Free Artificial Intelligence Essentials (AI|E) practice test: 30 questions with answers

The activation function applies the non-linear transformation that lets networks represent complex relationships. The loss function, optimizer, and dataset split each play different roles in training rather than transforming the neuron's output.

The loss function quantifies prediction error, and the training process uses it to adjust the model's parameters. Storing data, displaying charts, and compressing files are separate tasks that the loss function does not perform.

Tokens are the pieces of text that language models process, and their count affects cost and context length. Cryptocurrency, login credentials, and graphics processors are unrelated to this meaning of token.

Attention computes how much each element should attend to others, which is the core mechanism of transformers. Lookup tables, hand-written rules, and random selection do not learn relationships between parts of a sequence.

Data sensitivity, requirements, and integration needs determine whether a tool is suitable and safe to use. A vendor's logo colour, founder preferences, and office location do not reflect whether the tool fits the project.

A hallucination is fluent output that asserts things that are false or unsupported, which is why generated answers need verification. The other options describe interface effects, backups, and schedules rather than an error in content.

Fairness is a core principle of responsible AI, because systems can otherwise treat groups unjustly. Secrecy, ignoring risk, and removing oversight each work against accountability and the protection of people affected by the system.

Transparency lets people understand, question, and appeal decisions, which is essential for accountability. It does not remove the need for testing, improve speed, or replace governance.

Under-representation in training data is a primary cause of biased performance, because the model learns less about those groups. Interface colour, server counts, and typefaces do not shape the patterns the model learns.

A governance framework defines policies, responsibilities, and oversight so that AI is used consistently and within acceptable risk. It does not replace legal counsel, ban AI outright, or remove the reviews that protect the organization.

The EU AI Act classifies systems by risk and attaches obligations, such as documentation and oversight, that grow with the level of risk. Building codes, tax rules, and energy standards do not regulate AI systems by risk.

Human review of high-impact decisions and ongoing monitoring of performance are core safeguards that keep AI use accountable. Undocumented risks, hidden limitations, and use of personal data without a legal basis each undermine responsible use.

Data is the set of examples that a system learns from, while a model is the function it learns and then uses to predict. The other options confuse hardware, code, and documentation with the data and the learned model.

The 2012 ImageNet results showed that deep convolutional networks, trained on large datasets with GPUs, could outperform earlier methods by a wide margin. The transistor, personal computers, and TCP/IP are important computing milestones but not the deep learning turning point.

Supervised learning trains on inputs paired with known outputs, while unsupervised learning finds groupings or patterns without labels. The reverse statement, the data type limits, and the claim that they are identical are all incorrect.

Overfitting means the model has learned the training examples too closely, including their noise, so it generalizes poorly. Underfitting describes a model that is too simple, while convergence and normalization are training and preprocessing concepts rather than descriptions of this gap.

AI is the broadest field, machine learning is a part of it that learns from data, and deep learning is a subset of machine learning built on multi-layer neural networks. The statements that reverse this hierarchy or deny that machines can learn from data are incorrect.

AI aims to build systems that carry out tasks such as perceiving, understanding language, and reasoning, which otherwise need human intelligence. Networks, hardware speed, and data storage are supporting technologies rather than the goal of the field.

Recommendation systems predict what a user is likely to want based on past behaviour, which is what the streaming example does. The calculator, paper clip, and thermometer perform fixed functions without learning from user behaviour.

AI-based detection learns normal patterns and flags deviations, which helps analysts focus on the most relevant events. It supports rather than replaces firewalls and human review, and disabling passwords or printing reports does not describe threat detection.

Predictive maintenance models sensor patterns to forecast failures, which reduces downtime and unplanned repairs. Painting, holiday scheduling, and contract drafting are not manufacturing uses of AI in the sense the question describes.

Automatic emergency braking uses models on sensor data to recognize hazards and act, which is an AI-supported safety function. Paper maps, cones, and bicycle bells are useful safety aids but do not involve AI.

Adaptive practice and feedback on written work are established uses of AI in learning, with teachers still in the loop. Replacing teachers without oversight, guaranteeing exam results, and printing textbooks are not valid uses of AI for personalized learning.

Voice assistants depend on speech recognition models that turn audio into text before any further processing. Clocks, timers, and posters perform fixed functions without interpreting speech.

A prompt is the instruction and context that steer a generative model's response. Labels, passwords, and drivers are unrelated objects that do not guide the model's output.

Few-shot prompting shows the model a handful of examples so it can infer the expected format and style. Zero-knowledge proofs, brute-force search, and random sampling are unrelated methods.

Clear instructions, the necessary context, and a stated output format give the model what it needs to produce a useful answer. Vague requests, missing context, and unstructured lists of tasks each make the result less reliable.

Keeping confidential data out of tools that are not approved for it is the basic control against exposure. Pasting client records, including passwords, and ignoring the policy each increase the risk of leaking sensitive information.

Chain-of-thought and role prompting are established techniques that shape how a model reasons and responds. Random characters, disabling context, and printing prompts are not prompt engineering techniques that improve the model's output.

A system prompt is set by the developer to define the assistant's behaviour and constraints for the whole session. User messages, operating system warnings, and keyboard shortcuts are different things that do not define the assistant's role.