The correct pairing is Harm is that generated output can exhibit gender and racial bias and mitigation is to run fairness evaluations and apply post processing filters before release.

Bias in generative models is a well known risk for code assistants and it can surface in comments, identifier names, or generated explanations that reflect stereotypes. Running fairness evaluations during development helps identify disparate outcomes and harmful patterns across groups. Applying post processing filters and safety settings in Vertex AI reduces exposure to biased content before users encounter it. Using evaluation plus filtering provides proactive and measurable safeguards during development and review, which aligns with the goal of reducing harm before release.

Harm is that Copilot style suggestions are sometimes incomplete and mitigation is to rely entirely on user feedback to fix issues is not appropriate because incompleteness is a routine product quality issue rather than a responsible AI harm. Relying entirely on user feedback is reactive and does not provide a systematic mitigation during development or review.

Harm is that code produced by the model fails to compile and mitigation is to increase the model temperature in Vertex AI is incorrect because compilation failures are a functional quality problem rather than an ethical or safety risk. Increasing temperature raises randomness and would likely worsen consistency rather than improve compilability.

Harm is that generated code may contain copyrighted snippets and mitigation is to enable Cloud DLP scanning on responses is not suitable because Cloud DLP detects sensitive data such as PII or secrets and it does not detect copyright status. Addressing copyright concerns requires provenance controls, licensing policies, or specialized content provenance checks rather than DLP.

Full AWS Practitioner Certification Question