Cloud concept · Updated

Bias and Fairness - Bias and fairness describe how AI systems can produce systematically skewed outcomes and how those outcomes are assessed for equitable treatment of people and groups.

In short

Bias can enter an AI system through its data, labels, design choices, or operating environment, potentially creating unjustified differences in outcomes. Fairness means evaluating and addressing those differences using criteria appropriate to the application, affected people, and applicable requirements.

Cloud Concepts Hub →

6 key aspects of Bias and Fairness

Bias and fairness require lifecycle-wide evaluation, context-specific metrics, governed data use, deliberate mitigation, and continued oversight.

01
Bias can enter throughout the lifecycle

Inspect data collection, labeling, feature selection, model training, and deployment conditions. A representative training dataset alone does not prevent an inappropriate decision policy from producing unfair outcomes.

02
Fairness requires explicit criteria

Choose metrics that match the decision and its potential harms, such as differences in approval rates or missed eligible applicants. Document why those criteria are appropriate rather than treating one metric as universally correct.

03
Subgroup evaluation reveals hidden disparities

Compare performance across relevant groups and, when sample sizes permit, intersecting groups. Report uncertainty because small evaluation samples can make disparities difficult to estimate reliably.

04
Mitigation involves measurable trade-offs

Evaluate data improvements, reweighting, model constraints, and decision-policy changes against both fairness and utility requirements. Satisfying one fairness definition may worsen another, so record the resulting trade-offs.

05
Sensitive attributes need careful governance

Removing sensitive attributes does not eliminate proxy discrimination. Where lawful and appropriate, protect the attributes needed for fairness measurement with restricted access and clear retention policies.

06
Fairness needs continuous oversight

Reassess models when production populations, data, or decision processes change. Combine automated checks with accountable review, documented limitations, and mechanisms for addressing harmful outcomes.

Cloud services for Bias and Fairness

SageMaker Clarify helps assess bias in training data and model predictions before release. SageMaker Model Monitor extends supported workflows into production by checking for bias drift against configured baselines. For generative applications, Amazon Bedrock evaluation workflows can support assessment of response fairness using representative prompts and human review. Together, these capabilities support measurement across development and operation, but teams still define affected groups, select appropriate criteria, protect sensitive evaluation data, and decide how to mitigate findings. Reports and alerts provide evidence for review rather than certification that a system is fair.

Amazon Bedrock

Amazon Bedrock model evaluation supports evaluating generative model responses, including human evaluation for criteria such as fairness. Use representative prompts and explicit evaluation rubrics; access to a managed foundation model does not guarantee equitable outputs.

Bias and Fairness exam questions

Real practice exam questions on Bias and Fairness, with the correct answer and its explanation. Answer them with every option in the exam simulator.

AIB-C01 (beta)

A retail bank is developing AI literacy training for branch managers who will evaluate AI proposals. In a workshop, managers conclude that a customer outreach model cannot produce unfair outcomes because it does not use protected characteristics as inputs. The model does use postal codes, purchasing patterns, and account history. The learning director wants managers to recognize the limitations of this conclusion and practice evaluating supporting evidence without becoming model developers. Which training activity BEST meets this objective?

Show answerCORRECT ANSWER

Use a controlled exercise with approved synthetic customer data in which managers examine how correlated attributes can preserve group differences, compare outreach outcomes across relevant groups, and document findings that require further investigation.

This exercise demonstrates how apparently neutral inputs can convey information related to protected characteristics. It develops practical skills in examining group outcomes and identifying issues for investigation without assuming that every difference proves unfairness or requiring managers to build models.

MLA-C02 (beta)

A property insurer is preparing a multimodal repair-cost regression dataset in Amazon SageMaker AI. Each record contains numeric damage measurements, an adjuster's report, property photographs, and a historical repair-cost target. The dataset has adequate coverage of each region, report language, and camera type.

Independent experts have produced verified reference costs for a representative audit sample by reviewing all three modalities. Historical targets have nearly identical mean absolute errors across regions. However, complaints suggest that targets systematically underestimate costs for records from one region when reports use a minority language and photographs come from older cameras. Valid repair costs differ across regions, so the company must not force regional target distributions to be identical.

Which approach should an ML engineer take before training?

Show answerCORRECT ANSWER

Use a SageMaker Processing job to calculate historical-target-minus-reference-cost errors for region, report-language, and camera-type combinations, conditioning on comparable damage severity. Investigate combinations with systematic negative errors and obtain expert-approved target corrections before repeating the audit.

Signed errors reveal systematic underestimation that mean absolute error cannot distinguish from overestimation. Intersectional analysis can expose a problem hidden by regional averages, while conditioning on damage severity helps separate target errors from legitimate cost differences. Expert-approved corrections address the identified target bias without imposing identical regional cost distributions.

AIF-C01

A lender trains a loan-screening model in Amazon SageMaker AI. The model achieves 95% overall accuracy on a labeled evaluation dataset, but some age groups have far fewer examples than others. Before deployment, the lender wants to assess whether applicants who meet its lending criteria are incorrectly rejected at different rates across age groups. Which approach best meets this requirement?

Show answerCORRECT ANSWER

Use SageMaker Clarify to compare false-negative rates across age groups on labeled evaluation data and investigate any disparities.

A false negative represents an applicant who meets the lending criteria but is incorrectly rejected. SageMaker Clarify supports post-training bias analysis, including comparisons of false-negative rates across groups. These comparisons address the lender's concern directly; disparities require investigation in the context of the application.

Practice Bias and Fairness in a full exam

Questions on Bias and Fairness in each practice exam, with every answer option and an explanation for each:

Start a free practice exam

What is Bias and Fairness

3 min read

Bias in AI commonly refers to systematic patterns in data, predictions, or decisions that can disadvantage particular people or groups. Fairness concerns whether a system treats people equitably under explicitly chosen criteria. These concepts are related but not interchangeable: a measured difference between groups requires investigation, while deciding whether that difference is unfair depends on the task, context, and consequences. In other machine learning contexts, bias can also mean a model's simplifying assumptions or the error component in the bias–variance trade-off; those meanings do not automatically imply discrimination.

Unfair outcomes can originate before model training. A speech recognition dataset might underrepresent certain accents, leading to higher transcription error rates for those speakers. Historical hiring labels might encode past discrimination, so learning to reproduce those labels can perpetuate inequity. Measurement errors, subjective annotations, missing records, and deployment to a population different from the training population can also introduce or amplify disparities.

Fairness assessment begins by defining the decision, the affected groups, and the harms that matter. For a loan screening model, engineers might compare approval rates, false-positive rates, and false-negative rates across relevant groups. Demographic parity compares positive prediction rates, while equal opportunity compares true-positive rates among people who actually meet the target condition. These definitions answer different questions and cannot always be satisfied simultaneously, particularly when underlying outcome rates differ. Overall accuracy alone can hide poor performance for a smaller group.

In a cloud machine learning workflow, teams can version datasets, run subgroup evaluations during training pipelines, and retain assessment reports alongside model artifacts. They can block deployment when agreed thresholds are exceeded and monitor production outcomes as populations or behavior change. For example, a cloud-hosted document classifier might perform well overall but misclassify documents written in a regional dialect. Evaluating that subgroup separately can reveal a problem that an aggregate dashboard misses. Reliable assessment may require delayed ground-truth labels and enough examples to distinguish meaningful differences from sampling noise.

Mitigation can involve collecting more representative data, improving labels, reweighting training examples, changing the learning objective, or adjusting decision policies. Removing sensitive attributes does not guarantee fairness because other inputs, such as location, can act as proxies. Conversely, appropriately governed access to sensitive attributes may be necessary to measure disparities. Collection and use of those attributes require privacy safeguards, access controls, and review of applicable requirements.

Generative AI also requires fairness testing. An LLM can reproduce stereotypes, associate occupations with particular demographic groups, or provide uneven response quality across languages. Teams can test comparable prompts that vary demographic references, use representative evaluation datasets, and involve human reviewers. A biased response is not necessarily a hallucination: an output can be factually grounded yet selectively framed or unfair.

Fairness is an ongoing engineering and governance responsibility, not a property guaranteed by using a managed service or passing one metric. Mitigations can affect accuracy, cost, latency, and user experience, and fairness criteria can conflict with one another. Document the chosen criteria, residual limitations, and escalation process. For high-impact decisions, provide appropriate human oversight and a way to challenge outcomes, while recognizing that human reviewers can also introduce bias.

Learn Bias and Fairness in lessons

5 short lessons on Bias and Fairness in 3 certification courses.

Open the lessons