DataTerminology.comNEW
Systematic and repeatable errors in algorithmic systems that produce unfair outcomes across individuals or groups.
DESCRIPTION
Algorithmic bias arises from data, modeling, measurement, or deployment choices that lead to inequitable outcomes. It can manifest as disparate performance across populations or systematic errors favoring certain groups. Sources include skewed sampling, label noise, proxies for sensitive attributes, and feedback loops. Mitigation involves better data practices, fairness-aware training, post-processing, and human oversight. Continuous monitoring and impact assessments are necessary to catch drift and context changes.
ALTERNATIVE DEFINITIONS
Wikipedia
Algorithmic bias is systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others.
Stresses systematic errors producing unfair outcomes for certain groups.
NIST SP 1270
Bias in AI can be understood as a systemic error in the output of AI systems that results in unfair outcomes.
Frames bias as systemic error linked to unfair outcomes and standardization needs.
RELATED TERMS
Fairness Sampling Bias Label Bias Proxy Bias Measurement Bias Disparate Impact Fairness Metrics Debiasing Explainability Accountability
COMPETING TERMS
RELATIONS
- Broader: Bias
KNOWLEDGE AREAS
TARGET AUDIENCE
Data Scientist AI Engineer MLOps Engineer AI Product Manager Data Governance Manager Chief Data Officer
EXTERNAL SOURCES
- Wikipedia - Algorithmic bias
- Google Developers - Overview of fairness and bias
- Barocas, Hardt, Narayanan - Textbook on fairness and bias
- NIST - Bias in AI and mitigation
TOP QUESTIONS
What causes Algorithmic Bias?
Skewed data, measurement errors, proxies for sensitive attributes, model choices, and deployment context can all introduce bias.
Category: Causes
How do you detect Algorithmic Bias?
Use fairness metrics and disaggregated performance analysis across relevant subgroups, combined with domain reviews.
Category: Detection
What are common mitigation strategies for Algorithmic Bias?
Data balancing, adversarial debiasing, constraint-based training, threshold adjustments, and human-in-the-loop controls are common approaches.
Category: Mitigation
Why is Algorithmic Bias a persistent risk?
Data and context change over time, while models may entrench historical patterns, requiring ongoing monitoring and updates.
Category: Risks
Is Algorithmic Bias always illegal?
Not always; legality depends on jurisdiction and sector. However, many biased outcomes breach ethical norms and can still cause harm.
Category: Compliance
How does Algorithmic Bias relate to fairness metrics?
Fairness metrics formalize specific notions of equal treatment or outcomes to assess and compare biases.
Category: Relation
What is a common misunderstanding about Algorithmic Bias?
That removing protected attributes eliminates bias; proxies and data processes can still encode sensitive information.
Category: Misunderstanding
How do regulations impact Algorithmic Bias management?
Laws and guidance require risk assessments and mitigation, with sector-specific standards emerging.
Category: Regulation
When should bias checks be performed?
Throughout the lifecycle: data design, training, evaluation, deployment, and continuous monitoring stages.
Category: Lifecycle