AIF-C01 Guidelines for Responsible AI Guidelines for Responsible AI Practice Questions

Benchmark datasets are standardized, curated datasets created specifically to evaluate models for bias and fairness, so they let the platform assess the model's outputs consistently with very little administrative effort.

User-generated content and moderation logs would need substantial manual labeling and review and do not provide a consistent fairness baseline. Internal content moderation guidelines are policy documents, not a dataset that can be used to measure bias in outputs.

Reusing generative AI output as one's own work without attribution is plagiarism, which is a recognized responsible AI and intellectual-property concern.

Toxicity refers to harmful or offensive content, hallucinations are fabricated but plausible-sounding outputs, and privacy concerns the handling of personal data. None of these describes copying generated text and presenting it as original work.

Gathering bite photos across different genders, ethnicities, and regions is aimed at preventing the model from working better for some groups than others, which reflects the fairness principle of responsible AI.

Explainability and transparency concern revealing how the model reaches its outputs, and governance concerns policies and oversight. None of these describes the act of assembling a representative, balanced dataset to avoid biased outcomes.

Amazon Bedrock Guardrails let the company define and enforce safety, content, and bias controls on generative AI responses, which supports responsible AI for an application that answers patients' medical questions.

Amazon Inspector scans for software vulnerabilities, Amazon Rekognition analyzes images and video, and AWS Trusted Advisor checks accounts against best practices. None of these enforces responsible AI controls on generated content.

When a generative AI model produces outputs that closely resemble copyrighted works, it raises the risk of intellectual property infringement, which is a legal risk of working with generative AI that companies must address through proper data curation and output review.

Model hallucination refers to generating plausible but factually incorrect content, not reproducing existing copyrighted material. Data drift describes changes in input data patterns over time. Overfitting means a model memorizes training data too closely and generalizes poorly, which is a different technical problem from IP infringement.