AI Bias Categories and Definitions

According to ISA/IEC 42001, AI bias enters at multiple points:

1.      Historical bias (training data reflects past discrimination),

2.      Measurement bias (proxies correlating with protected characteristics),

3.      Aggregation bias (works well on average, fails for subgroups),

4.      Deployment bias (used outside the intended context).

Bias can be invisible in overall metrics while causing severe harm to specific groups. Disaggregated testing (across demographic groups) is essential.

Ongoing monitoring includes:

• performance metrics (accuracy, precision, recall disaggregated by group),

• bias indicators (demographic parity, equalized odds),

• drift detection (distribution shift, concept drift),

Establish clear thresholds and escalation procedures for when metrics indicate problems.

Bias, as defined by NIST Special Publication 1270, is:

1. Systemic AI bias occurs when an AI system consistently produces unfair or discriminatory outcomes for certain groups of people because it reflects and amplifies deeper societal, institutional, or historical biases present in its training data, design, or real-world use. It is widespread and structural (not random or isolated) and often comes from historical data that already contains prejudice.

2. Statistical/Computational bias occurs when an AI system produces unfair, skewed, or inaccurate results due to problems in its data, algorithms, or mathematical design — not because of deeper societal prejudices. Purely technical flaws in data or algorithms that cause skewed results (can exist even in neutral settings).

3. Human biases, whether conditioned socially or unconscious cognitive bias, are factors in data selection, curation, preparation, and analysis processes

Here is a clear, practical table summarizing the most common types of bias in AI systems:

Category	Type of Bias	Simple Definition
Systemic / Societal	Systemic Bias	Bias embedded in society, institutions, or history that the AI learns and amplifies (e.g., historical discrimination reflected in data).
Systemic / Societal	Historical Bias	Bias that comes from past societal inequalities is captured in training data.
Data-Related	Representation / Sampling Bias	Training data does not accurately represent the real-world population (due to under- or over-representation of groups).
Data-Related	Measurement / Label Bias	Errors or inconsistencies in how data is measured, collected, or labeled (proxies or flawed ground truth).
Data-Related	Selection Bias	Data is selected non-randomly, skewing the dataset.
Data-Related	Reporting Bias	Certain events or outcomes are reported more frequently than others in the data.
Algorithmic / Computational	Algorithmic Bias	Bias is introduced by the model design, optimization process, or mathematical assumptions.
Algorithmic / Computational	Aggregation Bias	The model treats all data as a single group even when subgroups behave differently.
Algorithmic / Computational	Optimization / Objective Bias	The goal function or loss function favors certain outcomes over others.
Human / Cognitive	Confirmation Bias	The AI (or its developers) favors information that confirms existing beliefs or patterns.
Human / Cognitive	Automation Bias	Humans over-rely on AI outputs and fail to question them.
Human / Cognitive	Implicit / Cognitive Bias	Unconscious human prejudices are introduced during data labeling, feature selection, or model design.
Emergent / Deployment	Feedback Loop Bias	Bias that worsens over time because the AI’s predictions influence future data it is trained on.
Emergent / Deployment	Evaluation / Benchmark Bias	Bias from using flawed or unrepresentative test metrics or benchmarks.
Emergent / Deployment	Deployment Bias	Bias that arises when the model is used in a context different from the one in which it was trained.

Quick Notes:

• Many biases overlap (e.g., historical bias often leads to systemic bias).

• Systemic bias is often the most harmful because it scales societal inequalities.

• Statistical/Computational bias focuses on technical flaws in data and algorithms.

• Human bias enters at every stage through people's decisions.