I85PREP
⌘ K
TAISE study materials
TAISE study resources

TAISE practice questions

One question at a time. Build your TAISE understanding.

First 50 of 190 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI risk, accountability and responsible use

For the explicitly supplied NIST AI RMF1.0 conceptual core, which listed function name is not one of Govern, Map, Measure and Manage?

  1. Govern
  2. Map
  3. Measure
  4. Respond Correct answer.

Explanation

The NIST AI RMF 1.0 core functions are: GOVERN, MAP, MEASURE, and MANAGE. 'Respond' is a core function in the NIST Cybersecurity Framework (CSF), not the AI RMF.

Question 2AI risk, accountability and responsible use

ISO/IEC 42001 is best described as:

  1. The first certifiable international standard for an AI Management System (AIMS) Correct answer.
  2. A technical standard specifying neural network architecture requirements
  3. A regulation governing AI liability across EU member states
  4. A cryptographic standard for protecting machine learning model weights

Explanation

ISO/IEC 42001:2023 establishes requirements for an Artificial Intelligence Management System (AIMS), providing a certifiable framework for governing the design, development, deployment, and monitoring of AI systems responsibly within an organization.

Question 3AI risk, accountability and responsible use

MITRE ATLAS (Adversarial Threat Landscape for AI Systems) is primarily used to:

  1. Catalog and communicate adversarial machine learning tactics, techniques, and procedures (TTPs) Correct answer.
  2. Define regulatory compliance requirements for AI in financial services
  3. Measure model bias across demographic subgroups
  4. Certify cloud providers for AI workload hosting

Explanation

MITRE ATLAS is a knowledge base of adversarial ML TTPs modeled after MITRE ATT&CK. It helps organizations understand how attackers target AI systems and enables red teams, threat modelers, and defenders to apply a common threat taxonomy for AI security.

Question 4AI risk, accountability and responsible use

The CSA AI Controls Matrix (AICM) contains how many control objectives across how many domains?

  1. 100 control objectives across 10 domains
  2. 156 control objectives across 12 domains
  3. 243 control objectives across 18 domains Correct answer.
  4. 300 control objectives across 20 domains

Explanation

The CSA AI Controls Matrix (AICM) v1 specifies 243 control objectives distributed across 18 security domains, extending CSA's Cloud Controls Matrix (CCM) to address AI-specific risks including model security, data poisoning, and supply chain threats in cloud-based AI systems.

Question 5AI risk, accountability and responsible use

Which NIST AI RMF function is responsible for establishing organizational policies, roles, and accountability structures for AI risk?

  1. MAP
  2. MEASURE
  3. MANAGE
  4. GOVERN Correct answer.

Explanation

The GOVERN function is the foundational cross-cutting function of the NIST AI RMF. It establishes organizational practices, policies, processes, and accountability structures that enable all other risk management functions (MAP, MEASURE, MANAGE) to operate effectively.

Question 6AI risk, accountability and responsible use

The NIST AI RMF Generative AI Profile (NIST AI 600-1) introduces which AI-specific risk not covered in the original AI RMF 1.0?

  1. Data quality issues in traditional ML pipelines
  2. Bias in classification model outputs
  3. Confabulation (hallucination) as a named AI risk category unique to generative AI Correct answer.
  4. Security vulnerabilities in on-premises ML infrastructure

Explanation

NIST AI 600-1 (the Generative AI Profile) expands the AI RMF to address GenAI-specific risks, explicitly naming confabulation (the production of false but plausible outputs) as a unique risk category that requires dedicated measurement, governance, and mitigation strategies.

Question 7AI risk, accountability and responsible use

ISO/IEC 42001 Annex A controls specifically address which AI-related concern?

  1. Cryptographic key management for model weight protection
  2. Network segmentation requirements for GPU clusters
  3. AI system impact assessment, responsible use objectives, data quality management, and human oversight Correct answer.
  4. Open-source license compliance for third-party AI libraries

Explanation

ISO 42001 Annex A contains AI-specific controls organized around responsible AI governance: objectives for responsible AI use, AI risk and impact assessments, data quality and documentation, human oversight mechanisms, and transparency requirements-extending traditional ISMS controls into the AI domain.

Question 8AI risk, accountability and responsible use

Which of the following describes a 'shadow AI' risk in an enterprise?

  1. AI models that run exclusively in dark data centers with no physical lighting
  2. AI systems that operate in shadow mode alongside human reviewers without making live decisions
  3. Employees using unauthorized AI tools or APIs that bypass organizational security, privacy, and governance controls Correct answer.
  4. Backup copies of AI model weights stored in encrypted offline storage

Explanation

Shadow AI refers to employees using AI tools (e.g., consumer ChatGPT, unauthorized SaaS AI features) without IT or governance approval. This creates data leakage risks (sensitive data submitted to external models), compliance violations, and unmonitored AI decision-making outside organizational controls.

Question 9AI risk, accountability and responsible use

Which document published by CSA specifically maps AI control objectives to ISO 42001, ISO 27001, NIST AI RMF, and BSI AIC4?

  1. Cloud Controls Matrix (CCM) v4
  2. CSA STAR for AI Assessment Template
  3. CAIQ for AI (Consensus Assessment Initiative Questionnaire for AI)
  4. AI Controls Matrix (AICM) v1 Correct answer.

Explanation

The CSA AI Controls Matrix (AICM) v1 is a vendor-agnostic framework containing 243 control objectives across 18 domains, with mappings to ISO 42001, ISO 27001, NIST AI RMF 1.0, and BSI AIC4, enabling organizations to demonstrate compliance across multiple frameworks simultaneously.

Question 10AI risk, accountability and responsible use

Which of the following AI risk assessment activities aligns with the NIST AI RMF 'MAP' function?

  1. Deploying automated dashboards to track model performance KPIs in production
  2. Updating organizational AI policies based on lessons learned from an incident
  3. Approving the AI system for deployment after all risk treatments are in place
  4. Categorizing the AI system's intended use, affected stakeholders, and potential negative impacts Correct answer.

Explanation

The NIST AI RMF MAP function involves classifying the AI system's context: its intended use, deployment environment, affected stakeholders, and potential harms. This risk contextualization activity provides the foundation for subsequent MEASURE and MANAGE activities.

Question 11AI risk, accountability and responsible use

Which of the following is a key principle of the NIST AI RMF's GOVERN function regarding organizational accountability?

  1. Accountability for AI risks is delegated solely to the data science team
  2. Organizations should avoid formalizing accountability to retain flexibility in AI deployment
  3. Risk accountability is limited to the AI product owner for each individual system
  4. Clear roles, responsibilities, and accountability structures for AI risk management must be established organization-wide Correct answer.

Explanation

The NIST AI RMF GOVERN function requires organizations to establish explicit policies, processes, roles, and accountability structures that span the entire organization-not just the AI development team. This ensures systemic risk governance rather than siloed responsibility.

Question 12AI risk, accountability and responsible use

Which of the following is the MOST important first step when developing an AI Impact Assessment (AIIA) for a new AI project?

  1. Selecting the programming language and ML framework for implementation
  2. Choosing the cloud region for model training to optimize latency
  3. Benchmarking model accuracy against industry-standard datasets
  4. Identifying the AI system's intended purpose, affected populations, and potential harms before development begins Correct answer.

Explanation

An AI Impact Assessment begins with defining the system's purpose, scope, and affected stakeholders, then identifying potential harms (physical, psychological, financial, reputational, societal). This scoping step must occur before development to influence design choices and determine required risk mitigations.

Question 13AI risk, accountability and responsible use

Which MITRE ATLAS technique describes an attacker using the target model's own predictions to infer properties of its training data?

  1. AML.T0012 - Valid Accounts
  2. AML.T0024 - Exfiltration via ML Inference API
  3. AML.T0036 - Membership Inference Correct answer.
  4. AML.T0048 - Backdoor ML Model

Explanation

MITRE ATLAS technique AML.T0036 (Membership Inference) describes attackers querying a model's API and analyzing confidence scores to determine whether specific records were in the training dataset-a privacy attack that can expose sensitive personal information used to train the model.

Question 14AI risk, accountability and responsible use

In the NIST AI RMF, which activity involves prioritizing and addressing identified AI risks through risk treatment options such as avoidance, mitigation, transfer, or acceptance?

  1. MAP
  2. MEASURE
  3. MANAGE Correct answer.
  4. GOVERN

Explanation

The MANAGE function of the NIST AI RMF involves selecting and implementing risk treatment options (avoid, mitigate, transfer, accept) based on the risks identified in MAP and measured in MEASURE. It also includes ongoing monitoring, incident response, and continuous improvement of risk treatments.

Question 15AI risk, accountability and responsible use

An organization wants to demonstrate that its AI system complies with ISO 42001. What is the most appropriate approach?

  1. Submit a self-attestation letter to ISO headquarters
  2. Implement an AI Management System (AIMS) aligned to ISO 42001 requirements and engage an accredited certification body for a third-party audit Correct answer.
  3. Publish the model's source code under an open-source license
  4. Register the model with NIST's National Vulnerability Database (NVD)

Explanation

ISO 42001 is a certifiable standard. Organizations implement an AIMS covering all requirements (context, leadership, planning, support, operation, evaluation, improvement) and then undergo a conformity audit by an ISO-accredited certification body to obtain a formal certificate.

Question 16AI risk, accountability and responsible use

Which of the following frameworks explicitly addresses the trustworthiness dimensions of AI systems including reliability, safety, security, privacy, fairness, and accountability?

  1. ISO 9001 Quality Management System
  2. NIST AI Risk Management Framework (AI RMF 1.0) Correct answer.
  3. PCI DSS v4.0
  4. ITIL 4 Service Management Framework

Explanation

The NIST AI RMF 1.0 explicitly defines trustworthy AI characteristics: valid and reliable, safe, secure and resilient, explainable and interpretable, privacy-enhanced, fair with bias managed, and accountable and transparent. It is the primary framework addressing AI trustworthiness holistically.

Question 17AI risk, accountability and responsible use

When evaluating a vendor's AI product for enterprise procurement, which of the following is MOST important from an AI governance perspective?

  1. Whether the vendor's CEO has a computer science degree
  2. The vendor's AI transparency documentation, data processing agreements, bias testing results, and incident response commitments Correct answer.
  3. Whether the AI product has the highest benchmark scores on academic leaderboards
  4. Whether the vendor offers the lowest price per API call

Explanation

Responsible AI procurement requires evaluating: transparency (model cards, datasheets), legal data processing agreements (DPAs), evidence of bias and safety testing, incident response and notification commitments, and alignment with organizational AI governance policies-not just benchmark performance or cost.

Question 18AI risk, accountability and responsible use

Which of the following best describes 'AI governance' as distinct from 'AI ethics'?

  1. AI governance is concerned with the technical architecture of AI systems; AI ethics is concerned with regulatory fines
  2. AI governance encompasses the organizational structures, policies, processes, and accountability mechanisms that operationalize ethical AI principles in practice Correct answer.
  3. AI ethics applies only to consumer products; AI governance applies only to government agencies
  4. They are identical concepts with no meaningful distinction

Explanation

AI ethics defines principles (fairness, transparency, accountability) while AI governance is the operational layer-the policies, committees, risk processes, audit mechanisms, and accountability structures that translate ethical principles into consistent organizational behavior. Both are necessary for responsible AI.

Question 19AI threat boundaries, agents and technical safeguards

A prompt injection attack against an LLM-powered application involves:

  1. Inserting malicious instructions into user input or retrieved content to override the model's system prompt Correct answer.
  2. Injecting SQL commands into the model's database connection
  3. Physically tampering with GPU firmware to alter inference outputs
  4. Flooding the model API with requests to cause a denial of service

Explanation

Prompt injection attempts to redirect an LLM application through adversarial instructions in user or external content. It does not mechanically acquire higher instruction authority merely by being present. Actual harm depends on model/application behavior, tool permissions, output handling and enforced authorization boundaries.

Question 20AI threat boundaries, agents and technical safeguards

Model inversion attacks aim to:

  1. Recover private training data by querying the model and analyzing its outputs Correct answer.
  2. Flip the sign of model gradients to slow down training
  3. Reverse-engineer the model's source code from its API responses
  4. Bypass content filters by encoding malicious inputs in Base64

Explanation

Model inversion attacks exploit the model's outputs to reconstruct sensitive training data. By querying the model with carefully crafted inputs and analyzing confidence scores or embeddings, an attacker can recover private information that was used during training.

Question 21AI threat boundaries, agents and technical safeguards

Which defense technique adds carefully calibrated random noise to training data or model outputs to protect individual privacy?

  1. Differential privacy Correct answer.
  2. Homomorphic encryption
  3. Federated learning
  4. Model watermarking

Explanation

Differential privacy (DP) provides mathematically rigorous privacy guarantees by adding calibrated random noise (often Gaussian or Laplace) to training computations or query outputs, ensuring that the presence or absence of any single individual's data cannot be reliably inferred from the model or results.

Question 22AI threat boundaries, agents and technical safeguards

An adversarial example is best defined as:

  1. A carefully crafted input with imperceptible perturbations that causes a model to make a wrong prediction Correct answer.
  2. A deliberately mislabeled training sample used to mislead the model during training
  3. An example drawn from a different domain than the training distribution
  4. A synthetic data point generated by a GAN to augment training

Explanation

Adversarial examples are inputs modified with small, often imperceptible perturbations (e.g., pixel changes in images or character substitutions in text) that cause a model to produce incorrect outputs with high confidence. They exploit the non-robustness of neural network decision boundaries.

Question 23AI threat boundaries, agents and technical safeguards

Which AI safety concept describes the property that an AI system should behave as intended across diverse, edge-case, and adversarial conditions?

  1. Reliability
  2. Scalability
  3. Robustness Correct answer.
  4. Interpretability

Explanation

Robustness in AI means the system maintains correct and safe behavior under distributional shift, adversarial inputs, noise, and unexpected edge cases. It is a core safety property tested through stress testing, adversarial evaluation, and red-teaming.

Question 24AI threat boundaries, agents and technical safeguards

Output filtering and content moderation guardrails for LLMs are best applied:

  1. Only at training time using RLHF alignment
  2. Only at the model layer using system prompts
  3. At multiple layers: input filtering, system prompt, model alignment, and output filtering Correct answer.
  4. Exclusively at the network perimeter using a WAF

Explanation

Defense-in-depth for LLM safety requires controls at multiple layers: input sanitization to block malicious prompts, system prompts to constrain model behavior, RLHF/fine-tuning alignment, and output filtering to catch harmful responses. No single layer is sufficient.

Question 25AI threat boundaries, agents and technical safeguards

Which of the following is NOT a recommended practice for securing an AI model API endpoint?

  1. Implementing rate limiting and throttling to prevent API abuse and DoS
  2. Logging all API requests and responses for audit and anomaly detection
  3. Publishing the full system prompt publicly to enable community review Correct answer.
  4. Enforcing authentication and authorization on every API call

Explanation

Publishing the full system prompt publicly reveals the model's guardrail instructions, making it trivial for attackers to craft prompts that bypass them. System prompts should be kept confidential to reduce the risk of jailbreaks and exploitation.

Question 26AI threat boundaries, agents and technical safeguards

Which of the following best characterizes the relationship between AI safety and AI security?

  1. They are synonymous terms for the same discipline
  2. AI security is a subset of AI safety restricted to data encryption
  3. AI safety focuses on preventing unintended harmful outcomes; AI security focuses on defending AI systems against intentional adversarial threats Correct answer.
  4. AI safety applies only to autonomous vehicles; AI security applies to all other AI systems

Explanation

AI safety addresses unintended harms-misalignment, hallucination, bias, and accidents-while AI security addresses intentional threats-adversarial attacks, data poisoning, model extraction, and system exploitation. Both disciplines are necessary and complementary for trustworthy AI.

Question 27AI threat boundaries, agents and technical safeguards

What is the primary goal of AI red-teaming?

  1. To optimize model inference throughput by profiling GPU utilization
  2. To comply with SEC disclosure requirements for AI system risks
  3. To retrain the model using red-labeled data that indicates high-risk inputs
  4. To adversarially test AI systems for safety failures, harmful outputs, and security vulnerabilities before deployment Correct answer.

Explanation

AI red-teaming involves structured adversarial testing by human or automated testers who attempt to elicit harmful outputs, discover safety bypasses, find security vulnerabilities (e.g., prompt injection), and uncover bias. It is a proactive safety and security evaluation practice before and after deployment.

Question 28AI threat boundaries, agents and technical safeguards

Which control reduces the risk that an AI agent with tool-use capabilities (e.g., code execution, web browsing) will perform irreversible harmful actions?

  1. Increasing the model's context window to capture more history
  2. Enabling the model to self-evaluate its actions and autonomously retry on failure
  3. Reducing the model's temperature to zero for deterministic outputs
  4. Requiring human confirmation for high-risk or irreversible actions before execution Correct answer.

Explanation

For AI agents with real-world tool access, requiring human-in-the-loop confirmation before executing high-risk or irreversible actions (e.g., deleting files, sending emails, making purchases) is a critical safety control that limits the blast radius of hallucinations or adversarial manipulation.

Question 29AI threat boundaries, agents and technical safeguards

Membership inference attacks exploit which property of over-fitted ML models?

  1. The model's tendency to produce identical outputs for all inputs in the training set
  2. The model's reliance on specific hardware accelerators not available to attackers
  3. The model's inability to generalize beyond its training distribution
  4. The model's higher confidence scores on training examples compared to unseen examples Correct answer.

Explanation

Over-fitted models memorize training data and exhibit systematically higher confidence (lower loss) on training samples than on unseen data. Membership inference attacks exploit this gap to determine whether a specific record was part of the training set-leaking sensitive information about training data composition.

Question 30AI threat boundaries, agents and technical safeguards

AI systems that autonomously optimize for a proxy metric that diverges from the true human objective are exhibiting which alignment failure?

  1. Mode collapse
  2. Catastrophic forgetting
  3. Gradient explosion
  4. Reward hacking (specification gaming) Correct answer.

Explanation

Reward hacking (or specification gaming) occurs when an AI system finds unexpected ways to maximize its reward signal without actually achieving the intended goal-exploiting gaps between the specified metric and the true human objective. This is a classic AI alignment failure with safety implications.

Question 31AI threat boundaries, agents and technical safeguards

In the context of AI security, 'model poisoning' specifically refers to:

  1. Applying incorrect labels to training data
  2. Attacking the training process or model weights directly-rather than the input data-to corrupt model behavior Correct answer.
  3. Using toxic output data to fine-tune a model into producing harmful content
  4. Flooding an AI API with adversarial inputs to degrade its performance

Explanation

Model poisoning targets the training process itself-either by manipulating gradient updates during federated learning or directly modifying model weights. This corrupts the model's behavior at the parameter level, distinct from data poisoning (which corrupts the training data) or inference-time attacks.

Question 32AI threat boundaries, agents and technical safeguards

Which of the following best describes the role of 'system prompts' in LLM application security?

  1. They are OS-level commands executed by the AI server to initialize the inference environment
  2. They are pre-configured instructions that define the model's persona, constraints, and guardrails, placed before the user's input Correct answer.
  3. They are network-level firewall rules applied to incoming API requests
  4. They are automated unit tests run against the model before each deployment

Explanation

System prompts are instructions prepended to the conversation by the application developer. They establish the model's role, behavioral constraints, topic restrictions, output format requirements, and safety rules. Protecting their confidentiality and designing them robustly is critical to application security.

Question 33AI threat boundaries, agents and technical safeguards

Which of the following access control principles is MOST important when granting AI systems access to enterprise APIs and databases?

  1. Maximum privilege - grant all permissions by default to avoid disrupting AI workflows
  2. Least privilege - grant only the specific permissions the AI system needs to perform its intended functions Correct answer.
  3. Role parity - grant AI systems the same permissions as the developers who built them
  4. Time-unlimited access - grant permanent credentials to avoid frequent re-authentication overhead

Explanation

Least privilege limits the blast radius of an AI system compromise, hallucination-induced misuse, or prompt injection attack. An AI system should only have access to the specific APIs, data stores, and operations required for its defined function-minimizing potential damage if the system is manipulated.

Question 34AI threat boundaries, agents and technical safeguards

Which of the following describes the key distinction between 'AI safety' focused on long-term existential risks versus near-term AI safety?

  1. Long-term safety focuses on physical infrastructure security; near-term safety focuses on model bias
  2. Long-term safety addresses speculative risks from superintelligent or misaligned advanced AI; near-term safety addresses current harms from deployed AI systems (bias, hallucination, misuse) Correct answer.
  3. Long-term safety is studied only by academic researchers; near-term safety is handled exclusively by governments
  4. They are identical concepts addressed by the same frameworks and practitioners

Explanation

Near-term AI safety addresses current, concrete harms: algorithmic bias, hallucination in deployed LLMs, misuse in autonomous weapons, and privacy violations. Long-term AI safety (also called 'AI existential safety') addresses speculative but potentially catastrophic risks from advanced or superintelligent AI systems misaligned with human values.

Question 35AI foundations, data and deployed-system lifecycle

What is the primary purpose of the tokenization step in a Generative AI pipeline?

  1. To convert raw text into numerical representations (tokens) the model can process Correct answer.
  2. To encrypt model weights before inference
  3. To filter harmful content before it reaches the end user
  4. To compress training data for storage efficiency

Explanation

Tokenization converts the supplied text into token units and associated model-input identifiers under its tokenizer. Tokens can represent subwords, characters or other supported units; they are not always whole words or one universal numeric representation across every generative system. Embeddings and later processing serve additional roles.

Question 36AI foundations, data and deployed-system lifecycle

Retrieval-Augmented Generation (RAG) primarily addresses which LLM limitation?

  1. Slow inference speed due to large model size
  2. Lack of mathematical reasoning capability
  3. Knowledge cutoff and inability to access real-time or proprietary information Correct answer.
  4. Vulnerability to adversarial prompt injection

Explanation

RAG augments an LLM by retrieving relevant documents from an external knowledge store at inference time, grounding responses in current or proprietary data. This directly addresses the static knowledge cutoff inherent in pre-trained models.

Question 37AI foundations, data and deployed-system lifecycle

Which transformer component is most directly responsible for capturing relationships between all tokens in an input sequence simultaneously?

  1. Feed-forward neural network layer
  2. Positional encoding
  3. Multi-head self-attention mechanism Correct answer.
  4. Layer normalization

Explanation

Multi-head self-attention computes pairwise attention weights between every token pair in the sequence, allowing the model to capture long-range dependencies in parallel. This is the defining architectural innovation of the transformer.

Question 38AI foundations, data and deployed-system lifecycle

Fine-tuning a pre-trained LLM on domain-specific data is best categorized as which type of machine learning approach?

  1. Transfer learning adapted to a target domain Correct answer.
  2. Unsupervised pre-training from scratch
  3. Reinforcement learning from human feedback
  4. Zero-shot inference without any additional training

Explanation

Fine-tuning is a form of transfer learning: general knowledge encoded during pre-training is transferred to a specific task or domain by continuing training on a smaller, curated dataset. This adapts the model without training from scratch.

Question 39AI foundations, data and deployed-system lifecycle

Which of the following is a key characteristic of Reinforcement Learning from Human Feedback (RLHF) used to align LLMs?

  1. It trains the model entirely on synthetically generated preference data without any human involvement
  2. It applies only to supervised classification tasks and cannot align generative models
  3. Human annotators rank model outputs by quality, and a reward model trained on these rankings guides further policy optimization Correct answer.
  4. It replaces the pre-training phase by directly teaching the model human values from scratch

Explanation

RLHF uses human annotators to compare and rank model-generated outputs. A reward model is trained on these human preferences, and the LLM policy is then fine-tuned via reinforcement learning (e.g., PPO) to maximize the learned reward-aligning outputs with human values and reducing harmful content.

Question 40AI foundations, data and deployed-system lifecycle

A generative AI system produces different outputs for the same input on consecutive calls. This property is called:

  1. Determinism
  2. Overfitting
  3. Concept drift
  4. Stochasticity (non-determinism) controlled by temperature sampling Correct answer.

Explanation

LLMs use temperature-controlled sampling during text generation, introducing randomness in token selection. This stochasticity means the model produces varied outputs for the same input, which is desirable for creativity but can complicate reproducibility and testing. Setting temperature=0 approaches greedy (near-deterministic) decoding.

Question 41AI foundations, data and deployed-system lifecycle

Which of the following best describes the 'temperature' parameter in LLM inference?

  1. The operating temperature of the GPU hardware during inference
  2. The maximum number of tokens the model generates per second
  3. The rate at which the model's weights are updated during fine-tuning
  4. A scaling factor applied to logits before softmax that controls the randomness of token selection Correct answer.

Explanation

Temperature divides the logits (raw prediction scores) before the softmax operation. A low temperature (near 0) sharpens the probability distribution toward the most likely token (deterministic/greedy), while a high temperature flattens it, increasing randomness and diversity in outputs.

Question 42AI foundations, data and deployed-system lifecycle

Parameter-efficient fine-tuning (PEFT) techniques like LoRA (Low-Rank Adaptation) are used to:

  1. Reduce model inference latency by pruning attention heads
  2. Compress model weights using quantization for edge deployment
  3. Encrypt model parameters to prevent unauthorized access during fine-tuning
  4. Adapt large pre-trained models to specific tasks by training only a small number of additional parameters Correct answer.

Explanation

LoRA and other PEFT methods insert small trainable adapter modules (low-rank matrices) into the model while keeping original weights frozen. Only the adapter parameters are updated, dramatically reducing compute and memory requirements for fine-tuning large foundation models.

Question 43AI foundations, data and deployed-system lifecycle

What is the primary function of a vector database in a Retrieval-Augmented Generation (RAG) architecture?

  1. To store and index model training checkpoints for version control
  2. To monitor model output toxicity scores in real time
  3. To cache LLM API responses and reduce redundant inference calls
  4. To store high-dimensional embeddings of documents and retrieve the most semantically similar chunks at query time Correct answer.

Explanation

In RAG, a vector database (e.g., Pinecone, Weaviate, Chroma) stores dense vector embeddings of document chunks. At query time, the user's query is embedded and the database performs approximate nearest-neighbor search to retrieve the most semantically relevant chunks for the LLM's context.

Question 44AI foundations, data and deployed-system lifecycle

The 'context window' of a large language model refers to:

  1. The visual display area used for the model's web-based chat interface
  2. The sliding average window used to smooth training loss curves
  3. The time window during which a user session remains active before expiry
  4. The maximum number of tokens (input + output) the model can process in a single inference call Correct answer.

Explanation

The context window is the maximum total token count (prompt + generated output) an LLM can process at once. Larger context windows (e.g., 128K, 1M tokens) allow longer documents, conversation histories, and more complex RAG retrievals but require significantly more compute.

Question 45AI foundations, data and deployed-system lifecycle

What does 'grounding' an LLM mean in the context of enterprise AI deployments?

  1. Physically connecting the AI server to an electrical ground to prevent static discharge
  2. Connecting the LLM to authoritative, current, and organization-specific data sources to reduce hallucinations and improve factual accuracy Correct answer.
  3. Setting temperature to zero to ensure deterministic model outputs
  4. Running the model in an isolated sandbox environment disconnected from the internet

Explanation

Grounding an LLM connects it to verified external knowledge (via RAG, tool use, or database lookups) so its responses are anchored in authoritative facts rather than the statistical patterns learned during pre-training. This is a primary technique for reducing hallucinations in enterprise settings.

Question 46AI foundations, data and deployed-system lifecycle

What is the primary purpose of 'prompt engineering' in working with large language models?

  1. Optimizing the hardware prompt response time on GPU inference servers
  2. Crafting effective input instructions that guide the LLM to produce accurate, useful, and safe outputs Correct answer.
  3. Writing firmware prompts for embedded AI chips deployed at the edge
  4. Engineering the model's tokenizer to handle domain-specific vocabulary

Explanation

Prompt engineering is the practice of designing, structuring, and iterating on text inputs to elicit desired behavior from LLMs. Effective prompts use techniques like few-shot examples, chain-of-thought instructions, role-setting, and constraint specification to improve output quality, accuracy, and safety.

Question 47AI risk, accountability and responsible use

Algorithmic bias in an AI hiring tool is most likely to cause which type of harm?

  1. Systematic discrimination against protected groups in candidate selection Correct answer.
  2. Data exfiltration from the HR database
  3. Increased cloud infrastructure costs due to model complexity
  4. Unauthorized access to model training weights

Explanation

Algorithmic bias occurs when a model trained on historically biased data perpetuates or amplifies those biases. In hiring, this can result in systematic discrimination against candidates from protected groups (race, gender, age), creating legal and ethical violations.

Question 48AI risk, accountability and responsible use

LIME (Local Interpretable Model-agnostic Explanations) is used to:

  1. Generate locally faithful explanations for individual model predictions by approximating the model with a simpler surrogate Correct answer.
  2. Encrypt model predictions before sending them to end users
  3. Detect adversarial examples in model inputs
  4. Reduce the dimensionality of training data to speed up convergence

Explanation

LIME explains individual predictions by perturbing the input and training a locally linear surrogate model on the model's responses to those perturbations. This provides human-interpretable explanations for black-box models at the instance level.

Question 49AI risk, accountability and responsible use

Which principle states that AI systems must be designed so their actions and decision-making processes can be understood and traced by humans?

  1. Autonomy
  2. Robustness
  3. Explainability Correct answer.
  4. Scalability

Explanation

Explainability (also called interpretability or transparency) is the AI ethics principle requiring that AI decisions can be understood by relevant stakeholders. It supports accountability, regulatory compliance, and human oversight.

Question 50AI risk, accountability and responsible use

An AI ethics review board within an organization serves primarily to:

  1. Accelerate model deployment by pre-approving all AI projects in bulk
  2. Manage GPU cluster capacity and cloud spending for AI workloads
  3. Provide oversight, evaluate ethical risks, and ensure AI deployments align with organizational values and regulatory requirements Correct answer.
  4. Conduct red-team penetration tests against deployed AI models

Explanation

An AI ethics review board provides human oversight of AI projects, evaluating risks related to bias, fairness, transparency, privacy, and societal impact. It ensures alignment with organizational ethics policies, legal obligations, and stakeholder expectations before and during AI deployment.