Free EXIN BCS Artificial Intelligence Foundation Sample Questions

Correct answer: B — Non-maleficence

• A. Beneficence focuses on promoting well-being and doing good. The scenario describes causing harm rather than failing to do good, so this principle is not the most directly violated.
• B. Non-maleficence means 'do no harm.' The excessive water consumption causing shortages for nearby communities is a clear harm, making this principle the most directly violated.
• C. Autonomy concerns individuals' ability to make independent choices. The issue here is resource depletion harming communities, not a restriction of personal autonomy.
• D. Explicability involves transparency and the ability to explain AI decisions. The harm from water usage is not due to a lack of explainability.

Correct answer: B — High efficiency in performing specific tasks

• A. Incorrect. Narrow AI (Weak AI) is designed for specific tasks and lacks the ability to generalize or transfer knowledge across different domains. This flexibility is a characteristic of General AI, not Narrow AI. Narrow AI operates within a limited scope and does not effectively apply knowledge from one domain to unrelated domains.
• B. Correct. Narrow AI excels at performing specific, well-defined tasks with high efficiency, accuracy, and speed. It is optimized for narrow applications such as image recognition, language translation, recommendation engines, or spam filtering. This focused specialization is the primary strength of Narrow AI systems.
• C. Incorrect. Complex, abstract reasoning is a hallmark of General AI or Strong AI, not Narrow AI. Narrow AI operates within predefined parameters and does not possess human-like reasoning capabilities. It lacks the ability to handle diverse, open-ended problems outside its trained task.
• D. Incorrect. Human-like emotional understanding and empathy are traits associated with Artificial General Intelligence (AGI) or Artificial Superintelligence (ASI), not Narrow AI. Narrow AI lacks consciousness or emotional intelligence; such abilities would require advanced social and cognitive intelligence that narrow systems do not possess.

Correct answer: C — A spam filter that classifies emails accurately

• A. Incorrect. A robot capable of any household chore would require general intelligence to adapt to many unrelated tasks, which is beyond Narrow AI's scope of performing a specific task.
• B. Incorrect. Engaging in open-ended dialogue implies a general understanding and flexibility characteristic of General AI, whereas Narrow AI is limited to a specific context or task.
• C. Correct. A spam filter is a classic example of Narrow AI because it is designed specifically to classify emails as spam or not spam, performing well within that narrow domain without general intelligence.
• D. Incorrect. Generating new scientific theories requires high-level reasoning and creativity, which are hallmarks of General AI, not the focused capabilities of Narrow AI.

Correct answer: B — Machine Learning

• A. Incorrect. Artificial Intelligence is a broader field encompassing the simulation of human intelligence, but it does not specifically refer to algorithms that improve through experience.
• B. Correct. Machine Learning is the study of computer algorithms that improve automatically through experience and by the use of data, matching the description exactly.
• C. Incorrect. Human Intelligence refers to the cognitive abilities of humans, not to computer algorithms or machine learning.
• D. Incorrect. The Scientific Method is a systematic approach to inquiry, not a term for algorithms that improve through experience.

Correct answer: B — A measure of data center energy efficiency.

• A. Incorrect. PUE is not related to AI model accuracy or performance over time. It is a metric used to evaluate data center energy efficiency, not AI model metrics.
• B. Correct. Power Usage Effectiveness (PUE) measures the energy efficiency of a data center by comparing total facility power to the power used by IT equipment. A lower PUE indicates higher efficiency.
• C. Incorrect. PUE does not measure carbon offsets or the contribution of renewable energy. It is an energy-efficiency ratio that compares total power to IT equipment power, not a carbon accounting metric.
• D. Incorrect. While cooling systems contribute to overall power consumption, PUE measures the overall energy efficiency of the entire data center facility, not just the cooling system.

Correct answer: B — Fairness

• A. Accountability refers to the assignment of responsibility for AI system outcomes. While important for governance, it does not specifically address equitable treatment or bias prevention.
• B. Correct. Fairness is the principle focused on ensuring AI systems treat individuals and groups equitably, avoiding the creation or reinforcement of bias, and preventing discriminatory outcomes.
• C. Transparency involves making AI systems understandable and explainable to stakeholders. It is crucial for trust but does not directly concern equitable treatment or bias.
• D. Robustness ensures that AI systems perform reliably under varying conditions. It relates to safety and resilience, not directly to fairness or bias.

Correct answers: A, D — Implement strict data minimization and anonymization techniques.; Train the model with more diverse faces to improve accuracy.

• A. Correct. Strict data minimization and anonymization reduce privacy risks and align with ethical principles like fairness and transparency by ensuring only necessary data is collected and processed.
• B. Incorrect. Replacing facial recognition with license plate recognition does not inherently address ethical concerns and may introduce other privacy or surveillance issues.
• C. Incorrect. Publishing a report after deployment is reactive and does not proactively mitigate ethical risks during development or implementation; prevention is key.
• D. Correct. Training with more diverse faces improves model accuracy and reduces bias, addressing ethical concerns related to fairness and discrimination proactively.

Correct answer: B — Based on risk to health, safety, and fundamental rights

• A. Incorrect. Revenue is not a primary factor in AI regulation; regulatory frameworks focus on the impact and risks of the AI system rather than the size or income of the developer.
• B. Correct. AI systems are often classified based on their risk to health, safety, and fundamental rights, as seen in frameworks like the EU AI Act. Higher-risk AI systems are typically subject to stricter requirements.
• C. Incorrect. The programming language or framework used does not determine regulatory classification. Regulation is concerned with the behavior, use, and impact of the system, not the technical tools chosen to build it.
• D. Incorrect. While server location may influence data protection laws (e.g., GDPR), it is not the primary basis for AI system classification. AI regulation usually depends on the risk and intended use of the system rather than where servers are physically located.

Correct answer: B — A sector-based framework relying on existing regulators

• A. While a single comprehensive AI Act like the EU's provides consistency, it is not the only approach and may lack flexibility across rapidly evolving AI use cases. Many regions prefer hybrid or sector-specific frameworks instead of relying solely on one overarching act.
• B. A sector-based framework leverages existing regulators in domains such as healthcare, finance, transport, and employment, allowing rules to be tailored to the specific risks of each domain. This is often seen as a flexible and proportionate way to regulate AI.
• C. Complete deregulation with industry self-governance is generally not considered appropriate for AI due to risks related to safety, fairness, privacy, and accountability. Most policy discussions emphasize some form of oversight rather than leaving governance entirely to industry self-regulation.
• D. Mandatory government approval for every AI product would be too burdensome and impractical, especially for low-risk applications. Regulation typically aims to be risk-based, applying stronger controls where potential harm is greater rather than requiring approval for all products.

Correct answer: B — To prioritize risks visually by plotting likelihood against impact.

• A. Incorrect. A risk matrix does not calculate exact probabilities; it is a qualitative or semi-quantitative tool for categorizing risks based on likelihood and impact, not precise numerical probability. In AI risk management, probabilities are often estimated qualitatively due to uncertainty.
• B. Correct. A risk matrix is used to prioritize risks by visually plotting their likelihood against their impact, helping organizations focus on the most critical risks first. This is a standard technique in AI risk management to assess and compare risks.
• C. Incorrect. While documenting risks is part of risk management, a risk matrix is not merely a linear list; it provides a structured, visual way to assess and prioritize risks. A linear list is useful for tracking but does not by itself indicate priority.
• D. Incorrect. Assigning monetary values to incidents is part of quantitative risk assessment, but a risk matrix typically uses qualitative or semi-quantitative measures (e.g., high/medium/low) rather than exact monetary values. Some AI risks, such as ethical or reputational harms, are difficult to quantify financially.

Correct answer: A — A process to evaluate an AI system's effects on individuals and society.

• A. Correct. An AI impact assessment evaluates an AI system's potential effects on individuals, groups, and society, identifying risks such as bias, privacy violations, safety issues, and broader social impacts. This is a key part of risk management in AI.
• B. Incorrect. A security audit focuses narrowly on technical vulnerabilities like access control and system hardening. While security is important, it does not cover the full range of ethical, societal, and operational risks addressed by an AI impact assessment.
• C. Incorrect. Financial analysis examines costs, benefits, and return on investment, which is useful for business decisions but does not assess ethical, social, or human impacts central to an AI impact assessment.
• D. Incorrect. A legal review for intellectual property compliance addresses copyright and patents, but it is only one possible legal aspect and does not capture the broader purpose of evaluating an AI system's overall effects on people and society.

Correct answer: A — Chatbot

• A. Correct. A chatbot is an AI application that simulates conversation with users using natural language processing (NLP) to understand and respond to queries. This matches the described system handling customer queries about orders and returns.
• B. Incorrect. An email client is used to send, receive, and manage emails, and does not typically provide conversational, automated question-and-answer support on a website.
• C. Incorrect. A database management system stores, organizes, and retrieves data. While it may support a chatbot's backend, it is not itself a customer-facing conversational AI system.
• D. Incorrect. A web server delivers web content and handles HTTP requests, but does not inherently understand natural language or generate conversational responses.

Correct answer: A — Labeled training data

• A. Correct. Supervised learning requires labeled training data, where each input example is paired with the correct output (label). This allows the model to learn the mapping from inputs to outputs and make predictions on new, unseen data.
• B. Incorrect. Unlabeled training data is used in unsupervised learning, where the model finds patterns or structures without predefined labels. In supervised learning, labels are essential for guiding the learning process.
• C. Incorrect. A reward function is a key component of reinforcement learning, where an agent learns by interacting with an environment and receiving rewards or penalties. Supervised learning relies on labeled examples, not rewards.
• D. Incorrect. Training data is fundamental to supervised learning; without it, the model has no examples from which to learn. Specifically, the data must be labeled to provide the correct outputs.

Correct answer: C — Supervised: labeled data; unsupervised: unlabeled data.

• A. Incorrect. Supervised learning does not always produce more accurate models; accuracy depends on the quality and quantity of labeled data, the algorithm, and the problem. In some cases, unsupervised learning may be more appropriate.
• B. Incorrect. Unsupervised learning does not inherently require more computational power than supervised learning; it depends on the dataset size, algorithm, and task complexity.
• C. Correct. Supervised learning uses labeled data where input-output pairs are provided, while unsupervised learning works with unlabeled data to find patterns or clusters without predefined labels.
• D. Incorrect. Supervised learning models are specifically designed for prediction tasks, such as classification and regression, by learning from labeled examples.

Correct answer: C — The compromise between fitting training data well and generalizing to unseen data.

• A. Incorrect. While model complexity and training data size can influence bias and variance, the tradeoff is not directly about balancing these two factors. It focuses on the error components of the model—bias (error due to simplifying assumptions) and variance (error due to sensitivity to data fluctuations).
• B. Incorrect. Training error and validation error are related to model performance, but this is not the definition of the bias-variance tradeoff. The tradeoff is about balancing underfitting (high bias) and overfitting (high variance) by adjusting model complexity.
• C. Correct. The bias-variance tradeoff describes the compromise between a model's ability to minimize error on training data (low bias) and its ability to perform well on unseen data (low variance). High bias leads to underfitting, while high variance leads to overfitting.
• D. Incorrect. Choosing between linear and non-linear models is a model selection decision that can affect bias and variance, but it does not directly define the bias-variance tradeoff. Both types can have different levels of bias and variance depending on their complexity.

Correct answer: D — Collaborative robots designed for safe work with human staff.

• A. Incorrect. Humanoid robots are designed to replicate human appearance and actions, but the scenario describes robots working collaboratively with humans in a warehouse, not necessarily humanoid in form. The key feature is safe physical collaboration, not human-like form or behavior.
• B. Incorrect. Traditional industrial robots are usually separated from human workers by cages or barriers due to safety risks. The question explicitly states the robots operate without safety barriers and adjust behavior near people, which is characteristic of collaborative robots, not industrial ones.
• C. Incorrect. Autonomous mobile robots (AMRs) can navigate using maps and move items, but the defining detail here is the safe interaction with human pickers in close proximity. AMRs may or may not have collaborative safety features; the scenario emphasizes collaboration and safety, which is the hallmark of collaborative robots.
• D. Correct. Collaborative robots (cobots) are specifically designed to work safely alongside human staff without safety barriers. They can sense nearby humans and slow down or stop to prevent injury, exactly as described in the scenario.

Correct answer: A — Adapting a pre-trained model on a domain-specific dataset to improve its performance.

• A. Correct. Fine-tuning involves taking a model that has already been pre-trained on a large general corpus and further training it on a smaller, domain-specific dataset. This process adapts the model's weights to perform better on particular tasks, styles, or domains, leveraging the knowledge from pre-training while specializing for the target application.
• B. Incorrect. Adjusting hyperparameters (like learning rate) is a part of training setup, but it does not constitute fine-tuning. Fine-tuning specifically refers to additional training on new data, not merely changing settings to avoid overfitting.
• C. Incorrect. Pruning reduces model size by removing unnecessary parameters for efficient deployment; it is a model compression technique. Fine-tuning, by contrast, involves further training on domain-specific data rather than just removing parameters.
• D. Incorrect. Evaluating a model's precision on a validation dataset is a performance assessment activity, not a training process. Fine-tuning modifies the model through additional training, distinguishing it from mere evaluation.

Correct answer: A — Fine-tune the model on a large corpus of legal texts and documents.

• A. Correct. Fine-tuning a pre-trained LLM on a large corpus of legal texts and documents is the most effective way to adapt it for generating legal briefs. This process leverages the existing knowledge of the model while specializing it for the legal domain, adapting its style, terminology, and structure.
• B. Incorrect. Training a new model from scratch is inefficient and time-consuming compared to fine-tuning. It would lack the general language understanding and contextual knowledge of a pre-trained LLM, and is unnecessary for most adaptation tasks.
• C. Incorrect. While prompt engineering can guide the model, it does not change the model's underlying behavior or deeply adapt it to legal brief generation. Repeating lengthy legal definitions in every query is not scalable or reliable, and is less effective than fine-tuning on relevant data.
• D. Incorrect. A rule-based filter can help with compliance or formatting, but it does not adapt the model to generate legal briefs. The model still lacks domain-specific learning unless fine-tuned or otherwise trained on legal data.

Correct answer: A — Volume, Velocity, Variety

• A. Correct. The original 'Three Vs' of big data are Volume (the amount of data), Velocity (the speed at which data is generated and processed), and Variety (the different types and sources of data). These three characteristics are widely recognized as the foundational definition of big data.
• B. Incorrect. While Veracity (data accuracy) and Value (business value) are sometimes added as additional Vs in extended definitions, they are not part of the original 'Three Vs'. The standard trio is Volume, Velocity, and Variety.
• C. Incorrect. Virtualization is not one of the 'Three Vs' of big data. Velocity and Volume are correct, but the third V should be Variety, not Virtualization.
• D. Incorrect. Visibility is not a recognized characteristic in the 'Three Vs' of big data. Variety and Volume are two of the correct Vs, but the third V should be Velocity, not Visibility.

Correct answer: B — To gain insights for improvement

• A. Incorrect. While big data does involve storing large volumes of data, storage alone is not its primary purpose. The main value lies in analyzing the data to support better decisions and actions.
• B. Correct. The primary purpose of big data is to analyze large and complex datasets to uncover patterns, trends, and insights that drive decision-making, operational improvements, and strategic advantages.
• C. Incorrect. Big data solutions typically require significant investment in infrastructure and tools, so cutting technology expenditures is not the primary purpose. Any cost savings come indirectly from insights gained.
• D. Incorrect. While big data can assist with regulatory compliance and reporting, this is a secondary benefit. The primary purpose is to analyze data and create business value.

Correct answer: B — Accuracy

• A. Incorrect. Timeliness describes how up-to-date the data is and whether it is available when needed, not whether the data is correct or error-free.
• B. Correct. Accuracy means the data correctly represents the real-world values it is intended to describe, without errors. Accurate data is essential for reliable AI outputs and model performance.
• C. Incorrect. Consistency refers to data being uniform across different records, systems, or time periods where it should match, not specifically to correctness or absence of errors.
• D. Incorrect. Completeness means all required data values are present and no essential information is missing, which is different from accuracy which focuses on correctness.

Correct answer: B — Completeness

• A. Incorrect. Consistency refers to data being uniform and free from contradictions across different records or sources. While important for data quality, it does not describe whether all required data is present.
• B. Correct. Completeness means the data contains all required values or records needed for the AI task, with no important gaps. Incomplete data can lead to biased, inaccurate, or unusable models.
• C. Incorrect. Timeliness means data is up to date and available when needed for decision-making or model use. It is important but does not indicate whether all necessary data elements are present.
• D. Incorrect. Validity means data conforms to expected formats, rules, or allowable values. It focuses on correctness against constraints rather than the presence of all required data.

Correct answer: C — Line chart

• A. Incorrect. A pie chart is used to show parts of a whole at a single point in time, not changes over time. It cannot effectively display a continuous trend such as stock price movement.
• B. Incorrect. A scatter plot can reveal patterns but is primarily used to show relationships between two variables. It is not the best choice for displaying a continuous trend over time; line charts are preferred for time-series data.
• C. Correct. A line chart is the standard choice for showing trends over time because it connects data points in sequence. This makes it easy to see rises, falls, and overall movement in stock prices over a continuous period.
• D. Incorrect. A bar chart compares discrete categories or values across time periods but is less effective than a line chart for showing continuous trends. Stock price movement over time is best visualized with a line chart.

Correct answer: C — To clean and transform raw data for effective modeling.

• A. Incorrect. While increasing dataset size can sometimes improve generalization, that is not the primary purpose of data preprocessing. Preprocessing focuses on improving data quality, not quantity.
• B. Incorrect. Converting unstructured data into a relational database is a specific data structuring task, not the general goal of data preprocessing. Preprocessing involves broader steps like cleaning, normalization, and encoding to make data suitable for modeling.
• C. Correct. Data preprocessing involves cleaning (e.g., handling missing values, removing duplicates) and transforming (e.g., normalization, encoding) raw data to make it suitable for modeling, which improves model accuracy and efficiency.
• D. Incorrect. Choosing an algorithm is part of model selection, not data preprocessing. Preprocessing is about preparing data, not selecting algorithms.

Correct answer: D — Investigate the missing data pattern to choose a strategy.

• A. Incorrect. Removing all rows with missing values can lead to significant data loss and may introduce bias if the missingness is not random. It is sometimes appropriate but not as a default first step.
• B. Incorrect. Replacing missing values with the average square footage is a simple imputation method but may not account for the underlying pattern or cause of missingness, potentially skewing results. You should first understand why data is missing.
• C. Incorrect. Applying a complex imputation algorithm immediately without understanding the missing data pattern can lead to inappropriate assumptions and poor model performance. The choice of method should depend on the missingness pattern.
• D. Correct. Investigating the missing data pattern (e.g., random, systematic) is crucial to determine the most appropriate strategy, such as imputation, deletion, or modeling the missingness itself. This helps reduce bias and ensures the data preparation suits the machine learning task.

Correct answer: B — To generate new content based on prompts.

• A. Classification is a task associated with discriminative AI, not generative AI. Generative AI focuses on creating new content rather than categorizing existing data.
• B. Correct. Generative AI produces new original content such as text, images, audio, or code based on patterns learned from training data and user input prompts. This distinguishes it from discriminative models.
• C. Data compression is a separate field in computer science; generative AI is not primarily used for compression but for content creation.
• D. Optimizing data retrieval speed relates to database management or indexing, not the core functionality of generative AI, which is to generate content.

Correct answer: A — A linear sequence of distinct phases is used

• A. Correct. The Waterfall methodology is characterized by a linear, sequential process where each phase (e.g., requirements, design, implementation, testing, deployment) must be completed before the next begins. This makes it a plan-driven approach with distinct phases.
• B. Incorrect. Short, time-boxed iterations are a characteristic of Agile methodologies such as Scrum, not Waterfall. Waterfall follows a sequential, non-iterative approach and does not deliver work incrementally.
• C. Incorrect. Requirements evolving based on ongoing feedback is associated with Agile approaches. In Waterfall, requirements are typically defined upfront and remain fixed throughout the project lifecycle.
• D. Incorrect. Self-organizing teams managing their own tasks is an Agile principle. Waterfall projects usually follow a hierarchical, top-down management structure with centralized planning and control.

Correct answer: A — Waterfall, because it suits fixed requirements and regulatory scrutiny

• A. Correct. Waterfall is ideal for projects with fixed, well-defined requirements and strict regulatory compliance. Its sequential phases, clear documentation, and formal sign-offs provide the traceability and control needed in healthcare.
• B. Incorrect. Agile is designed for evolving requirements and frequent feedback, but its iterative nature makes it less suitable for fixed requirements and heavily regulated environments where changes can complicate compliance validation.
• C. Incorrect. Spiral focuses on risk analysis and prototyping, which can be valuable for complex or uncertain projects. However, for a project with fixed requirements and strict compliance, it introduces unnecessary iterations and lacks the straightforward documentation of Waterfall.
• D. Incorrect. Prototyping is useful for early validation and user feedback, but it is not a comprehensive project management methodology. It lacks the structured phases and documentation required to meet regulatory standards in a clinical setting.

Correct answer: C — The chatbot makes racially insensitive comments from biased training data, harming company reputation.

• A. A temporary server outage is an operational or technical risk. It affects service availability and reliability but does not directly harm the company's public image or reputation.
• B. Using an outdated catalog is a data accuracy or functional risk. It may cause frustration or customer dissatisfaction, but it is not inherently a reputational risk unless it leads to broader trust issues.
• C. Racially insensitive comments due to biased data are a serious ethical and reputational risk. They can cause public backlash, loss of customer trust, legal exposure, and long-term brand damage.
• D. A modest increase in response time is a performance risk. While it may slightly impact user experience, it does not typically result in reputational harm.

Correct answer: C — Use predictive maintenance AI to analyze sensor data and forecast breakdowns

• A. Installing more sensors provides additional data but does not leverage AI to analyze existing data; it is a data collection improvement, not an AI opportunity.
• B. Training workers is a human-centered improvement and does not utilize AI to automate or enhance analysis; AI can detect patterns beyond human capability.
• C. Predictive maintenance AI can analyze sensor data in real-time to forecast breakdowns, reducing downtime and costs; this is a classic AI use case.
• D. Replacing machines is a capital investment, not an AI application; AI can improve existing processes without hardware upgrades.

Correct answer: A — Retrain the model and boost lifecycle monitoring

• A. Correct. The model is suffering from drift as customer behavior changes. Retraining with recent data adapts to new preferences, and boosting lifecycle monitoring detects future degradation early—key governance activities for maintaining model accuracy.
• B. Incorrect. Increasing storage capacity does not address model performance decline or outdated training data. Governance requires monitoring and retraining, not infrastructure scaling.
• C. Incorrect. Migrating to a different cloud provider does not solve the problem of outdated data or accuracy drop. The issue is data-related, not hosting-related.
• D. Incorrect. Reducing features may worsen accuracy and does not reflect changed preferences. The correct action is to update the model with new data and strengthen monitoring.

Correct answer: B — Customer service chatbot

• A. Incorrect. Predictive analytics systems forecast trends or outcomes based on historical data, but are not designed to handle real-time customer queries or provide conversational support.
• B. Correct. Customer service chatbots are specifically built to handle repetitive, high-volume queries and can operate 24/7, providing instant responses and reducing agent workload.
• C. Incorrect. Autonomous vehicle systems are used for self-driving navigation and driving tasks, which are unrelated to customer support interactions.
• D. Incorrect. Language translation platforms convert text or speech between languages, but they do not primarily handle customer support inquiries or provide automated question answering.

Correct answer: C — Reduction of human error via automation

• A. Incorrect. While AI may create new types of jobs, it also displaces many existing roles, particularly in repetitive or routine tasks. The net effect on job opportunities is not universally positive across all sectors; some sectors may see reduced demand for labor.
• B. Incorrect. AI can help reduce bias if designed and trained properly, but it cannot completely eliminate decision bias. AI systems can inherit or amplify biases present in their training data, algorithms, or human use, leading to biased outcomes.
• C. Correct. A key societal benefit of AI is reducing human error by automating repetitive, data-intensive, or rule-based tasks. This improves accuracy, safety, and efficiency in areas such as medical diagnostics, manufacturing quality control, and transportation.
• D. Incorrect. While AI can assist in creative tasks and generate novel outputs, it does not possess unlimited creative capabilities. AI creativity is constrained by its training data, algorithms, and prompts, and it lacks true human-like originality, intuition, or emotional depth.

Correct answer: B — Expand education in digital literacy and promote lifelong learning.

• A. Incorrect. Banning AI research would stifle innovation and economic growth, and it is impractical to protect jobs indefinitely. It also does not address the need to adapt workers for changing roles.
• B. Correct. Expanding digital literacy and promoting lifelong learning helps workers adapt to new roles and technologies. This supports reskilling and upskilling, enabling people to move into roles created or transformed by AI.
• C. Incorrect. Subsidizing businesses that avoid AI adoption would hinder competitiveness and innovation. It also fails to prepare the workforce for AI integration across industries.
• D. Incorrect. Mandating early retirement reduces the workforce but does not solve economic and social impacts. It limits opportunity rather than helping people transition into new careers.

Correct answer: A — Failing to disclose AI nature erodes trust and violates honest commerce.

• A. Correct. Transparency is a key ethical principle in AI. Failing to disclose the AI nature of a system can be deceptive, erode trust, and violate principles of honest commerce, as users have a right to know they are interacting with an AI.
• B. Incorrect. While response speed can affect customer experience, it is not a primary ethical concern related to transparency. The issue at hand is lack of disclosure, not speed of response.
• C. Incorrect. Difficulty with accents is a technical or accessibility issue, not directly tied to ethical transparency. It does not address the core concern of disclosing AI involvement.
• D. Incorrect. Reducing human agents and cost savings are business and service design considerations, not ethical concerns about transparency. The ethical focus is on honest disclosure of AI interaction.