulatory transparency.