Fairness and bias evaluation: AIP-C01 study notes and practice questions
What to know about Fairness and bias evaluation for the AIP-C01 exam, from the AI Safety, Security, and Governance domain (20% of the exam guide), then 1 free question to test it. Part of the AIP-C01 study guide.
AIP-C01 asks how to detect and reduce unequal treatment in model outputs, and how to keep that testing going after launch. Exam questions usually describe a difference in answers for similar requests, or a training set that under-represents a group, and ask for the test or control that finds and fixes it.
Counterfactual testing
A counterfactual test sends pairs of prompts that are identical except for one attribute, such as a name, gender, age group or region. Compare the outputs for quality, tone and the decision they recommend. Run many pairs across a test set. Differences larger than an agreed tolerance are a signal to investigate. Removing protected attributes from the prompt is not enough, because other words can act as proxies, so test the outputs.
LLM-as-a-judge with fairness criteria
Bedrock evaluations can use a judge model with custom metrics that score each response for equal treatment and for stereotyping. This scales to thousands of responses. Judges have their own biases, so calibrate them against a human-labeled sample before relying on the scores. See LLM as a judge.
Comparing prompt variants
Prompt management variants let you test two wordings of the same prompt side by side. Measure each variant on the counterfactual set and choose the one with smaller gaps and acceptable quality.
Human review for sensitive outputs
For high-stakes uses, such as hiring, lending or healthcare triage, use human evaluation jobs with your own work team. People review samples and rate fairness, and their ratings become the reference for the automated judge.
Metrics over time
Publish fairness scores as CloudWatch custom metrics from each scheduled evaluation run. Alarm when the gap between groups crosses the threshold, so a regression after a model or prompt change is caught.
Dataset balance before fine-tuning
Check the training data before customization. Count examples per group and per label, and compare the distribution with the population the model will serve. Rebalance by collecting more examples or by careful augmentation. Glue Data Quality rules can check distributions as part of the data pipeline.
How to choose
Scenario wording
Choose
“The same loan question gets different answers for different names”
Counterfactual prompt pairs with a tolerance
“Score many answers for equal treatment at scale”
LLM-as-a-judge with fairness criteria
“Compare two prompt wordings for bias”
Prompt management variants with an A/B evaluation
“High-stakes hiring or lending output”
Human evaluation with your own work team
“Fine-tuning data underrepresents one group”
Dataset balance check before training
Common traps
Assuming removal of protected attributes ensures fairness. Proxy words still carry signal.
Trusting one judge model without human calibration.
Testing once before launch. Fairness can drift with new prompts and models.
Using SageMaker Clarify for new work. It is closed to new customers, so use Bedrock evaluations.
Measuring only average quality. An average can hide a large gap for one group.
A staffing agency is about to release a model-based tool that drafts short candidate summaries for recruiters. The summaries are built from résumés and interview notes.
The agency's ethics review asks whether the drafts describe equally qualified candidates in the same way, regardless of name, gender or age. A single overall accuracy score on a general benchmark will not answer that, because differences can hide in a few subgroups.
The team must produce evidence before launch, using the model it will actually run, and must be able to repeat the same test after each prompt change.
Which approach meets these requirements?
Choose one answer.
AA. Build a paired test set in which each candidate profile appears twice with only the name, gender or age changed, run both versions through the model, and compare the scores.
BB. Run the general text quality benchmark on the model, and publish its overall accuracy score, which the vendor reports as high, as the evidence that the summaries are fair to candidates.
CC. Add a word filter that blocks gendered pronouns and age-related words in every summary, so the output contains no such terms, and record the filter as the fairness control for the review.
DD. Ask the model in the system prompt to treat all candidates equally, and have a recruiter read one summary per week to confirm that no bias is visible.
B. Wrong: An overall score averages across groups and can hide a gap for one subgroup. It measures general quality, not whether equally qualified candidates are described the same way.
D. Wrong: A prompt instruction is a request, not evidence, and reading a few summaries each week cannot reveal small, consistent gaps. The option gives no repeatable measurement for the ethics review.
Explanation
Paired counterfactual prompts change only the attribute being tested, so any consistent difference in the summaries becomes measurable, and the same set can be rerun after each change. An overall benchmark score can hide subgroup gaps, and word filters, prompt instructions and occasional reading do not measure fairness.
Want more on Fairness and bias evaluation? The 5 full-length practice sets have 7 more Fairness and bias evaluation questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →