Free C1000-206 practice test: 30 questions with answers

Free-form emails and photos have no predefined schema, so they are unstructured data. Ledger rows and fixed-schema database columns are structured because they follow a predefined table format.

Semi-structured data, such as JSON or XML, uses tags or keys to organize records without a fixed table schema. The other options describe unstructured data (video, scanned images) or structured data (a fixed grid of values).

Predictive AI learns from historical data to estimate a future or unknown outcome, which is exactly what default-risk scoring does. Generative AI creates new content, RPA automates repetitive clicks and form entry, and a rule-based expert system follows hand-written rules rather than learning from data.

Generative AI produces new content such as text or images from a prompt or from attributes like these. Forecasting, summarizing past sales, and clustering customers do not create new written content.

Training uses example data to tune the model's parameters so it captures patterns that generalize. Deployment, storing data, and compressing a model for mobile are real steps, but none of them is the learning process itself.

Inference is the stage where an already trained model is applied to new, unseen inputs to produce predictions. Labeling happens before training, feature engineering is a data preparation step, and backpropagation is part of training rather than prediction.

Labeled historical outcomes with a yes/no target call for supervised classification. Clustering ignores labels, reinforcement learning learns from rewards in an interactive setting, and dimensionality reduction only compresses features.

Linear regression is a supervised method that predicts a continuous numeric value. Predicting a class label is classification, assigning unlabeled records to groups is clustering, and choosing actions for reward is reinforcement learning.

Discovering natural groups in data that has no labels is the job of unsupervised clustering. The supervised options assume labels already exist, and reinforcement learning needs an agent that acts and receives rewards, which this scenario does not describe.

Learning from rewards and penalties through trial and error is the defining feature of reinforcement learning. Labeled positions point to supervised learning, grouping states points to clustering, and fixed if-then rules are not learned at all.

Hidden layers sit between the inputs and the output and learn progressively more useful intermediate features through their weights. The other options describe storing labels, holding only the final output, or input preprocessing, none of which is the job of hidden layers.

Backpropagation calculates how each weight contributes to the loss, and gradient-based updates then adjust the weights to reduce error. The other two statements are false, because the output layer is kept for inference and hidden layers can contain any number of neurons.

A large language model generates text by repeatedly predicting the most likely next token given the tokens so far. Database querying, image comparison and intrusion monitoring are different tasks that a language model is not built to do.

Hallucination is when a generative model produces fluent, plausible output that is not grounded in fact, such as a fabricated citation. Overfitting, data drift and tokenization are real concepts, but they describe memorization, changing inputs and text splitting respectively.

Few-shot prompting places a handful of worked input-output examples in the prompt so the model can follow the pattern without any weight updates. Zero-shot prompting omits examples, while distillation and hyperparameter tuning both involve training work that the scenario rules out.

Retrieval-augmented generation fetches the current policy passages at question time and places them in the prompt, so the documents can change without retraining. The other options either require repeated retraining or do not generate answers from the documents.

Computer vision is the field that interprets images and video, so detecting pedestrians in camera frames is a direct example. Transcribing calls is speech recognition, while recommendations and demand forecasting work on purchase or meter data rather than images.

Fine-tuning continues training a pretrained model on labeled examples so its weights absorb the desired style. Changing sampling temperature only alters output randomness, while sharding and seed generation concern serving and reproducibility rather than learning.

Recruitment tools that affect who gets hired fall into the high-risk category under the EU AI Act, so the company should assess the risk and test for bias before rollout. Launching first, removing only names, or outsourcing fairness decisions each skip the evaluation that the risk calls for.

When training data under-represents some groups, the model learns weaker patterns for them, producing uneven accuracy. Drift, label-noise overfitting and prompt injection are real problems, but none of them explains performance that depends on who appears in the training images.

Auditing data for representation gaps and measuring results across subgroups are standard ways to find and reduce bias. Making a model larger does not correct skewed training data, and hiding subgroup results prevents the problem from being detected or addressed.

Prompt injection uses crafted input to override or leak the instructions a language model is given at runtime. Data poisoning happens during training, denial of service floods a service, and model inversion reconstructs training data rather than overriding instructions.

Data minimization, which means collecting only the personal data a stated purpose needs, is a core principle in privacy regulations such as the GDPR. Keeping records indefinitely, sharing raw data without contracts, and hiding sources all weaken compliance.

Telling an applicant the main reasons behind a decision is an explainability and transparency requirement. Data minimization, robustness and availability are important principles, but they do not ensure that a person can understand why a decision was made.

A first AI project works best with a clear task, usable historical data and a measurable outcome, since those allow a pilot to be tested and judged. The other options lack a clear goal, an owner or a proportionate scope.

Integration depends on clear ownership: who acts on each prediction, what follows, and who monitors outcomes over time. Replacing systems all at once, judging only training accuracy, or delaying staff training all ignore how the model will be used in practice.

A use case should promise real business value that can be measured, and it needs data that is available and good enough for the task. The project's name and the team's conference attendance say nothing about whether an AI solution will work.

Sizing the cost of the problem and checking that relevant data exists are the basic tests of a good use case. Buying a model first, following an announcement, or piloting everywhere at once skips the problem definition and feasibility checks.

Teaching agents clear criteria for using suggestions builds trust and consistent judgment, which drives adoption. Replacing systems without a transition, rewarding acceptance counts alone, and hiding accuracy data each undermine the team's ability to use the tool well.

Business value shows up in operational measures such as cycle time or cost per transaction, and in outcomes such as retention or conversion. Parameter counts and tool purchases describe the technology or spending rather than whether the business is better off.