AI Bias & Fairness Cheat Sheet

Updated 2026-04-28

Next Topic: AI Ethics and Responsible AI Cheat Sheet

AI bias and fairness are critical dimensions of responsible AI that address how machine learning systems may produce discriminatory outcomes based on protected attributes like race, gender, or age. Bias emerges from multiple sources — biased training data, flawed algorithmic design, or problematic evaluation methods — and manifests as systematic unfairness toward specific demographic groups. Fairness aims to ensure equitable treatment and outcomes through mathematical constraints, mitigation techniques, and ongoing monitoring, though trade-offs often exist between different fairness definitions and between fairness and accuracy. By 2026, these challenges have expanded beyond traditional predictive models to encompass large language models, generative AI, and agentic systems — where existing fairness frameworks developed for classification tasks no longer fully suffice.

What This Cheat Sheet Covers

This topic spans 16 focused tables and 117 indexed concepts. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Types of Bias in AI SystemsTable 2: Group Fairness MetricsTable 3: Individual and Counterfactual FairnessTable 4: Intersectional and Subgroup FairnessTable 5: Fairness-Accuracy Trade-offsTable 6: Pre-Processing Bias MitigationTable 7: In-Processing Bias MitigationTable 8: Post-Processing Bias MitigationTable 9: Causal FairnessTable 10: Bias Detection and EvaluationTable 11: Fairness Tools and FrameworksTable 12: Feedback Loops and Dynamic FairnessTable 13: Case Studies and Real-World ApplicationsTable 14: Regulatory and Ethical FrameworksTable 15: Explainability for FairnessTable 16: Fairness in LLMs and Generative AI

Table 1: Types of Bias in AI Systems

Type	Example	Description
Historical bias	Training on past hiring data with male preference	Arises when training data reflects past societal inequalities and discriminatory practices — even with perfect sampling, the data encodes historical injustices.
Representation bias	Dataset with 90% images of light-skinned faces	Occurs when training data does not accurately reflect the real-world population distribution — certain groups are over- or under-represented.
Measurement bias	Using credit score as proxy for creditworthiness	Arises from using flawed proxies or imperfect features to measure a concept — the measurement itself systematically differs across groups.
Label bias	Inconsistent annotations by human labelers	Introduced when human annotators apply subjective judgments or stereotypes during data labeling — different labelers may tag the same data differently.
Aggregation bias	Single diabetes model for all ethnic groups	Results from applying a one-size-fits-all model to populations with different data-generating processes — assumes homogeneity when subgroups differ.
Selection bias	Survey data collected only from smartphone users	Arises when data collection systematically excludes or oversamples certain populations — the sampling process itself introduces skew.

Table 1: Types of Bias in AI Systems

Type	Example	Description
Historical bias	Training on past hiring data with male preference	Arises when training data reflects past societal inequalities and discriminatory practices — even with perfect sampling, the data encodes historical injustices.
Representation bias	Dataset with 90% images of light-skinned faces	Occurs when training data does not accurately reflect the real-world population distribution — certain groups are over- or under-represented.
Measurement bias	Using credit score as proxy for creditworthiness	Arises from using flawed proxies or imperfect features to measure a concept — the measurement itself systematically differs across groups.
Label bias	Inconsistent annotations by human labelers	Introduced when human annotators apply subjective judgments or stereotypes during data labeling — different labelers may tag the same data differently.
Aggregation bias	Single diabetes model for all ethnic groups	Results from applying a one-size-fits-all model to populations with different data-generating processes — assumes homogeneity when subgroups differ.
Selection bias	Survey data collected only from smartphone users	Arises when data collection systematically excludes or oversamples certain populations — the sampling process itself introduces skew.