Which step does most to reduce bias in an AI system?
Why D is the answer
Test outcomes across groups, and fix what you find
Overall accuracy can hide unfair results for one group.
In detail
- Test across groups, then fix what you find.
- Repeat the check after each change.
Use case
An illustrative scenario, not a real incident.
Overall accuracy looks good, but testing by group shows a problem for one, which the team fixes.
See it in different ways
flowchart LR A["Consent for data"] --> B["Balanced training data"] B --> C["Test across groups"] C --> D["Guardrails in production"] D --> E["Monitor and explain"]
mindmap
root((Trustworthy AI))
Consent
Before collecting data
Fairness
Test across groups
Transparency
Say that it is AI
Safety
NeMo Guardrailsflowchart LR S0["Split results by group"] --> S1["Compare"] --> S2["Fix"] --> S3["Test again"]
Key point: Test outcomes across groups.
Why the other options are wrong
- A. Remove the test set and rely on overall accuracy
- An overall figure can hide unfair results.
- B. Increase the model size until accuracy reaches its peak
- A bigger model does not remove bias.
- C. Hide the training data from the people who review it
- Hiding data prevents review.