Human biases

Human cognitive limitations and biases that can produce AI systems whose decisions are affected by systemic errors in human judgment.

Source: AIEOG AI Lexicon (Feb 2026), adapted from NIST SP 1270

Human biases are the systematic patterns in human thinking that lead to distorted judgment. In the AI context, human biases matter because they can transfer into AI systems at every stage of development: from defining the problem, to selecting and labeling training data, to designing model features, to interpreting outputs.

AI does not develop biases independently. It learns them from the data and decisions humans provide. Every dataset reflects the choices, assumptions, and blind spots of the people who collected and curated it. Every feature selection reflects the designer’s assumptions about what matters.

Common human biases that affect AI systems include confirmation bias, anchoring bias, availability bias, and automation bias (over-trusting AI outputs because they come from a computer).

Human biases that enter AI systems can produce discriminatory, inaccurate, or unfair outcomes at scale. What starts as an individual human bias can become an institutional bias when embedded in an AI model that makes thousands of decisions per day.

In financial services, human bias can lead to discriminatory lending decisions, unfair fraud detection if investigators’ past decisions were influenced by bias, biased risk ratings reflecting subjective human judgments, and inappropriate customer treatment.

Automation bias is particularly relevant for compliance teams. As AI systems become more prevalent, there is a tendency to over-rely on their outputs and reduce critical human review.

1. Educate teams on human bias. Train model developers, data curators, validators, and end users on common human biases and how they affect AI systems.
2. Diversify development teams. Diverse teams are more likely to identify biases that homogeneous teams may overlook.
3. Audit training data for human bias. Examine historical data used for model training for patterns that reflect human bias rather than objective truth.
4. Guard against automation bias. Establish processes that encourage critical evaluation of AI outputs rather than automatic acceptance.
5. Include human bias assessment in validation. Model validation should explicitly consider how human biases may have influenced model development.
6. Monitor for bias propagation. Track whether AI system outputs exhibit patterns consistent with known human biases.

Bias, Training data, Data quality/validity, Responsible AI, AI risk assessment, Explainability