Generative AI Leader — Practice Questions and Mock Exam

Google's Secure AI Framework (SAIF) is a conceptual framework providing industry-wide standards for addressing security and risk concerns of AI systems (docs.cloud.google.com/architecture/framework/security/use-ai-securely-and-responsibly). Its six core elements cover expanding security foundations to the AI ecosystem, contextualizing AI system risks, and adapting controls for faster feedback loops. SAIF does not define LLM technical specifications—that is model documentation. It is not a pricing framework—that is Cloud billing. It does not certify model performance benchmarks—that is evaluation tooling like ML.EVALUATE. SAIF is about responsible AI security governance.

Document AI is Google Cloud's platform for extracting structured information from documents including PDFs, scanned images, and handwritten forms (docs.cloud.google.com/document-ai/docs/overview). It offers processors for OCR, entity extraction, classification, and layout parsing—ideal for medical records processing. Speech-to-Text API converts audio to text, not documents. Cloud Vision API handles general image analysis but lacks document-specific parsing like key-value extraction. Natural Language API analyzes text sentiment and entities but cannot process PDFs or scanned documents. Document AI is purpose-built for document understanding workflows.

Human-in-the-loop (HITL) involves human oversight to review, validate, and correct AI outputs, ensuring accuracy and safety for critical decisions (developers.google.com/machine-learning/glossary/generative). Google's ML glossary defines HITL as a strategy for ensuring people shape, evaluate, and refine model behavior. HITL is not limited to initial training—it remains essential during deployment for ongoing quality assurance. It does not replace automated processing; it supplements it with human judgment at critical decision points. HITL does not mean manually typing every response—it uses efficient review workflows with confidence thresholds to route uncertain outputs for human review.

When selecting a foundation model for a chatbot handling text, image, and audio inputs, the most critical factor is whether the model supports those modalities (docs.cloud.google.com/docs/ai-ml/generative-ai/develop-generative-ai-application). Multimodal models like Gemini process text, images, video, and audio natively. Training data size in terabytes is an internal metric not exposed for selection. Parameter count indicates model size but not cross-modal capability. Diffusion architecture is for image generation, not multimodal chatbots. Modality support directly determines whether the model can accept required input types.