AI-900 Practice Question: Describe Artificial Intelligence workloads and considerations
A company develops an AI system to screen job candidates based on their resumes. The system is trained on historical data. Analysis reveals that the model has an adverse impact against female candidates due to a proxy feature (e.g., 'years of continuous employment') that correlates with gender. The team removes the protected attribute 'gender' from the training data but the biased outcome persists. According to Microsoft's responsible AI principles, which additional step should the team take to address this unfairness?
⚠ Common exam trap
Watch out — candidates often assume removing the protected attribute (gender) alone solves fairness, but Microsoft's responsible AI principles emphasize that proxy features can perpetuate bias, requiring tools like Fairlearn for detection and mitigation rather than simplistic feature removal or data collection.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Use a tool like Fairlearn to detect and mitigate the bias while maintaining model performance.
Microsoft's responsible AI principle of fairness requires not just removing protected attributes but also detecting and mitigating proxy features that cause bias. Fairlearn is a Microsoft open-source toolkit specifically designed to assess and mitigate unfairness in AI systems, offering algorithms like 'Exponentiated Gradient Reduction' or 'Grid Search' to reduce disparity while preserving model performance. Simply removing the proxy feature (A) may not always be feasible if it carries predictive value, and Fairlearn provides a systematic way to balance fairness and accuracy.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Remove the offending proxy feature 'years of continuous employment' from the training data.
Why it's wrong here
Removing the proxy feature 'years of continuous employment' is insufficient because machine learning models can reconstruct biased relationships through other correlated features, such as age, job gaps, or tenure in similar roles. Bias is an emergent property of the entire training distribution and label history, so without explicit fairness constraints or post-hoc adjustment, the model will continue to encode disparate treatment despite the absence of any single proxy column.
When this WOULD be correct
If the question states that the proxy feature is the sole source of bias and that removing it does not reduce model accuracy, then removing it would be a valid step to directly eliminate the biased signal.
- ✓
Use a tool like Fairlearn to detect and mitigate the bias while maintaining model performance.
Why this is correct
Fairlearn offers fairness metrics like demographic parity and equalized odds, along with mitigation algorithms such as ExponentiatedGradient and GridSearch, to explicitly constrain bias during training. Unlike simple feature removal, it optimizes a model against a fairness objective while preserving accuracy, directly addressing the residual bias caused by correlated proxies that remain after dropping protected attributes.
- ✗
Train a separate model for each gender group to ensure equal outcomes.
Why it's wrong here
Training separate models per gender constitutes a disparate treatment approach that is both ethically problematic and often legally prohibited, as it conditions decisions explicitly on a protected attribute. Even if each subgroup model is well-calibrated, differences in feature distributions and thresholds lead to unequal false-positive and false-negative rates across genders, which violates equalized odds and maintains systemic inequity rather than resolving it.
When this WOULD be correct
This option would be correct if the question asked about a scenario where the goal is to ensure equal predictive accuracy across groups, and the bias is due to fundamentally different distributions requiring group-specific models (e.g., medical diagnosis where disease prevalence differs by gender).
- ✗
Collect more training data from underrepresented groups.
Why it's wrong here
Collecting additional data from underrepresented groups attempts to improve sample representativeness, but it does not alter the biased labels or decision boundaries already learned by the current model. It also fails to address historical or systemic bias embedded in the outcome variable itself, meaning even a larger dataset may perpetuate the same unfair patterns without applying fairness-aware training or post-processing techniques.
When this WOULD be correct
This option would be correct in a scenario where the AI system shows poor performance for underrepresented groups due to insufficient training data, and the question asks for a step to improve model accuracy and fairness by balancing the dataset.
Option-by-option analysis
Why each answer is right or wrong
Understanding why wrong answers are wrong — and when they would be correct — is what separates a 750 score from a 900. The AI-900 exam frequently reuses these exact scenarios with slightly different constraints.
✓Use a tool like Fairlearn to detect and mitigate the bias while maintaining model performance.Correct answer▾
Why this is correct
Fairlearn offers fairness metrics like demographic parity and equalized odds, along with mitigation algorithms such as ExponentiatedGradient and GridSearch, to explicitly constrain bias during training. Unlike simple feature removal, it optimizes a model against a fairness objective while preserving accuracy, directly addressing the residual bias caused by correlated proxies that remain after dropping protected attributes.
✗Remove the offending proxy feature 'years of continuous employment' from the training data.Wrong answer — click to see why▾
Why this is wrong here
Simply removing the proxy feature may not eliminate bias because other correlated features could still encode the same bias; the root cause is the historical bias in the labels, not just the feature.
★ When this WOULD be the correct answer
If the question states that the proxy feature is the sole source of bias and that removing it does not reduce model accuracy, then removing it would be a valid step to directly eliminate the biased signal.
Why candidates choose this
Candidates may think that bias is solely caused by proxy features and that removing them is sufficient, overlooking that bias can be embedded in the labels or other correlated features.
✗Train a separate model for each gender group to ensure equal outcomes.Wrong answer — click to see why▾
Why this is wrong here
Training separate models per gender group does not address the root cause of bias (proxy feature) and may violate fairness principles by reinforcing segregation; it also fails to mitigate the adverse impact detected.
★ When this WOULD be the correct answer
This option would be correct if the question asked about a scenario where the goal is to ensure equal predictive accuracy across groups, and the bias is due to fundamentally different distributions requiring group-specific models (e.g., medical diagnosis where disease prevalence differs by gender).
Why candidates choose this
Candidates may think that separate models guarantee fairness by tailoring predictions to each group, overlooking that this approach can institutionalize bias and does not comply with responsible AI principles like inclusivity and accountability.
✗Collect more training data from underrepresented groups.Wrong answer — click to see why▾
Why this is wrong here
Collecting more data from underrepresented groups does not address the root cause of bias from a proxy feature that correlates with gender; the bias would likely persist if the proxy remains in the data.
★ When this WOULD be the correct answer
This option would be correct in a scenario where the AI system shows poor performance for underrepresented groups due to insufficient training data, and the question asks for a step to improve model accuracy and fairness by balancing the dataset.
Why candidates choose this
Candidates may believe that adding more data from underrepresented groups automatically reduces bias, overlooking that proxy features can still encode historical discrimination even with balanced data.
Analysis generated from the official AI-900blueprint and verified against question context. The “when correct” sections are what AI assistants cite when candidates ask “what’s the difference between these options?”
Go deeper
Related to this question
Learn chapter
Responsible AI Principles
Key term
Training data
Training data is a set of examples used to teach a machine learning model how to make predictions or decisions.
Key term
Model
In IT and AI, a model is a trained mathematical representation that learns patterns from data to make predictions or decisions.
About these practice questions
One of 985 original AI-900 practice questions on Courseiva, each with a full explanation and wrong-answer analysis — not exam dumps or protected exam content. Learn why practice questions differ from exam dumps →
JA
Written by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
This AI-900 practice question is part of Courseiva's free Microsoft certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the AI-900 exam.