Free Generative AI Leader practice test: 30 questions with answers

Privacy and data protection cover how personal data is collected, used, logged and exposed. Fairness concerns differential outcomes, interpretability concerns explaining decisions, and latency concerns performance, so none of them addresses personal data exposure directly.

Safety filters screen inputs and outputs against policy categories and block or flag content that is harmful or disallowed. Response length, speed and weight compression are performance or formatting concerns and are not what a safety filter is for.

Prompt injection is an attack where untrusted text, typed by a user or hidden in a retrieved document, tries to change the instructions the model follows. Training speed, caching and token pricing are unrelated to that attack.

Routing by complexity sends easy requests to a cheaper model and keeps the costly model for work that needs it, which lowers average cost while protecting quality where it matters. Using the largest model everywhere raises cost, longer outputs add tokens, and disabling grounding can reduce accuracy.

A cross-functional group covers the legal, risk, data and business accountability that a regulated deployment needs, and it can set policies that apply across teams. A vendor, a single developer or individual staff cannot take that accountability for the business.

A representative evaluation set, agreed before launch, lets the team measure quality and compare changes objectively. Using production data without review creates privacy and quality risks, a pretrained model still needs testing on your own tasks, and feedback gathered after closure cannot improve the pilot.

Expected value with measurable outcomes and the practicality of data, risk and integration decide which use cases are worth pursuing first. Advertised parameter counts, the newest release and interface colour are not reasons to prioritise a business use case.

A route to a person for sensitive questions and continuous monitoring of answer quality and safety are standard controls for customer-facing assistants. Publishing without review, disabling safety settings and avoiding all logging each create a risk the business would have to carry.

Hallucination is the term for fluent output that is presented as fact but is not supported by the data or the world. Overfitting describes a training problem rather than a confident invented answer, and tokenisation drift concerns how text is split, not whether it is true. Latent bias is about systematic skew, which is a different risk.

Temperature controls how much the model's sampling favours less likely tokens, so raising it gives more varied output. Setting top-k to 1 makes output close to deterministic, and length caps and stop sequences do not change how varied the wording is.

Tokens are the units of text that a language model reads and produces, and they are what usage and context limits are measured in. A signed key is a credential, a pixel belongs to image input, and a document-level vector is an embedding rather than a token.

Foundation models are trained on large, broad datasets and then adapted through prompting, tuning or grounding to many tasks. Traditional models are typically trained for one narrow task on labelled data, and foundation models can generate content as well as classify it.

Language models generate text by predicting the next token, and sampling settings such as temperature mean that the same prompt can produce different outputs. Models do not reliably store documents verbatim, and summarisation is a standard use of generation.

Grounding ties answers to approved, current documents retrieved at query time, which addresses the wrong answer directly and keeps the answer current. Raising temperature makes errors more likely, removing the documents leaves the model with less to go on, and retraining from scratch is slow and costly for a documentation problem.

The context window is the total number of tokens the model can take into account for one request, including the prompt and any retrieved material. User counts, response time and allowed topics are different things that a team might configure or measure.

A multimodal model accepts more than one kind of input, such as text, images, audio or video, within the same model. The other options describe deployment or data-access patterns rather than the model's input types.

Clear instructions, relevant context and a worked example usually make outputs more accurate and better formatted. Length alone does not help, and many models handle several languages. Prompts remain one of the main levers after deployment.

Few-shot prompting places a small number of examples in the prompt so the model imitates their structure and style, without any retraining. Batch scoring is an execution pattern, and sweeps and gradient clipping are training-time techniques.

Retrieval-augmented generation fetches the current catalogue at query time and supplies it to the model, so answers stay current without retraining. Weekly retraining is expensive and still lags behind changes, lowering temperature does not add knowledge, and a smaller model with no context cannot know the catalogue.

Grounding and human review lower the rate of hallucinations, but no single control removes them completely, so high-stakes outputs still need checks. Hallucinations occur in managed and open models alike, and raising temperature tends to increase them.

Groundedness checks whether each claim in the summary is supported by the source, which is what faithfulness means. Word count, latency and vocabulary counts say nothing about whether the content is accurate.

A system prompt sets standing instructions such as the assistant's role, tone, scope and constraints, and it applies across the conversation. Credentials, the user's first message and training logs are different things.

Relevant examples, grounding in trusted documents and explicit format constraints all steer the model toward accurate, well-formed output. Maximum temperature makes factual answers less reliable, and removing instructions leaves formatting and scope to chance.

Pay-per-use LLM services generally bill for the tokens sent in and generated out by each request, so token volume is the cost driver to track. User counts, page counts and model file size may affect other things but are not what is billed per call.

Vertex AI provides access to and management of foundation models, and Model Garden helps teams choose and deploy models from its catalogue. Lifecycle policies, Cloud Armor and slot reservations are real Google Cloud features, but they do not build generative AI solutions.

Gemini for Google Workspace brings generative assistance into the productivity apps staff already use, with the organisation's data protections applied. Model Garden and BigQuery ML serve builders rather than end-user productivity, and the Vision API analyses images for labels rather than assisting with documents.

Vertex AI Agent Builder provides managed components for search, grounding and conversational agents over enterprise content, which reduces the custom work. Functions and storage are building blocks rather than a finished search and chat experience, and raw GPUs leave the team to build everything itself.

NotebookLM is designed around the user's own sources, so answers come from the uploaded material and can be traced back to it. Google Trends, Ads and Looker Studio do not answer questions from a user's own document set.

Google AI Studio is the quickest route to test prompts against Gemini and obtain an API key for prototyping, before a team moves the work onto a production platform. Logging, Looker and Cloud Run are useful in other stages of a system but do not provide that prototyping surface.

Drafting and summarising inside existing tools saves time without leaving the workflow, and summarising long threads and meeting notes is a common everyday use. Assistance does not remove the need for human review, billing is an administrator decision, and storing passwords is not a Workspace AI feature.