Free CT-AI practice test: 30 questions with answers

Narrow AI, also called weak AI, is designed for a specific task or limited domain. Option B describes general AI, option C attributes a consciousness that the narrow label does not imply, and option D is wrong because narrow AI often uses machine learning and where a system runs does not define its category.

The AI effect describes how established AI capabilities lose their label as AI once they become commonplace. The other three are real machine learning or human factors terms, but none of them describes that shift in how people classify a capability.

Computer vision is the AI field concerned with interpreting images and video, which is what scan analysis requires. Fuzzy logic, expert systems and search and optimisation are AI techniques too, but none of them is the main technology for extracting meaning from pixel data.

AI-based systems often have no complete specification, and they can change behaviour as they learn from new data, which are two of the main differences from conventional systems. Conventional systems are deterministic and testable, and the claim that AI is always more accurate is not true.

Reading free-text complaints, extracting product mentions and judging tone are core natural language processing tasks. Computer vision and robotics work with images and physical systems, and fuzzy logic can model graded categories but does not read or interpret language on its own.

Evolution covers how an AI-based system changes as it learns, so each retrained version needs checking against the agreed expectations. Autonomy, flexibility and interpretability are also quality characteristics, but they do not describe change over time caused by retraining.

Scores that consistently disadvantage a group that is under-represented in the training data are a bias problem. Autonomy, reward hacking and transparency are separate quality concerns, and a lack of transparency could accompany the bias but does not explain the unequal scores.

Interpretability concerns how well a person can understand cause and effect in a model, and explainability concerns giving understandable reasons for a specific decision. Accuracy is a different characteristic, AI-based systems are not transparent by default, and interpretability does not depend on using a deep network.

Acceptance criteria for AI systems need measurable targets on agreed data, linked to the business goal. Asking for judgement, demanding perfect accuracy on every input, or citing the newest technique does not give the testers anything they can measure.

The robot maximises the reward signal rather than the intended outcome, which is reward hacking and a side effect of poorly specified rewards. Autonomy, interpretability and concept drift are real quality concerns, but none of them explains the exploitation of the reward.

Learning a mapping from inputs to known outputs using labelled examples defines supervised learning. The other options describe learning without labels, learning from reward signals, and reusing a pretrained model, none of which matches labelled email data.

Strong results on training data with weak results on unseen data is the classic sign of overfitting. Underfitting would show poor results on training data too, and data augmentation and reinforcement learning describe training methods rather than a failure pattern.

Precision equals true positives divided by all predicted positives, so it answers how many flagged emails were really spam. Recall and specificity answer different questions, and F1 is the harmonic mean of precision and recall, not an arithmetic mean of accuracy and recall.

Hyperparameters, such as the learning rate, are set before training and control how training proceeds, while model parameters such as weights are learned from the training data. Testers do not normally tune parameters after deployment, and hyperparameters are not learned from test results.

When no reliable expected result exists for an input, a test oracle cannot be defined, which is the test oracle problem that is common in AI-based systems. The other options name real testing issues, but none of them describes a missing expected result.

Metamorphic testing uses metamorphic relations, such as the same sentiment after a name swap, to check outputs when no expected label is available. Mutation testing, exploratory testing and boundary value analysis are real techniques, but none of them relies on relations between related inputs.

Back-to-back testing runs the same inputs through two independent implementations and treats agreement as a pseudo-oracle, so differences are investigated. Metamorphic testing relies on relations between transformed inputs, mutation testing seeds faults, and exploratory testing does not compare outputs.

Fluent, confident answers that are not supported by the authoritative policy are hallucinations, a known failure mode of large language models. The other options describe memorised training data, changed relationships over time, and malicious user instructions, none of which the scenario indicates.

When test records also appear in training data, the test result no longer measures performance on unseen data, which is data leakage. Underfitting, concept drift and class imbalance are real issues, but identical records across the split point directly to leakage.

Good daytime results with failures at night point to a gap between the training data and the operational conditions. The other options contradict the evidence: daytime success argues against a network that is too small, and a learning rate or test set size problem would not produce this pattern.

Input data testing checks data quality such as missing values, duplicates and label consistency, and compares production inputs with training data. Inference speed, architecture review and interface contrast are useful activities, but they do not test the input data.

A shift in the distribution of input features without a change in the input-target relationship is data drift. Concept drift requires the relationship itself to change, which the scenario rules out, and leakage and overfitting describe different problems.

Choosing a model by its test score means the test set has guided selection, so the final score no longer estimates performance on truly unseen data. A separate validation set should be used for selection, with the test set kept for the final evaluation.

Adversarial examples are crafted perturbations that exploit model sensitivity, which makes them central to robustness testing. Boundary cases, synthetic samples and held-out instances are real terms, but they describe input limits, class balancing and evaluation data respectively.

Equal overall accuracy can hide unequal outcomes for subgroups, so the tester should report the disparity and investigate its causes in the data and features. Accepting the model, retraining on a removed test set, or ignoring the gap would each hide the problem.

Relying on a watermark rather than the object means the model learned a spurious correlation that may not hold in real use, so measured accuracy can be misleading. The other options misread the evidence: a shortcut is not good training, and the problem lies in the feature rather than in the dataset size or network depth.

Unit tests with known inputs and expected outputs check that preprocessing code behaves correctly before any training is run. Accuracy targets, drift checks and architecture choices are important, but they are assessed at other levels and are not what these unit tests verify.

Fitting a tiny fixed batch is a common sanity check: if the code cannot learn that batch, the pipeline likely has a bug. Regression, load and security tests are valid test types, but none of them checks whether the training code can learn at all.

Reproducibility depends on controlling randomness and recording everything that defines a run, including code, data, libraries and hyperparameters. Re-running is needed to check reproducibility, automatic upgrades introduce new variation, and deleting checkpoints removes evidence.

Stochastic training produces small run-to-run variation, so tests should check results against an agreed tolerance rather than exact equality. Demanding exact matches would fail healthy runs, while skipping the check or comparing only training loss leaves the quality question unanswered.