Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Algorithmic Fairness
Fundamental principle aimed at ensuring that artificial intelligence systems produce fair and non-discriminatory outcomes for all individuals or groups, regardless of their protected characteristics.
Algorithmic Bias
Systematic distortion in an algorithm's predictions or decisions that unfairly favors or disadvantages certain groups, often resulting from biased training data or inappropriate design.
Demographic Parity
Fairness criterion requiring that the proportion of positive outcomes be identical across different demographic groups, regardless of actual individual characteristics.
Equalized Odds
Fairness principle ensuring that true positive and false positive rates are equal across different groups, guaranteeing similar predictive performance for all populations.
Calibration Fairness
Fairness measure requiring that for any given prediction score, the actual probability of the outcome be the same for all relevant demographic groups.
Counterfactual Fairness
Fairness approach evaluating whether a prediction would remain unchanged if an individual's protected characteristics were different, while keeping all other attributes constant.
Individual Fairness
Principle stating that similar individuals should receive similar treatments or predictions from the AI system, ensuring consistency at the individual level.
Group Fairness
Fairness measure evaluating whether different demographic groups receive statistically similar outcomes from the AI system, without consideration of specific individual characteristics.
Disproportionate Impact
Unequal negative or positive effect of an algorithm on different demographic groups, measured by the statistical gap in outcome rates for each group.
Distributive Justice
Philosophical concept applied to AI ensuring a fair distribution of resources, opportunities, or benefits generated by algorithmic systems among all concerned groups.
Procedural Fairness
Principle ensuring that algorithmic decision-making processes are transparent, consistent, and allow for recourse, regardless of the outcomes produced.
Predictive Parity
Criterion guaranteeing that the proportion of correct predictions is the same for all groups, ensuring equitable reliability of predictions across different populations.
Algorithmic Privilege
Systematic advantage granted by an algorithm to certain groups or individuals, resulting from implicit biases in the model's design or training.
Fair Treatment
Fundamental principle requiring that AI systems treat all people impartially and consistently, without discrimination based on protected or sensitive characteristics.
Fairness Through Awareness
Methodology that explicitly incorporates knowledge of potential biases and protected characteristics into the design and evaluation of models to ensure fair outcomes.
Algorithmic Discrimination
Differentiated and unfair treatment of certain groups by an AI system, resulting from algorithmic decisions that create or perpetuate systemic inequalities.
Longitudinal Equity
Principle ensuring that the fairness of an AI system is maintained over time, preventing the emergence of new biases or the amplification of existing inequalities as the model evolves.