Free AIB-C01 practice test: 30 questions with answers

Turning a long transcript into a short summary is a summarization task that generative models handle well. Anomaly detection, image classification, and forecasting solve different problems.

Question answering over internal documentation with citations is a natural generative AI use case that improves productivity. The other options describe different AI problems.

Drafting and summarizing are language tasks where generative models add clear value with human review. Final approvals, legal review, and compensation decisions carry risks that require human accountability.

Creating images from text prompts is image generation, a generative modality. Speech recognition, tabular regression, and anomaly detection do not produce images.

Forecasting and fraud detection predict numbers or classes from historical patterns, which is classical predictive ML. Drafting, slogans, and summaries create or condense text, which suits generative models.

Lending decisions carry regulatory and fairness obligations, so criteria, risk assessment, and ownership must be set before deployment. Launching first, choosing by price, or maximizing model size skips those controls.

Clinical communication requires human oversight and transparency about AI involvement to reduce harm. Publishing prompts, unsupervised sending, and removing logs each undermine safety or accountability.

Testing across groups and collecting reports from users both surface bias that a single team might miss. Model size does not guarantee fairness, and removing demographic data from evaluation prevents measurement.

Traceable records of inputs, versions, sources, and approvals let the company reconstruct how an output was produced. Deleting content or keeping no records prevents that explanation.

Output filtering and escalation stop harmful content before it reaches customers. Removing feedback or relying only on training leaves gaps, and publishing offensive text is harmful.

Few-shot prompting is the cheapest first step and fine-tuning is considered only if measured quality falls short. Training a foundation model from scratch is far more costly than the need justifies.

Retrieval-augmented generation pulls the latest content into the prompt when needed, so updates do not require retraining. Monthly fine-tuning is slow and costly, and a static prompt goes out of date.

Latency and cost requirements usually favor smaller models when they meet accuracy goals. Larger models add latency and cost, and weekly batches fail the real-time requirement.

Scoped retrieval keeps each request limited to authorized documents, which prevents cross-client leakage. Shared fine-tuning or shared prompts mix confidential data across matters.

Accuracy on a labeled sample of real complaints shows whether the categories are assigned correctly. Dashboard counts, demo speed, and marketing names do not measure classification quality.

Users adopt tools that fit their workflow and that they were trained to use, and feedback loops make improvements visible. Forced use without training and removing alternatives often increases resistance.

A central team that sets guardrails and shared platforms, while units own use cases, scales adoption safely. Unmanaged tools and personal accounts expose data and create inconsistent controls.

Providing an approved alternative and clear policy addresses the data exposure directly, while controls limit recurrence. Waiting or ignoring the behavior leaves confidential data at risk.

Explanations and the ability to override build calibrated trust, and tracking outcomes shows whether the tool helps. Hidden logic, forced compliance, and removing people early all undermine trust.

A sponsor with a cross-functional governance group resolves trade-offs and owns risk decisions. Unstructured team choices and vendor sales roles do not provide that accountability.

ROI is measured by business outcomes compared against full cost, including model usage and people time. Token counts, slide counts, and model counts measure activity rather than value.

Usage-based pricing scales with request volume and with the tokens in each request, so these drive most recurring cost. UI color, wiki names, and room bookings have no effect on model spend.

Total cost per acceptable outcome combines price per token with how many tokens each option needs. A lower unit price can still cost more overall when more tokens are required.

Percentile latency under realistic load shows what users actually experience at peak. A single nighttime average hides tail delays, and laptop hardware and prompt length do not measure the target.

Scaling cost depends on how many requests each user makes and how large each request is. Office or hardware facts do not drive model consumption.

Generative AI is a subset of machine learning focused on producing new content, while many classic ML models classify or forecast. The other statements either reverse the relationship or make absolute claims that are false.

Hallucination refers to plausible but incorrect or unsupported generated output. Overfitting, labeling, and feature scaling are other ML concepts that do not describe fabricated fluent text.

Foundation models are large pretrained models adaptable to many tasks through prompting, retrieval, or fine-tuning. Rules engines, indexes, and dashboards do not match that definition.

Fine-tuning updates a pretrained model using task-specific examples. Sharding, caching, and load balancing are infrastructure techniques and do not change model behavior.

Models process text as sub-word tokens, and pricing is commonly based on input and output token counts. Hardware chips, identity signatures, and developer names are unrelated.