AIF-C01 Question #140: Real Exam Question with Answer & Explanation

Question

Which technique can a company use to lower bias and toxicity in generative AI applications during the post-processing ML lifecycle?

Options

• AHuman-in-the-loop
• BData augmentation
• CFeature engineering
• DAdversarial training

Explanation

Human-in-the-loop (HITL) is correct because it involves incorporating human review and feedback after the model generates outputs, allowing people to flag, filter, or correct biased or toxic responses before they reach end users - making it a post-processing technique. This ongoing human oversight is specifically designed to monitor and mitigate harmful AI outputs during deployment.

Why the distractors are wrong:

• Data augmentation (B) is a pre-processing technique that increases training data diversity to improve model performance, not a post-processing bias/toxicity filter.
• Feature engineering (C) is also a pre-processing step focused on selecting and transforming input variables to improve model training - it doesn't address outputs.
• Adversarial training (D) occurs during model training by exposing the model to challenging examples to improve robustness; it is not a post-processing lifecycle technique.

Memory tip: Think of the ML lifecycle in stages - pre, during, and post training. "Human-in-the-loop" is the only option where a human reviews the output after generation, making it the clear post-processing choice. If you remember "post = a human checks the result," you'll lock in this answer.

No community discussion yet for this question.