Removing Stereotypes from Training Data to Mitigate Bias
A media company uses a generative AI model to automatically create image captions for user-uploaded photos. During quality assurance, testers discover that the model sometimes generates captions that include stereotypes based on gender and race, even when the photos do not contain people. For example, a photo of a kitchen produces captions like 'woman cooking,' and a photo of a sports car generates 'man driving.' The company wants to launch the feature soon but recognizes the reputational risk. They have a limited budget and need to implement a solution that reduces harmful stereotypes without overly restricting the captions' creativity. The team has access to the model's training data, which is a large public dataset of image-caption pairs. Which approach should the team prioritize?
Quick Answer
The correct approach is to filter the training data to remove or downweight pairs with stereotypes, then fine-tune the model. This directly addresses the root cause of bias by cleansing the dataset of harmful associations before the model learns from them, which is far more effective than applying superficial post-hoc filters that might also suppress valid creative captions. On the AWS Certified AI Practitioner AIF-C01 exam, this scenario tests your understanding of bias mitigation at the data preparation stage, a key concept in the Responsible AI domain. A common trap is to assume a post-processing filter is sufficient, but the exam emphasizes that modifying training data is the most direct and cost-effective way to remove stereotypes from training data without overly restricting model creativity. Memory tip: think “clean the source, not the output”—data filtering before fine-tuning is like weeding a garden before planting new seeds.
⚠ Common exam trap
AIF-C01 often tests the misconception that post-processing filters or larger models can solve bias, when the most effective and sustainable solution is to address the training data itself.
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
✓
Filter the training data to remove or downweight pairs with stereotypes, then fine-tune the model
The root cause of the stereotypical captions is the training data itself, which contains biased image-caption pairs (e.g., kitchens associated with women, sports cars with men). Filtering or downweighting these biased pairs and then fine-tuning the model directly addresses the source of the bias, reducing harmful stereotypes while preserving the model's generative creativity. This approach is cost-effective because it leverages the existing model and dataset, and it aligns with AWS responsible AI practices of data-centric debiasing. Post-processing or model replacement would not fix the underlying bias and could either over-restrict or fail to generalize.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Replace the generative model with a simpler classification model that only describes objects
Why it's wrong here
A classification model that only describes objects cannot generate captions, which is the core requirement; it outputs predefined labels rather than natural language, so it fails to produce any captions at all, let alone creative ones. This option is tempting because a classifier would reliably avoid stereotypes by ignoring social context, making it a correct choice if the goal were simply to tag objects without generating descriptive text.
- ✗
Use a different pre-trained generative model that is larger and more accurate
Why it's wrong here
A larger pre-trained model inherits the same biased image-caption correlations from its training corpus, so stereotypes persist regardless of accuracy gains. It is tempting because swapping models is a quick fix, and would be correct if the issue were caption quality or generalisation rather than demographic bias.
- ✓
Filter the training data to remove or downweight pairs with stereotypes, then fine-tune the model
Why this is correct
Filtering or downweighting stereotyped image-caption pairs before fine-tuning reduces the biased associations the model learns, directly targeting the training data's skew. This satisfies the low-budget, creativity-preserving constraint better than prompt filtering or output blocklists, which suppress symptoms without correcting learned associations.
- ✗
Add a post-processing filter that checks captions for known stereotype patterns and blocks them
Why it's wrong here
Pattern-matching filters only catch stereotypes already enumerated, blocking captions after generation while leaving the model's underlying bias intact and harming creativity. It is tempting because it is cheap and fast to deploy, and would be correct as a temporary guardrail alongside data or fine-tuning remediation.
Go deeper
Related to this question
About these practice questions
One of 862 original AIF-C01 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
2 more ways this is tested on AIF-C01
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. A company uses an AI system to automate loan approvals. The model uses demographic features and achieves high accuracy, but the company wants to ensure compliance with responsible AI guidelines. Which practice best balances performance and fairness?
hard- A.Use demographic features but with minimal monitoring
- B.Use a complex black-box model and rely on post-hoc explanations
- ✓ C.Remove sensitive attributes and monitor for proxy bias
- D.Optimize the model solely for accuracy on historical data
Why C: Removing sensitive attributes (e.g., race, gender) from the training data directly addresses fairness by preventing the model from explicitly using these features. However, simply removing them is insufficient; monitoring for proxy bias (e.g., zip code or income correlating with race) is critical to ensure the model does not inadvertently learn discriminatory patterns through correlated features. This approach balances performance by retaining predictive power from non-sensitive features while actively auditing for fairness violations.
Variation 2. A hospital uses an AI system to prioritize patients for organ transplant based on predicted survival rates. The system was trained on historical data that includes socioeconomic factors. A review reveals that the system systematically assigns lower priority to patients from lower-income neighborhoods, even when medical urgency is similar. The hospital's ethics board demands an immediate remedy. The data science team is small and must act quickly. What should the hospital do to address this fairness issue most effectively?
easy- A.Discontinue the AI system and have all prioritization done by a human committee
- ✓ B.Retrain the model with only medically relevant features, after removing socioeconomic factors and correlated proxies
- C.Apply a re-weighting penalty to boost priority for low-income patients
- D.Use a different model type, such as a random forest instead of gradient boosting, on the same data
Why B: The bias stems from socioeconomic features and their correlated proxies leaking into the model, so the most effective remedy is to retrain using only medically relevant features after removing socioeconomic variables and any correlated proxies (e.g., ZIP code, insurance type). This addresses the root cause of the disparate impact rather than masking it. It is also feasible for a small team acting quickly, since it is a data and feature-engineering change rather than a full system rebuild.
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official Amazon Web Services exam blueprint
This AIF-C01 practice question is part of Courseiva's free Amazon Web Services 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 AIF-C01 exam.