Three Effective Strategies to Mitigate Bias in AI Models
Which THREE strategies can help mitigate bias in an AI model? (Choose three.)
Quick Answer
The answer is applying fairness metrics during model evaluation, along with removing protected attributes from training data and ensuring diverse and representative training datasets. These three strategies directly address bias by first preventing the model from learning direct correlations with sensitive features like race or gender—a pre-processing technique that reduces direct discrimination, though proxy variables must still be monitored. Fairness metrics then quantitatively assess model outcomes across demographic groups during evaluation, catching disparities that raw accuracy might miss. On the Salesforce AI Associate exam, this question tests your understanding of bias mitigation as a lifecycle process, not a single fix; a common trap is assuming removing protected attributes alone suffices, ignoring that correlated proxies (e.g., zip code for race) can perpetuate bias. A useful memory tip is the "Three Ds": De-identify (remove protected attributes), Diagnose (apply fairness metrics), and Diversify (use representative data).
⚠ Common exam trap
Salesforce often tests the misconception that simply removing protected attributes or randomizing outputs is sufficient to eliminate bias, when in fact bias can persist through proxies and requires comprehensive fairness evaluation and diverse data.
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
✓
Remove protected attributes from training data
Removing protected attributes (e.g., race, gender) from training data reduces the risk of the model directly learning correlations with these sensitive features. This is a common pre-processing technique to prevent direct discrimination, though it may not eliminate indirect bias if correlated proxy features remain.
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 protected attributes from training data
Why this is correct
Removing attributes like race/gender can prevent direct discrimination.
- ✗
Focus training on majority group data for accuracy
Why it's wrong here
This would worsen bias against minorities.
- ✗
Randomize a portion of model outputs
Why it's wrong here
Randomization does not address underlying bias.
- ✓
Use diverse and representative training data
Why this is correct
Diverse data reduces underrepresentation.
- ✓
Apply fairness metrics during model evaluation
Why this is correct
Fairness metrics help detect and reduce bias.
Go deeper
Related to this question
About these practice questions
One of 753 original AI Associate 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 →
Same concept, more angles
1 more way this is tested on AI Associate
These questions test the same concept from different angles. Work through them to make sure you can recognise it however the exam phrases it.
Variation 1. Which TWO are best practices for mitigating bias in AI models?
hard- A.Using complex deep learning models.
- ✓ B.Using balanced training datasets.
- C.Removing all sensitive attributes.
- D.Training models on the most recent data only.
- ✓ E.Applying disparity analysis.
Why B: The correct answers are B and E. Using balanced training datasets (B) helps mitigate bias by ensuring the model is trained on a representative sample, reducing representation bias. Applying disparity analysis (E) helps identify and measure bias in model outcomes, allowing for targeted mitigation. Option A is incorrect because complex deep learning models can amplify bias and are harder to audit, not a best practice for mitigation. Option C is incorrect because while removing sensitive attributes may seem neutral, proxy variables can still introduce bias, and this approach does not address underlying biases in data. Option D is incorrect because training on the most recent data only does not guarantee unbiased data; historical bias or new biases may still exist.
JA
Written by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
This AI Associate practice question is part of Courseiva's free Salesforce 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 Associate exam.