This post is a part of the AI-901: Microsoft Azure AI Fundamentals Exam Prep Hub. 
This topic falls under these sections:
Identify AI concepts and capabilities (40–45%)
   --> Describe principles of responsible AI
      --> Describe considerations for fairness in an AI solution

Note that there are 10 practice questions (with answers and explanations) for each section to help you solidify your knowledge of the material. Also, there are 2 practice tests with 60 questions each available on the hub below the exam topics section.

Fairness is one of the core principles of Responsible AI and is an important topic for the AI-901 certification exam. Microsoft emphasizes that AI systems should treat all people fairly and avoid producing biased or discriminatory outcomes.

Understanding fairness in AI means understanding how bias can enter an AI system, how unfair outcomes can affect people, and what organizations can do to reduce those risks.

What Is Fairness in AI?

Fairness in AI means that an AI system should make decisions or recommendations without unjustly favoring or disadvantaging individuals or groups.

An AI solution is considered unfair if it produces biased outcomes based on characteristics such as:

Gender
Race or ethnicity
Age
Religion
Disability status
Nationality
Socioeconomic background

The goal is not simply technical accuracy. An AI model can be highly accurate overall while still treating certain groups unfairly.

Why Fairness Matters

AI systems increasingly influence important real-world decisions, including:

Hiring and recruiting
Loan approvals
Healthcare recommendations
Insurance pricing
Criminal justice assessments
School admissions
Customer service prioritization

If these systems are unfair, they can reinforce or amplify existing social inequalities.

For example:

A hiring AI might prefer resumes from men because historical company data reflects mostly male hires.
A facial recognition system may perform poorly for people with darker skin tones if training data lacked diversity.
A loan approval model may unfairly deny applications from certain neighborhoods because of biased historical lending patterns.

These outcomes can damage trust, create legal risks, and harm individuals.

How Bias Enters an AI System

Fairness problems usually originate from bias in data, design, or implementation.

1. Biased Training Data

AI models learn patterns from historical data. If the historical data reflects human bias, the AI may learn and repeat that bias.

Example

If a company historically hired mostly men for engineering roles, an AI recruiting tool trained on that data may incorrectly learn that male candidates are preferable.

This is one of the most common causes of unfair AI systems.

2. Underrepresentation in Data

Some groups may not be sufficiently represented in the training dataset.

Example

A speech recognition model trained mostly on American English speakers may perform poorly for people with different accents.

When data lacks diversity, the AI system may not generalize well to all users.

3. Labeling Bias

Humans often label training data. Human assumptions and prejudices can influence those labels.

Example

If reviewers consistently rate certain groups more negatively during data labeling, the AI model may inherit those patterns.

4. Feature Selection Bias

Sometimes developers unintentionally include features that correlate with protected characteristics.

Example

Using ZIP codes in a lending model could indirectly reflect race or income levels.

Even if race is not explicitly included, proxy variables can still create unfair outcomes.

5. Algorithmic Bias

Some algorithms may optimize for overall accuracy while ignoring fairness across groups.

Example

An AI model may achieve 95% accuracy overall but perform significantly worse for a minority population.

This demonstrates why fairness metrics matter alongside accuracy metrics.

Key Fairness Considerations

When evaluating fairness in an AI solution, organizations should consider several important areas.

Equal Treatment

AI systems should provide similar quality of service and outcomes across different demographic groups.

Example

A facial recognition system should work equally well for all skin tones and genders.

Avoiding Discrimination

AI should not unfairly disadvantage protected groups.

Example

A hiring system should evaluate applicants based on qualifications rather than demographic patterns found in historical data.

Inclusive Design

AI systems should be designed for diverse populations from the beginning.

This includes:

Diverse datasets
Diverse testing groups
Accessibility considerations
Multiple languages and accents
Cultural differences

Transparency and Explainability

Organizations should understand how AI systems make decisions and be able to explain those decisions when needed.

Example

If a loan application is denied, the organization should be able to explain the factors involved.

Explainability helps identify unfair behavior and improves accountability.

Continuous Monitoring

Fairness is not a one-time task.

AI systems should be continuously monitored because:

Data changes over time
User populations evolve
Biases may emerge after deployment

Organizations should regularly review model outputs and retrain models when necessary.

Trade-Offs in Fairness

Fairness in AI is complex because different definitions of fairness can conflict.

For example:

Maximizing overall accuracy may reduce fairness for smaller groups.
Equal outcomes across groups may require adjusting decision thresholds.
Removing sensitive attributes does not always eliminate bias.

There is often no perfect fairness solution, which is why ethical judgment and governance are important.

Microsoft’s Responsible AI Principles

Microsoft identifies fairness as one of six core Responsible AI principles.

The six principles are:

Fairness
Reliability and safety
Privacy and security
Inclusiveness
Transparency
Accountability

For the AI-901 exam, you should understand that fairness focuses on ensuring AI systems do not create unjust bias or discrimination.

Tools and Techniques for Improving Fairness

Organizations can reduce unfairness using several approaches.

Improve Data Quality

Use diverse and representative datasets
Remove biased or low-quality data
Balance underrepresented groups

Evaluate Fairness Metrics

Measure model performance across different groups instead of relying only on overall accuracy.

Example Metrics

False positive rates
False negative rates
Accuracy by demographic group

Human Oversight

Humans should remain involved in reviewing sensitive AI decisions.

Example

An AI hiring recommendation system might assist recruiters, but humans should make final hiring decisions.

Explainable AI

Explainability tools help organizations understand why models make certain decisions.

This can help detect hidden bias.

Responsible AI Governance

Organizations should establish policies, reviews, and ethical guidelines for AI development and deployment.

Real-World Example of Fairness

Scenario: AI-Based Hiring System

A company creates an AI model to screen resumes.

Potential Fairness Problem

Historical hiring data shows the company hired mostly men for technical roles.

The AI learns patterns associated with male candidates and begins ranking female candidates lower.

Possible Solutions

Use more diverse training data
Remove biased features
Audit model outputs regularly
Include human review
Test performance across demographic groups

This is a classic AI fairness scenario and aligns well with AI-901 exam objectives.

Azure and Responsible AI

Microsoft Azure AI Services and related AI platforms include Responsible AI guidance and tools to help developers:

Detect bias
Improve transparency
Monitor model behavior
Evaluate fairness metrics
Implement human oversight

Microsoft encourages organizations to adopt Responsible AI practices throughout the AI lifecycle.

Important AI-901 Exam Tips

For the exam, remember these key points:

Fairness means AI systems should avoid unjust bias and discrimination.
Bias often originates from training data.
High model accuracy does not guarantee fairness.
Diverse datasets help improve fairness.
Human oversight remains important.
Fairness is one of Microsoft’s six Responsible AI principles.
AI systems should be monitored continuously after deployment.
Transparency and explainability support fairness efforts.

Practice Exam Questions

Question 1

A company develops an AI system to screen job applicants. The system consistently ranks male applicants higher because historical hiring data mostly contains successful male candidates.

What is the MOST likely cause of this fairness issue?

A. Insufficient computing power
B. Biased training data
C. Excessive model transparency
D. Lack of cloud storage

Correct Answer

B. Biased training data

Explanation

The AI system learned patterns from historical hiring data that reflected past hiring bias. Because the training data was biased toward male candidates, the model inherited those unfair patterns.

This is one of the most common fairness problems in AI systems.

Why the Other Answers Are Incorrect

A. Insufficient computing power

Computing power affects performance and speed, not fairness.

C. Excessive model transparency

Transparency helps identify fairness problems rather than causing them.

D. Lack of cloud storage

Storage capacity does not create demographic bias in AI models.

Question 2

Which statement BEST describes fairness in AI?

A. AI systems should maximize profit for organizations
B. AI systems should make decisions without unjust bias
C. AI systems should eliminate all human involvement
D. AI systems should always make identical decisions for everyone

Correct Answer

B. AI systems should make decisions without unjust bias

Explanation

Fairness in AI focuses on preventing unjust discrimination and ensuring equitable treatment across different groups of people.

Fairness does not necessarily mean identical outcomes for everyone, but rather avoiding harmful or biased treatment.

Why the Other Answers Are Incorrect

A. AI systems should maximize profit for organizations

Profitability is unrelated to the Responsible AI principle of fairness.

C. AI systems should eliminate all human involvement

Human oversight is often important for maintaining fairness.

D. AI systems should always make identical decisions for everyone

Different circumstances may justify different outcomes. Fairness is about avoiding unjust bias.

Question 3

A speech recognition system performs poorly for users with certain accents because most training samples came from a single geographic region.

What fairness issue does this demonstrate?

A. Overfitting
B. Underrepresentation in training data
C. Excessive transparency
D. Encryption failure

Correct Answer

B. Underrepresentation in training data

Explanation

The training data lacked sufficient diversity, causing the model to perform poorly for underrepresented user groups.

Inclusive and representative datasets help improve fairness.

Why the Other Answers Are Incorrect

A. Overfitting

Overfitting occurs when a model memorizes training data rather than generalizing properly.

C. Excessive transparency

Transparency does not cause poor recognition accuracy for accents.

D. Encryption failure

Encryption relates to security, not fairness.

Question 4

Which Microsoft Responsible AI principle focuses on reducing bias and discrimination?

A. Accountability
B. Transparency
C. Fairness
D. Reliability and safety

Correct Answer

C. Fairness

Explanation

The Fairness principle focuses on ensuring AI systems do not unfairly disadvantage individuals or groups.

Why the Other Answers Are Incorrect

A. Accountability

Accountability concerns responsibility for AI systems and their outcomes.

B. Transparency

Transparency focuses on explainability and understanding AI decisions.

D. Reliability and safety

Reliability and safety focus on dependable and safe system operation.

Question 5

An organization removes race from a loan approval model, but the model still produces biased outcomes because ZIP code data indirectly reflects demographic patterns.

What does ZIP code represent in this scenario?

A. A fairness metric
B. A proxy variable
C. A transparency feature
D. A security control

Correct Answer

B. A proxy variable

Explanation

A proxy variable is a feature that indirectly correlates with sensitive attributes such as race, gender, or income level.

Even when protected attributes are removed, proxy variables can still introduce unfairness.

Why the Other Answers Are Incorrect

A. A fairness metric

Fairness metrics are measurements used to evaluate fairness.

C. A transparency feature

Transparency features help explain decisions, not indirectly encode demographic data.

D. A security control

Security controls protect systems and data.

Question 6

Why is human oversight important in AI systems that make sensitive decisions?

A. Humans can completely eliminate all bias
B. Humans can review and challenge potentially unfair outcomes
C. Humans increase automation speed
D. Humans reduce cloud costs

Correct Answer

B. Humans can review and challenge potentially unfair outcomes

Explanation

Human oversight helps organizations identify questionable or unfair AI decisions, especially in high-impact areas like hiring, healthcare, and finance.

AI systems should assist humans rather than fully replace judgment in sensitive scenarios.

Why the Other Answers Are Incorrect

A. Humans can completely eliminate all bias

Humans can reduce bias, but not completely eliminate it.

C. Humans increase automation speed

Human review usually slows processes rather than speeds them up.

D. Humans reduce cloud costs

Human oversight is unrelated to cloud pricing.

Question 7

An AI model achieves 98% accuracy overall but performs significantly worse for older adults than younger adults.

What does this scenario illustrate?

A. High accuracy guarantees fairness
B. Fairness and accuracy are always identical
C. An AI system can be accurate overall while still unfair
D. Transparency automatically prevents bias

Correct Answer

C. An AI system can be accurate overall while still unfair

Explanation

Overall accuracy can hide unequal performance across demographic groups. Fairness evaluations should measure outcomes for different populations separately.

Why the Other Answers Are Incorrect

A. High accuracy guarantees fairness

High accuracy does not guarantee equitable treatment.

B. Fairness and accuracy are always identical

These are different concepts and can conflict.

D. Transparency automatically prevents bias

Transparency helps identify issues but does not automatically eliminate them.

Question 8

Which action would BEST help improve fairness in an AI solution?

A. Limiting testing to a single user group
B. Using more diverse and representative training data
C. Hiding model outputs from reviewers
D. Reducing the amount of training data

Correct Answer

B. Using more diverse and representative training data

Explanation

Representative datasets improve an AI system’s ability to perform fairly across different populations and reduce bias caused by underrepresentation.

Why the Other Answers Are Incorrect

A. Limiting testing to a single user group

This increases the risk of bias and poor generalization.

C. Hiding model outputs from reviewers

Review and transparency help identify fairness issues.

D. Reducing the amount of training data

Less data often reduces model quality and fairness.

Question 9

Which of the following is an example of an unfair AI outcome?

A. A chatbot responding slowly during peak usage
B. A recommendation engine displaying duplicate products
C. A facial recognition system performing poorly for certain skin tones
D. A virtual machine running out of memory

Correct Answer

C. A facial recognition system performing poorly for certain skin tones

Explanation

Unequal performance across demographic groups is a classic fairness problem in AI systems.

This often results from insufficiently diverse training data.

Why the Other Answers Are Incorrect

A. A chatbot responding slowly during peak usage

This is a performance issue.

B. A recommendation engine displaying duplicate products

This is a recommendation quality issue.

D. A virtual machine running out of memory

This is an infrastructure issue.

Question 10

Why should AI systems be continuously monitored after deployment?

A. Fairness issues can emerge as data and user behavior change over time
B. AI systems never require updates after deployment
C. Monitoring removes the need for testing before deployment
D. Monitoring guarantees perfect fairness

Correct Answer

A. Fairness issues can emerge as data and user behavior change over time

Explanation

AI systems operate in changing environments. Data distributions, populations, and behaviors may evolve, creating new fairness risks after deployment.

Continuous monitoring is an important Responsible AI practice.

Why the Other Answers Are Incorrect

B. AI systems never require updates after deployment

AI systems often require retraining and adjustment.

C. Monitoring removes the need for testing before deployment

Pre-deployment testing remains essential.

D. Monitoring guarantees perfect fairness

No approach can guarantee perfect fairness in all situations.

Final Thoughts

Fairness is a foundational concept in Responsible AI and a critical topic for the AI-901 certification exam. Microsoft expects candidates to understand not only what fairness means, but also how bias enters AI systems and what organizations can do to reduce unfair outcomes.

As AI becomes more integrated into business and society, fairness is no longer optional—it is essential for building trustworthy and ethical AI solutions.

Go to the AI-901 Exam Prep Hub main page