Generative AI Leader Fundamentals of Generative AI Practice Question
Which THREE of the following are common techniques to reduce harmful biases in generative AI models? (Choose three.)
⚠ Common exam trap
Google Cloud often tests the distinction between hyperparameter tuning (like temperature) and actual bias mitigation techniques, so candidates mistakenly think lowering temperature reduces bias when it only affects output randomness.
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 reinforcement learning from human feedback (RLHF) with a reward model that penalizes biased or unfair outputs.
Option A is correct because RLHF with a reward model that explicitly penalizes biased or unfair outputs directly optimizes the generative model's policy toward fairness-aligned behavior, using human preference data to shape the reward signal. Option B is correct because curating diverse and balanced training datasets—deliberately overrepresenting underrepresented groups—counteracts sampling bias and spurious correlations learned from skewed data, which is a foundational bias-mitigation technique. Option D is correct because adversarial training that removes protected attribute information from hidden representations (e.g., via adversarial classifiers or fair representation learning) reduces the model's ability to encode and exploit sensitive attributes. Option C does not belong because lowering the temperature only makes sampling more deterministic; it changes randomness, not the underlying learned biases, and can even amplify the most probable (biased) output. Option E does not belong because a legal review of outputs is a post-hoc compliance check, not a technique that reduces bias within the generative model itself.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✓
Use reinforcement learning from human feedback (RLHF) with a reward model that penalizes biased or unfair outputs.
Why this is correct
RLHF directly targets bias by training a reward model that assigns lower scores to unfair outputs, then optimising the generative model against it. This satisfies the stem's requirement for a bias-reduction technique, since the penalty signal actively steers the model away from biased generations during fine-tuning rather than merely detecting them afterwards.
- ✓
Curate diverse and balanced training datasets that overrepresent underrepresented groups.
Why this is correct
Curating diverse, balanced datasets directly targets bias at its source: skewed training data. Overrepresenting underrepresented groups counteracts the statistical dominance of majority populations, reducing the model's tendency to learn and amplify those patterns. This satisfies the stem's requirement for a common bias-reduction technique during data preparation.
- ✗
Decrease the model's temperature parameter to make outputs more deterministic.
Why it's wrong here
Temperature controls sampling randomness, not the training data distribution that encodes bias; lowering it yields deterministic but equally biased outputs. It is tempting because temperature tuning is a genuine inference-time control, and it would be the right lever when the goal is reproducible, consistent responses rather than fairness.
- ✓
Apply adversarial training to remove protected attribute information from hidden representations.
Why this is correct
Adversarial training directly satisfies the debiasing requirement by training an adversary to predict protected attributes from hidden representations while the main model learns to prevent that prediction, stripping encoded demographic information. This targeted removal of protected attribute information from internal representations reduces bias-driven outputs, making it a recognised fairness technique for generative AI.
- ✗
Conduct a legal review of all generated outputs before release.
Why it's wrong here
Legal review screens outputs for compliance and liability, not statistical bias in model behaviour, and cannot scale to every generated response. It is tempting because human review is a real mitigation control, and it would be correct where regulatory or contractual sign-off on published content is the actual requirement.
Go deeper
Related to this question
About these practice questions
Courseiva writes every Generative AI Leader question from scratch — 1,008 in total, each with an explanation and a wrong-answer breakdown. None are copied from real exams or dumps. Learn why practice questions differ from exam dumps →
JA
Written by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
This Generative AI Leader practice question is part of Courseiva's free Google Cloud 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 Generative AI Leader exam.