Fundamentals of Generative AI →hardMultiple Select

Best Approach to Debias Generative AI While Retaining Performance

Which THREE are valid methods to reduce bias in generative AI outputs?

⚠ Common exam trap

Google Cloud often tests the misconception that increasing model size or using a single language (like English) can solve bias, when in reality these actions can worsen bias by amplifying existing skews or introducing new cultural blind spots.

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

✓

Using a more diverse training dataset

Option C is correct because a more diverse training dataset reduces representational bias by exposing the model to a wider range of demographics, cultures, languages, and viewpoints, so the learned distribution is less skewed toward a dominant group. Option D is correct because safety filters (e.g., content moderation classifiers, toxicity detectors, or bias guardrails applied to prompts and outputs) can detect and block biased or harmful generations, mitigating bias at inference time. Option E is correct because prompt engineering that explicitly instructs the model to be fair, neutral, or inclusive (e.g., system prompts specifying balanced representation) steers generation toward less biased outputs without retraining. Option A is not correct because restricting prompts to English narrows the linguistic and cultural scope, which can amplify bias rather than reduce it. Option B is not correct because increasing model size alone does not remove bias and can even amplify biases present in the training data.

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

✗
Using only English prompts
Why it's wrong here
Restricting prompts to English does not remove bias; training data skew persists regardless of prompt language, and non-English users simply receive degraded service. It is tempting because standardising inputs appears to control variability, and English-only prompting would be reasonable for a deliberately English-scoped deployment, not for bias mitigation.
✗
Increasing model size
Why it's wrong here
Scaling parameters does not target the biased correlations embedded in training data; larger models can reproduce or amplify them with greater fluency. It is tempting because bigger models score higher on many benchmarks, and increasing size would be correct when the constraint is general capability or reasoning quality rather than fairness.
✓
Using a more diverse training dataset
Why this is correct
A more diverse training dataset directly addresses representation bias by exposing the model to broader demographic, cultural and linguistic variation during pre-training, reducing the likelihood of skewed associations in generated outputs. This satisfies the stem's requirement for a valid bias-reduction method, since bias frequently originates from unrepresentative or historically skewed training corpora.
✓
Using safety filters
Why this is correct
Safety filters evaluate prompts and responses against configured thresholds, blocking harmful or biased content before it reaches users. This provides an output-side control that reduces biased generations regardless of what the underlying training data contained.
✓
Applying prompt engineering to instruct the model to be fair
Why this is correct
Prompt engineering steers generative output through explicit instructions, so fairness directives can constrain tone and framing at inference time without retraining. It satisfies the stem's requirement for a valid bias-reduction method, though it only shapes surface behaviour and cannot correct biased training data or Microsoft Entra ID-governed access decisions.

Go deeper

Best Approach to Debias Generative AI While Retaining Performance

Why each option matters

Option-by-option breakdown

Related to this question