Bias and fairness in AI — AIF-C01

What it is

Bias in AI is the presence of imbalances in data or disparities in the performance of a model across different groups. Fairness is one of eight core dimensions of responsible AI that AWS defines as considering impacts on different groups of stakeholders.

These are distinct but related ideas: bias is the measurable phenomenon; fairness is the design goal of reducing unequal impact.

Mental model

Think of bias as a signal-to-noise problem that compounds across the AI lifecycle. Imbalances can enter at the data stage (the training set does not represent all groups equally), at the algorithm stage (model design choices amplify existing patterns), and at the human-feedback stage (labels or reinforcement signals encode the labeler's assumptions). Because all three sources feed into model outputs, controlling for only one source does not guarantee a fair outcome.

The diagram above shows these three sources converging into model outputs as disparities across groups.

When to use it

Scenario	Term to reach for	Why
Training data does not represent all demographic groups equally	Data bias	The imbalance originates in the dataset, before any model is trained
A model's accuracy is lower for one group than another after training	Disparities in model performance	The bias has manifested in measurable output differences
A design decision in the model amplifies existing patterns	Algorithm bias	The source is architectural, not the dataset
An annotator's assumptions shape the training labels	Human bias	The source is the person providing feedback
The goal of reducing differential impact on stakeholders	Fairness	Fairness is the outcome target; bias describes the problem

Candidates often reach for "fairness" when describing the technical phenomenon and "bias" when describing the ethical goal — these are reversed. Bias describes what is measured; fairness describes what is designed for.

Common misconception

The trap: Bias is often assumed to be solely a data problem that disappears once "better data" is collected.

The official AWS responsible-AI framing explicitly identifies bias as something that can appear at multiple points: during data preparation, after model training, and in deployed models. This means a model trained on representative data can still produce biased outputs because of algorithmic choices or human-sourced labels, and a deployed model can develop new disparities as the real-world distribution shifts. Treating data quality as the only lever leads candidates to dismiss algorithm- and deployment-stage interventions as irrelevant — a mistake when a question describes a post-deployment scenario.

A second, related trap is treating fairness and explainability as the same concept. AWS defines them as separate dimensions: fairness concerns differential impact across groups, while explainability concerns understanding and evaluating system outputs. A model can be explainable (its decisions are interpretable) without being fair (its decisions still disadvantage a group).

How it shows up on the exam

The cognitive target is applying understanding of the responsible-AI dimensions, particularly distinguishing fairness from adjacent concepts such as explainability and transparency.

Watch for:

• Questions that describe a deployed model performing worse for one group and ask which responsible-AI dimension is at risk. The answer draws on the fairness definition ("considering impacts on different groups of stakeholders"), not on explainability or safety.
• Questions that present a data-quality improvement and ask whether it eliminates bias. Because AWS frames bias as detectable at data preparation, post-training, and in deployed models, data fixes alone are not a complete answer.
• Scenarios that describe making a model's decision logic interpretable and ask which dimension is served. That is explainability, not fairness — even if the underlying goal is equitable treatment.

Signal phrases in stems: "different groups," "disparate performance," "imbalances in data," "stakeholder impact."

Related concepts

• Responsible AI — the eight-dimension framework within which fairness and bias sit
• Explainable AI — the dimension most commonly confused with fairness; concerns output interpretability, not group impact
• Model transparency — enabling stakeholders to make informed choices; a governance concern distinct from fairness