Artificial Intelligence (AI) Fundamentals Questions and Answers
When an algorithm involves more than two classes, which technique is BEST suited to help a data scientist verify its accuracy?
Options:
Linear regression
Confusion matrix
K-NN
Answer:
BExplanation:
A confusion matrix is a table that cross-tabulates actual classes against predicted classes, and it extends naturally to any number of categories. For an algorithm with N classes it takes the form of an N-by-N grid in which the diagonal holds correct predictions and the off-diagonal cells show which classes are being mistaken for which. From it, a data scientist derives overall accuracy and per-class precision, recall, and F1-score, and can spot systematic confusion between particular pairs of classes. That diagnostic detail is what makes it the best choice for verifying accuracy in a multi-class setting, where a single accuracy percentage can conceal poor performance on smaller classes.
Linear regression is a modeling technique that predicts continuous numeric values. It is not an evaluation method, and it does not assess the correctness of class assignments.
K-nearest neighbors is itself a classification algorithm that labels a record based on the classes of its closest training examples. It produces predictions and therefore requires evaluation; it is not the tool used to evaluate them.
The question asks for a verification technique, and only the confusion matrix serves that role. It is a standard component of model validation, typically computed on held-out test data.
What is PowerShell?
Options:
Malware detection software
A team of sophisticated hackers
An advanced hacking tool
Answer:
CExplanation:
PowerShell is a command-line shell and scripting language created by Microsoft and bundled with Windows. Administrators use it to automate configuration and management tasks. The same capabilities give it deep access to operating system functions, the registry, networking, and remote systems, which makes it attractive to attackers. Because it is already installed and trusted, malicious scripts can run in memory without placing a conventional file on disk, a tactic known as fileless malware or "living off the land." Attackers use it for lateral movement, credential harvesting, downloading additional payloads, and evading signature-based antivirus. On that basis, security awareness material characterizes it as an advanced hacking tool.
PowerShell is not malware detection software. Detection products such as endpoint protection platforms are separate security controls, and PowerShell is more often an object of their monitoring than a component of them.
It is also not a team of sophisticated hackers. The expression describes people, whereas PowerShell is a piece of software. Advanced threat groups may exploit it, but the name does not refer to them.
Defenders mitigate the risk through script block logging, constrained language mode, application allowlisting, and monitoring for suspicious command patterns, recognizing that the tool is legitimate yet frequently abused.
How is the term "data owner" best characterized?
Options:
is defined as the organization aggregating data.
is defined as the organization collecting data.
can only be defined given context.
Answer:
CExplanation:
The term "data owner" can be defined only with reference to context. No single legal or technical definition applies universally. In a corporate setting it often denotes the business unit or executive accountable for a dataset and entitled to authorize its use. Under privacy law, such as the GDPR, the vocabulary differs, and the regulation speaks of data subjects, controllers, and processors, and not of owners. Many people also argue that the individual to whom personal data relates should be regarded as its owner. Contracts, industry practice, and jurisdiction can each assign ownership differently, and the party that collects, stores, or aggregates data may or may not hold ownership rights.
The statement that a data owner is the organization aggregating data is too narrow. An aggregator may compile information from many sources, but compilation does not automatically confer ownership, which may remain with the original source or be governed by agreement.
Equating the owner with the organization collecting the data fails for a similar reason. Collection may occur on behalf of another party, for example by a service provider acting for a client, and the collector's role does not settle who holds rights or accountability.
Because the meaning shifts with organizational, contractual, and regulatory circumstances, it is best characterized as definable only given context. Clarifying ownership explicitly is therefore a core task in data governance.
Which of the following represents a job type that is considered resistant to automation?
Options:
Teller
Cashier
Daycare worker
Answer:
CExplanation:
Daycare work is regarded as resistant to automation. The role demands social and emotional intelligence, responsiveness to children's changing needs, physical dexterity in unstructured and unpredictable settings, and continuous judgment about safety and wellbeing. Such abilities remain very difficult for machines to replicate. Parents also place a high value on human warmth and trust in the care of young children, which further limits the acceptance of automated substitutes. Studies of automation risk, including the well-known Frey and Osborne analysis of computerization, assign occupations built around personal care and caregiving a low probability of being automated.
Tellers perform structured transactions such as deposits, withdrawals, and balance inquiries. Those activities follow defined rules and are already handled by ATMs, mobile banking, and online platforms, which makes the role highly exposed to automation.
Cashiers process purchases, scan items, and take payment. Self-checkout kiosks, mobile payment, and cashierless store technology have replaced much of this routine, repetitive work and continue to expand.
The pattern across the three options is consistent: jobs made up of predictable, repeatable tasks are most vulnerable, while jobs dominated by human interaction and adaptable physical care are least vulnerable. Daycare worker is therefore the job type resistant to automation.
Classification models fall under which category of machine learning model?
Options:
Supervised learning
Reinforcement learning
Unsupervised learning
Answer:
AExplanation:
Classification belongs to supervised learning because the algorithm is trained on a labeled dataset in which every observation already carries a known category. During training, the model compares its predicted label with the true label, measures the error, and adjusts its parameters until the mapping from input features to output classes holds up on unseen records. Spam filtering, credit-risk grading, and image recognition all follow this pattern and all depend on historically labeled examples.
Reinforcement learning is a separate paradigm. An agent acts inside an environment and receives rewards or penalties, so there is no fixed answer key, only a feedback signal that shapes a policy toward cumulative reward. Unsupervised learning, by contrast, works on unlabeled data and looks for structure without guidance, for example by grouping similar records (clustering) or compressing features (dimension reduction).
Because classification needs predefined target classes supplied during training, it cannot be unsupervised, and because it learns from labeled examples rather than reward signals, it is not reinforcement learning. The presence or absence of labeled targets is the practical test for sorting any algorithm into these categories. Classification and regression are the two principal supervised tasks, separated only by whether the output is categorical or continuous.
Which of the following ethical challenges is AI use MOST likely to introduce?
Options:
Bias in content output
Unauthorized alteration of data
Data misuse
Answer:
AExplanation:
Bias in content output is the ethical challenge most directly tied to AI. Models learn from historical data, and when that data under-represents certain groups or encodes past discrimination, the model reproduces and frequently magnifies the pattern. Hiring screens that favor one demographic, lending models that price risk unevenly, and image generators that reinforce stereotypes are widely cited examples. Because outputs reach many users at scale and carry an appearance of objectivity, the harm spreads quickly and can be hard to trace to its origin.
Unauthorized alteration of data is an integrity violation. It can occur in any information system through tampering, malicious insiders, or weak access controls, and it is handled through security practices such as permissions, hashing, and audit logging. Nothing about it depends on AI.
Data misuse, meaning the use of information beyond its permitted purpose, is likewise a long-standing privacy and governance concern. AI can intensify it, but organizations faced the issue long before machine learning became common.
Output bias is the issue most specific to how AI learns and decides, so it is the one AI is most likely to introduce. Responsible practice responds with representative training data, fairness testing, human oversight, and documentation of known limitations.
The R² (coefficient of determination) notation can range from 0% to 100%. What type of relationship does an R² value of 0.845 indicate?
Options:
Neutral association
Positive linear association
Negative association
Answer:
BExplanation:
The coefficient of determination expresses the proportion of variance in the dependent variable that a regression model explains. A value of 0.845 means that roughly 84.5 percent of the variation is accounted for by the model, which is a strong fit. In the context of a simple linear model, that points to a strong linear relationship between the variables, and among the options offered, positive linear association is the appropriate characterization: as the independent variable rises, the dependent variable tends to rise in a consistent, near-linear manner.
A neutral association would correspond to an R² near zero, where the model explains almost none of the variability and the data points scatter without a discernible pattern. A value of 0.845 is far from that condition.
A negative association is not something R² can signal by itself. Because R² is the square of the correlation coefficient in the simple case, it is never negative, and its magnitude conveys strength but not direction. The sign of the slope or of the correlation coefficient reveals direction, and the question's emphasis on strength together with the answer choices supports the positive linear option.
Analysts should remember that a high R² does not prove causation and can be inflated by overfitting.
Which of the following can organizations use to detect possible anomalies in employee actions?
Options:
Deep neural networks
Confusion matrix
Behavioral AI
Answer:
CExplanation:
Behavioral AI is the approach organizations use to detect anomalies in employee actions. Commonly implemented as user and entity behavior analytics, it builds a baseline of normal activity for each person or role, covering login times, accessed resources, data volumes, and application usage, and then applies machine learning to flag meaningful deviations. An employee who suddenly downloads large quantities of files at an unusual hour, or accesses systems unrelated to their duties, produces a departure from baseline that the system highlights. Applications include insider threat detection, identification of compromised credentials, and monitoring for policy violations.
Deep neural networks are a general-purpose model architecture used for image recognition, language processing, and many other problems. They can be incorporated into an anomaly detection solution, but the architecture itself is not the purpose-built answer to behavioral monitoring, and the question calls for the tool defined by that function.
A confusion matrix is a table that summarizes a classifier's performance by comparing predicted and actual labels, showing true and false positives and negatives. It evaluates a model after the fact and does not detect anomalies in the workforce.
Behavioral AI matches the stated purpose of spotting unusual employee conduct. Because such monitoring involves personal data, organizations should pair it with transparent policies and privacy safeguards.
In an expert system, which component serves as the go-between connecting the human operator with the system's inference engine?
Options:
Knowledge base
Chatbot
User interface
Answer:
CExplanation:
An expert system consists of several cooperating components: a knowledge base that stores facts and rules, an inference engine that applies those rules to reach conclusions, and a user interface through which a person communicates with the system. The user interface is the intermediary between the human operator and the inference engine. It accepts questions, data, and commands from the operator, converts them into a form the engine can process, and then presents the engine's conclusions, and often the supporting explanation, in terms the operator understands. Without this layer, a non-specialist could not draw on the encoded expertise.
The knowledge base does not mediate between anyone. Its function is storage and organization of domain knowledge acquired from experts, and it is consulted by the inference engine rather than by the operator directly.
A chatbot is a particular style of conversational front end. Such a program can serve as the interface for some systems, but it is an implementation choice, not a named structural component of the classical expert system architecture. Many expert systems use forms, menus, or command screens instead.
Since the question asks for the architectural component that connects the operator with the inference engine, the user interface is the precise answer.
Which of the following can be considered an aspect of speech that natural language processing (NLP) may involve?
Options:
Translation
Morphology
Dictation
Answer:
BExplanation:
Morphology is the study of how words are built from smaller meaningful units such as roots, prefixes, and suffixes. NLP systems must analyze these components to understand language, because the form of a word signals its grammatical role and sense. Recognizing that "unhappiness" combines "un-", "happy", and "-ness," or that "running" and "ran" share a lemma, allows a system to tokenize text, normalize vocabulary, tag parts of speech, and handle languages with rich inflection. Morphological analysis sits alongside phonology, syntax, semantics, and pragmatics as one of the core linguistic layers that NLP addresses.
Translation and dictation are different in kind. Each is an end-user application assembled on top of several NLP capabilities. Machine translation draws on syntax, semantics, and statistical or neural modeling to convert text between languages, while dictation couples speech recognition with language modeling to produce written text from spoken input. Neither describes an intrinsic aspect of language itself; they describe tasks that a product performs.
The wording of the question, an "aspect of speech" that NLP may involve, calls for a linguistic property rather than a use case. Morphology is the only choice that names such a property, which is why it is the correct selection.
In deep learning, what is the MAIN advantage of using hierarchical feature learning?
Options:
It allows models to manually define features.
It reduces the need for labeled data.
It enables models to learn complex patterns from raw data.
Answer:
CExplanation:
Hierarchical feature learning is the mechanism by which deep networks build understanding in layers. The first layers detect elementary patterns, for example edges and color gradients in an image or phonemes in audio. Subsequent layers combine those outputs into progressively richer constructs such as textures, shapes, and object parts, and the deepest layers represent high-level concepts such as a face, a sentence's meaning, or a spoken word. Each stage uses the abstractions assembled by the one before it.
The main advantage is that the network can learn complex, non-linear patterns directly from raw data. Engineers no longer have to design features by hand, a laborious process that formerly limited traditional machine learning, because the network discovers useful representations itself during training. This is why deep learning has transformed image recognition, speech processing, and language modeling.
The statement that models manually define features is the reverse of the truth. Manual feature definition describes classical feature engineering, which hierarchical learning largely replaces.
The claim that it reduces the need for labeled data is also incorrect as a general matter. Deep models typically demand large quantities of data to train effectively, and hierarchical structure does not remove that requirement, although techniques such as transfer learning and self-supervision can ease it.
Learning complex patterns from raw inputs is therefore the principal benefit.
What is the MOST common reason enterprises adopt AI as a way to gain a competitive advantage?
Options:
Cost reductions
Risk remediation
Targeted marketing
Answer:
AExplanation:
Enterprises most often turn to AI to lower costs. Automating repetitive and labor-intensive work such as document processing, invoice matching, routine customer inquiries, and quality inspection reduces headcount hours, error rates, and cycle times. Savings of this kind are measurable, appear relatively quickly, and can be tied directly to a business case, which makes them the typical justification presented to executives when funding an AI program. Lower operating expense in turn lets a company price more aggressively or reinvest in growth, producing the competitive advantage.
Risk remediation is a genuine use case, covering fraud detection, compliance monitoring, and security analytics. It tends to be adopted by organizations with specific exposure rather than by the broad majority, and its benefits are harder to quantify because they consist of avoided losses.
Targeted marketing, including personalized pricing and advertising, delivers revenue benefits, but it is confined to customer-facing functions and depends on having rich customer data. Many enterprises do not have that data at the necessary scale.
Cost reduction is the common denominator across industries and functions, so it is the most frequently cited driver. Other motivations tend to appear as secondary objectives layered onto that foundation.
Which of the following options would help counter the loss of jobs due to AI?
Options:
Free healthcare
Creation of digital footprints
Continuous education
Answer:
CExplanation:
Continuous education is the option that helps counter job losses caused by AI. As automation takes over routine tasks, the occupations that remain or emerge call for new capabilities, including data literacy, oversight of AI tools, creativity, problem-solving, and interpersonal skills. Workers who keep learning through reskilling and upskilling programs can move into those roles, adapt as their existing duties change, and work alongside AI and not in competition with it. Employers, educational institutions, and governments increasingly promote lifelong learning for precisely this reason, because it addresses displacement directly by keeping the workforce's skills aligned with demand.
Free healthcare is a valuable social benefit that improves wellbeing and economic security, but it does not influence whether a person's job is automated or whether they have the skills to obtain new work. Its connection to employment loss from AI is indirect at best.
Creation of digital footprints refers to the trail of data that people leave through their online activity. Producing more of this data does not give anyone new skills or employment prospects, and it raises privacy concerns rather than easing displacement.
The remedy that tackles the cause, a gap between existing skills and the skills demanded by an AI-enabled economy, is ongoing learning. That is why continuous education is the answer, usually complemented by transition support and redesigned job roles.
Which of these clustering approaches does not scale well when applied to large data sets?
Options:
Fuzzy
Density-based
Hierarchical
Answer:
CExplanation:
Hierarchical clustering scales poorly because of the computation it requires. Agglomerative methods begin with every observation as its own cluster and repeatedly merge the closest pair, which requires a pairwise distance matrix whose size grows with the square of the number of records. Memory consumption is therefore quadratic, and typical implementations need between quadratic and cubic time depending on the linkage rule. Beyond tens of thousands of points the method becomes impractical, even though its dendrogram output is valued for exploring structure at multiple levels of granularity.
Density-based approaches such as DBSCAN identify regions of concentrated points and treat sparse regions as noise. With spatial indexing structures, they can approach near-linear performance, and they cope well with irregular cluster shapes and outliers on larger datasets.
Fuzzy clustering, exemplified by fuzzy c-means, assigns each record a degree of membership in several clusters. It operates iteratively in a manner similar to k-means, and its cost grows roughly in proportion to the number of records, so it remains workable at scale.
The deciding factor is the all-pairs distance bookkeeping inherent to hierarchical methods. When data volume is large, practitioners frequently switch to partition-based or density-based algorithms, or apply hierarchical clustering only to a sample or to preliminary cluster centers.
Which of the following is an example of a job that can now be performed by an AI?
Options:
Valet parking
Online Instruction
Commercial flying
Answer:
BExplanation:
Online instruction is a role that AI can now perform. Intelligent tutoring platforms present lessons, adapt difficulty to a learner's responses, answer questions in natural language, and grade assignments automatically. Conversational AI models can explain concepts, generate practice problems, and provide feedback at any hour for large numbers of students simultaneously. These capabilities are delivered entirely through digital channels, which suits the strengths of current AI: processing text, recognizing patterns in student performance, and producing instructional content.
Valet parking requires moving physically through a changing real-world environment, handling keys and vehicles, coordinating with people, and responding to unpredictable conditions. Although automated parking systems exist in controlled facilities, a general valet service performed independently by AI is not practical with present technology.
Commercial flying relies heavily on autopilot systems that manage cruise and some landing phases, but the role of airline pilot also demands judgment during emergencies, communication with air traffic control, and legal accountability. Regulations require certified human pilots, and public safety expectations make fully AI-operated passenger flights a distant prospect.
The distinguishing factor is how much of the work is informational as opposed to physical or safety-critical. Instruction conducted online is informational and well within the abilities of today's AI, so it is the example of a job that can now be performed by AI.
A class imbalance problem occurs when there are:
Options:
an unequal number of samples.
a lack of quality of samples.
a clustering balancing equation of samples.
Answer:
AExplanation:
A class imbalance problem arises when the classes in a dataset are represented by markedly different numbers of samples. A fraud dataset in which legitimate transactions outnumber fraudulent ones by several hundred to one is the standard illustration. The imbalance matters because most algorithms optimize overall accuracy, and a model can reach a very high accuracy score simply by predicting the majority class every time while failing to detect the rare class that the business actually cares about.
The second option, a lack of sample quality, describes a separate issue. Poor quality may involve noise, mislabeled records, missing values, or duplicates. A dataset can have excellent quality and still be imbalanced, and a balanced dataset can still be of poor quality, so the two concepts are independent.
The third option, a clustering balancing equation, is not a recognized concept. Clustering is an unsupervised technique for grouping unlabeled records, and it has no role in defining class distribution in labeled data.
Common remedies include resampling (oversampling the minority class or undersampling the majority class), synthetic generation such as SMOTE, cost-sensitive learning, and evaluation with precision, recall, and F1-score instead of plain accuracy.
What does the term XAI refer to?
Options:
The module explaining how the system reached its conclusion
The user interface of an AI system
An expert system's knowledge base
Answer:
AExplanation:
XAI stands for explainable artificial intelligence. The term covers methods and system components that make the reasoning behind an AI output understandable to people. In a classical expert system this takes the form of an explanation module that can display which rules fired and which facts supported a conclusion, answering questions like "why was this asked?" and "how was this result reached?" In modern machine learning the same goal is pursued through tools such as feature-importance scores, local surrogate models, and saliency maps that reveal which inputs drove a particular prediction.
The need for explainability arises from the opacity of complex models, especially deep neural networks, whose internal logic is difficult to inspect. Stakeholders in regulated fields such as lending, healthcare, and hiring must justify decisions, detect bias, and sustain trust, and an explanation that traces how the system arrived at its conclusion enables all three.
The user interface is simply the means by which a person interacts with the system. It governs presentation and input but not the justification of reasoning. An expert system's knowledge base holds the facts and rules that the system draws on, and it stores knowledge instead of explaining outcomes.
XAI therefore refers to the module or capability that explains how the system reached its conclusion.
What outcome characterizes a clustering method that is considered good?
Options:
Partitions
Intra-classes
Inter-classes
Answer:
BExplanation:
A good clustering result is judged by two criteria: cohesion within each cluster and separation between clusters. The first criterion, high intra-class similarity, means that members of the same cluster are close to each other according to the chosen distance or similarity measure, so within-cluster variation is small. The second criterion, low inter-class similarity, means different clusters remain clearly apart. Of the two outcomes named in the answer list, the one that characterizes a good method is strong similarity within classes.
Partitions describe how a clustering algorithm organizes data, namely by dividing it into non-overlapping groups as k-means does. Partitioning is a method category and a structure of the result, not a measure of quality. A poor clustering is also a partition.
Inter-classes would qualify only if the outcome were low similarity across clusters. As worded, the option does not state that distinction, and high similarity between classes would in fact indicate overlap and weak separation, which is the opposite of a good result.
Quality measures such as the silhouette coefficient, Davies-Bouldin index, and within-cluster sum of squares all operationalize this principle by rewarding tight clusters. Intra-class cohesion is therefore the characteristic outcome.