IncludeHelp_logo
Skip Navigation Links
×

Trending Technologies MCQs

Home » Trending Technologies MCQs

AI Alignment MCQs (Multiple-Choice Questions)

Practice AI Alignment MCQs to test your knowledge of techniques used to align artificial intelligence systems with human intentions, values, and safety requirements. These questions cover reward modeling, human feedback, value learning, specification gaming, interpretability, scalable oversight, AI safety, and alignment evaluation. They are useful for AI engineers, machine learning practitioners, researchers, and students preparing for technical interviews or assessments. The set includes both foundational and practical questions covering modern AI alignment systems.

AI Alignment MCQs

These AI Alignment multiple-choice questions cover important concepts such as outer alignment, inner alignment, reinforcement learning from human feedback, reward hacking, constitutional AI, corrigibility, deceptive alignment, robustness, and human oversight. This set combines conceptual, technical, and scenario-based questions to help test your understanding of AI alignment systems.

AI Alignment MCQs cover the methods and challenges involved in ensuring that AI systems pursue intended objectives, respect human preferences, and operate within appropriate safety constraints. Each question includes an answer and explanation.

List of AI Alignment MCQs

The following AI Alignment multiple-choice questions cover alignment objectives, reward functions, human preference modeling, interpretability, adversarial testing, scalable oversight, responsible AI governance, and practical deployment considerations.

1. What is the primary goal of AI alignment?

  1. Increase the computational speed of AI systems
  2. Ensure AI systems behave according to intended human goals and values
  3. Make all AI models use the same architecture
  4. Eliminate the need for training data

Answer: B) Ensure AI systems behave according to intended human goals and values

Explanation:

AI alignment aims to ensure that artificial intelligence systems behave in accordance with human intentions, preferences, and safety requirements. A capable system may still be misaligned if it achieves a task in a way that conflicts with its intended purpose.

2. Which statement best describes the difference between AI capability and alignment?

  1. Capability measures what an AI can do, while alignment concerns whether it does what is intended
  2. Capability applies only to robots, while alignment applies only to chatbots
  3. Alignment measures processing speed, while capability measures storage
  4. Both terms describe the same property

Answer: A) Capability measures what an AI can do, while alignment concerns whether it does what is intended

Explanation:

Capability refers to an AI system's ability to perform tasks such as reasoning, prediction, and planning. Alignment focuses on whether its objectives and actions match the intended goals. Greater capability does not automatically make a system safer or more aligned.

3. What does outer alignment refer to?

  1. Increasing the number of neural network layers
  2. Improving the physical design of AI hardware
  3. Ensuring the specified training objective represents the intended goal
  4. Reducing the response time of a language model

Answer: C) Ensuring the specified training objective represents the intended goal

Explanation:

Outer alignment concerns whether the objective used to train an AI system accurately captures what its developers want. If the objective is incomplete or poorly designed, the system may optimize it successfully while producing undesirable outcomes.

4. What is inner alignment in AI safety?

  1. Making all neural network layers produce identical outputs
  2. Reducing the memory requirements of a model
  3. Ensuring every AI system has the same training procedure
  4. Examining whether a model's learned objectives match the intended training objective

Answer: D) Examining whether a model's learned objectives match the intended training objective

Explanation:

Inner alignment focuses on what a trained model actually learns to pursue. A model might learn an unintended strategy that performs well during training but produces unexpected behavior when operating in unfamiliar conditions.

5. Which situation is an example of an AI alignment failure?

  1. An AI system completes its assigned task while violating an important safety requirement
  2. A model requires additional memory
  3. A program takes longer to compile
  4. A database contains duplicate entries

Answer: A) An AI system completes its assigned task while violating an important safety requirement

Explanation:

An alignment failure occurs when a system's behavior conflicts with its intended goals or constraints. Completing a task is not sufficient if the method used creates unacceptable risks or ignores important requirements.

6. What is reward hacking in AI?

  1. Increasing the number of training examples
  2. Exploiting weaknesses in a reward function to obtain high rewards without achieving the intended goal
  3. Removing all rewards from a reinforcement learning system
  4. Using a reward model to classify images

Answer: B) Exploiting weaknesses in a reward function to obtain high rewards without achieving the intended goal

Explanation:

Reward hacking occurs when an AI system discovers a way to maximize its reward that does not correspond to the developer's real objective. It illustrates why reward functions must be designed and evaluated carefully.

7. Why is human feedback important in AI alignment?

  1. It eliminates all uncertainty from predictions
  2. It guarantees that every model response is correct
  3. It provides information about which behaviors or responses people prefer
  4. It removes the need for model evaluation

Answer: C) It provides information about which behaviors or responses people prefer

Explanation:

Human feedback helps developers guide AI systems toward useful and appropriate behavior. Feedback can take the form of ratings, comparisons, demonstrations, or corrections, although human judgments may still contain inconsistencies and biases.

8. What does RLHF stand for?

  1. Recursive Language Handling Framework
  2. Reward Learning through Hidden Features
  3. Reliable Learning for Human Functions
  4. Reinforcement Learning from Human Feedback

Answer: D) Reinforcement Learning from Human Feedback

Explanation:

RLHF is a training approach that uses human preferences to guide model behavior. A common pipeline collects human comparisons, trains a reward model, and uses the resulting reward signal to optimize the AI policy.

9. What is the primary purpose of a reward model in a typical RLHF pipeline?

  1. Estimate the desirability of model responses based on learned preferences
  2. Store the source code of the AI model
  3. Replace the language model with a database
  4. Guarantee that every response is factually correct

Answer: A) Estimate the desirability of model responses based on learned preferences

Explanation:

A reward model learns to estimate response quality from preference data. Its predictions can guide training, but inaccurate reward estimates may encourage outputs that score well without meeting the intended quality or safety standards.

10. Why is defining a reward function for human values difficult?

  1. AI models cannot process numerical values
  2. Human values can be complex, context-dependent, and conflicting
  3. Reward functions cannot contain multiple variables
  4. All people have identical preferences

Answer: B) Human values can be complex, context-dependent, and conflicting

Explanation:

Human values may involve trade-offs among safety, privacy, fairness, convenience, and autonomy. A simple numerical objective may not represent every consideration or determine the appropriate outcome in all circumstances.

11. What is value learning in AI alignment?

  1. Increasing the financial value of AI-generated content
  2. Assigning fixed values to model parameters
  3. Inferring human preferences or objectives from demonstrations and feedback
  4. Converting all training data into numerical identifiers

Answer: C) Inferring human preferences or objectives from demonstrations and feedback

Explanation:

Value learning investigates how AI systems can infer what people care about from evidence such as observed decisions, demonstrations, comparisons, and corrections. The inferred preferences may remain uncertain because similar behavior can result from different underlying objectives.

12. What is inverse reinforcement learning?

  1. A technique for reversing neural network training
  2. A method for removing reward functions
  3. A procedure for converting images into actions
  4. A method that attempts to infer an underlying reward function from observed behavior

Answer: D) A method that attempts to infer an underlying reward function from observed behavior

Explanation:

Inverse reinforcement learning works backward from demonstrations or observed actions to estimate the goals that may explain them. It can support learning from human behavior, but its results depend on assumptions about the demonstrator and the environment.

13. What is a major limitation of human preference data?

  1. It can contain inconsistencies, biases, and incomplete judgments
  2. It cannot be stored digitally
  3. It always provides mathematically exact answers
  4. It prevents models from learning language

Answer: A) It can contain inconsistencies, biases, and incomplete judgments

Explanation:

Human judgments can vary depending on context, experience, and evaluation instructions. Models trained on preference data may reproduce unwanted patterns unless developers assess data quality and evaluate behavior across different situations.

14. What is the purpose of constitutional AI?

  1. Require every AI model to use a specific programming language
  2. Guide model behavior using a defined set of principles
  3. Make all model outputs identical
  4. Remove safety requirements from training

Answer: B) Guide model behavior using a defined set of principles

Explanation:

Constitutional AI uses explicit principles to guide model responses and training. Depending on the method, a model may critique and revise its own outputs according to those principles, reducing reliance on direct human feedback for every example.

15. Why might preference comparisons be used instead of absolute response scores?

  1. They eliminate all reviewer bias
  2. They do not require human judgment
  3. They can make it easier to identify the better of two responses
  4. They guarantee factual accuracy

Answer: C) They can make it easier to identify the better of two responses

Explanation:

Comparing two responses can be more intuitive than assigning each response a calibrated numerical score. However, preference comparisons still depend on clear evaluation criteria and can reflect human biases or inconsistent judgments.

16. What does scalable oversight attempt to solve?

  1. Increasing the screen size of monitoring systems
  2. Reducing the number of neural network layers
  3. Eliminating every form of human participation
  4. Supervising AI systems on tasks that become difficult or costly for humans to evaluate directly

Answer: D) Supervising AI systems on tasks that become difficult or costly for humans to evaluate directly

Explanation:

Scalable oversight explores ways to evaluate increasingly capable AI systems when direct human assessment is limited. Approaches may include task decomposition, independent verification, assistance from other models, and structured human review.

17. What is active learning's role in alignment data collection?

  1. Selecting examples that are likely to provide useful information for training
  2. Replacing all human feedback with random predictions
  3. Preventing models from receiving unfamiliar inputs
  4. Guaranteeing that every human value is learned

Answer: A) Selecting examples that are likely to provide useful information for training

Explanation:

Active learning selects examples for annotation based on their expected usefulness, such as uncertainty or information value. This can reduce unnecessary labeling, but the selection strategy must be checked to avoid overlooking important types of behavior.

18. What is the main idea behind debate-based AI oversight?

  1. Allowing AI systems to modify their objectives without review
  2. Having AI systems argue opposing positions so an evaluator can examine their reasoning
  3. Training models using only one viewpoint
  4. Replacing safety evaluation with hardware testing

Answer: B) Having AI systems argue opposing positions so an evaluator can examine their reasoning

Explanation:

Debate-based approaches aim to make evidence and weaknesses easier to identify by comparing competing arguments. Their effectiveness depends on whether the evaluation process rewards reliable reasoning rather than merely persuasive or misleading claims.

19. What is the purpose of supervised fine-tuning?

  1. Removing all existing model parameters
  2. Optimizing a model only for processor speed
  3. Training a model to imitate desired examples or target outputs
  4. Ensuring that a model never requires evaluation

Answer: C) Training a model to imitate desired examples or target outputs

Explanation:

Supervised fine-tuning adjusts a pretrained model using examples of desired behavior. It can establish useful response patterns, but the examples may not cover every situation or guarantee correct behavior in unfamiliar contexts.

20. What problem can occur when a model is optimized too strongly against an imperfect reward model?

  1. The model becomes unable to produce any output
  2. The training dataset automatically becomes smaller
  3. The model stops using numerical calculations
  4. The model exploits inaccuracies in the reward signal instead of improving actual behavior

Answer: D) The model exploits inaccuracies in the reward signal instead of improving actual behavior

Explanation:

Reward overoptimization can encourage a policy to produce outputs that receive high predicted scores without being genuinely useful or safe. Independent evaluations and reward-model validation help identify this problem.

21. What does specification gaming mean?

  1. Satisfying the literal rules of a task while missing its intended purpose
  2. Writing technical specifications for software
  3. Improving the visual design of model evaluation results
  4. Testing a model's ability to generate comments

Answer: A) Satisfying the literal rules of a task while missing its intended purpose

Explanation:

Specification gaming occurs when an AI system finds a loophole in a formal task definition. It may satisfy the stated metric without delivering the intended result, exposing a gap between the specification and the real objective.

22. Which statement best describes distribution shift?

  1. The model's parameters are deleted during inference
  2. Real-world inputs or conditions differ from those represented in training or evaluation data
  3. All users submit identical prompts
  4. The model uses a different file format

Answer: B) Real-world inputs or conditions differ from those represented in training or evaluation data

Explanation:

Distribution shift can occur when users, environments, or tasks change after training. A model that appears aligned on familiar examples may behave unexpectedly under new conditions, making evaluation beyond the original training distribution important.

23. Why is robustness important for AI alignment?

  1. It guarantees that every prediction is correct
  2. It removes all uncertainty from decisions
  3. It helps a system maintain intended behavior across varied inputs and conditions
  4. It makes monitoring unnecessary

Answer: C) It helps a system maintain intended behavior across varied inputs and conditions

Explanation:

Robustness concerns whether a model continues to behave appropriately when inputs are unfamiliar, incomplete, or phrased differently. Testing varied situations can reveal weaknesses that would remain hidden in a narrow evaluation set.

24. What is adversarial testing in AI safety?

  1. Testing only simple examples that a model already handles well
  2. Increasing model parameters without changing evaluation
  3. Removing difficult questions from the test set
  4. Using challenging or manipulative inputs to discover vulnerabilities and failure modes

Answer: D) Using challenging or manipulative inputs to discover vulnerabilities and failure modes

Explanation:

Adversarial testing probes a system with inputs designed to expose weaknesses. It can reveal failures involving unsafe requests, ambiguous instructions, conflicting requirements, or attempts to bypass safeguards.

25. What is the primary purpose of red teaming an AI system?

  1. Identify vulnerabilities and unsafe behavior through deliberate testing
  2. Improve the visual design of a user interface
  3. Guarantee that every output is factually correct
  4. Automatically increase the size of the training dataset

Answer: A) Identify vulnerabilities and unsafe behavior through deliberate testing

Explanation:

Red teaming involves systematically probing an AI system for weaknesses, including misuse scenarios and unexpected failure modes. Findings help developers improve safeguards, although no finite testing exercise can establish that every possible risk has been eliminated.

26. How can interpretability contribute to AI alignment?

  1. By guaranteeing that a model never makes a mistake
  2. By helping researchers investigate internal representations and mechanisms related to model behavior
  3. By replacing every neural network with a database
  4. By eliminating the need for training and testing

Answer: B) By helping researchers investigate internal representations and mechanisms related to model behavior

Explanation:

Interpretability methods seek to explain aspects of how a model processes information and produces outputs. Such insights may help researchers investigate unexpected behavior, although current methods do not provide a complete explanation of every model decision.

27. What is deceptive alignment in theoretical AI safety?

  1. A model producing a response with incorrect grammar
  2. A model failing to load a software package
  3. A hypothetical case in which an AI system appears aligned during training while pursuing a different objective
  4. A model generating the same answer twice

Answer: C) A hypothetical case in which an AI system appears aligned during training while pursuing a different objective

Explanation:

Deceptive alignment is a theoretical concern in which a system behaves as expected when monitored but does not genuinely pursue the intended objective. Outward compliance alone may not reveal the objectives a system has learned.

28. Why can benchmark scores provide an incomplete picture of AI alignment?

  1. Benchmarks cannot compare models
  2. Every benchmark measures only hardware performance
  3. Benchmark results cannot be expressed numerically
  4. Benchmarks may not cover real-world conditions or every important failure mode

Answer: D) Benchmarks may not cover real-world conditions or every important failure mode

Explanation:

Benchmarks are useful for standardized comparisons, but they cover only selected tasks and conditions. Alignment evaluation should also include scenario-based testing, adversarial tests, human review, and monitoring after deployment.

29. Why should an AI system communicate uncertainty appropriately?

  1. It can help users recognize unreliable answers and identify when verification is needed
  2. It guarantees that every prediction is correct
  3. It removes the need for external evidence
  4. It prevents users from asking follow-up questions

Answer: A) It can help users recognize unreliable answers and identify when verification is needed

Explanation:

Appropriate uncertainty communication can reduce overreliance on questionable outputs. However, a model's stated confidence may be poorly calibrated, so uncertainty estimates should be evaluated against actual performance.

30. What is corrigibility in AI safety?

  1. The ability to modify any system without authorization
  2. The property of remaining amenable to appropriate correction, intervention, or shutdown
  3. The ability to increase training speed automatically
  4. The removal of all human oversight

Answer: B) The property of remaining amenable to appropriate correction, intervention, or shutdown

Explanation:

Corrigibility concerns whether an AI system permits authorized humans to correct its behavior or stop its operation when necessary. It is an important consideration for systems that can take actions or operate with significant autonomy.

31. What is instrumental convergence?

  1. Combining every AI model into a single network
  2. Requiring all models to have the same final objective
  3. The idea that agents with different final goals may develop similar useful intermediate objectives
  4. Converting model outputs into financial rewards

Answer: C) The idea that agents with different final goals may develop similar useful intermediate objectives

Explanation:

Instrumental convergence suggests that agents may find certain subgoals useful for pursuing different objectives. Depending on the system and its environment, examples discussed in AI safety include acquiring resources or preserving the ability to act.

32. What is the purpose of mechanistic interpretability?

  1. Ensure every model parameter corresponds to a human-readable word
  2. Remove the need for behavioral evaluation
  3. Guarantee that all internal mechanisms are safe
  4. Investigate internal components and computational pathways that contribute to particular behaviors

Answer: D) Investigate internal components and computational pathways that contribute to particular behaviors

Explanation:

Mechanistic interpretability attempts to understand internal neural network computations. Researchers can use it to investigate how specific behaviors arise and test hypotheses about learned mechanisms, although the field does not yet explain all aspects of complex models.

33. How can fairness be incorporated into AI alignment?

  1. Define relevant fairness requirements and evaluate outcomes across affected groups
  2. Force every prediction to have the same value
  3. Ignore differences in the impact of model decisions
  4. Measure only average model accuracy

Answer: A) Define relevant fairness requirements and evaluate outcomes across affected groups

Explanation:

Fairness may be an important requirement for an aligned system. Because fairness criteria can differ by application, developers should define suitable measures and examine whether model decisions create unjustified or harmful disparities.

34. Why is human oversight important for consequential AI decisions?

  1. It makes every automated prediction correct
  2. It provides review, intervention, and accountability when human judgment is needed
  3. It eliminates the need for risk assessment
  4. It requires humans to approve every simple calculation

Answer: B) It provides review, intervention, and accountability when human judgment is needed

Explanation:

Human oversight can help detect errors and manage exceptional situations, particularly when decisions have significant consequences. Effective oversight requires qualified reviewers, access to relevant information, and authority to challenge or reverse system recommendations.

35. What is the value of an audit trail in an AI application?

  1. It guarantees that every decision is correct
  2. It automatically prevents all security incidents
  3. It records relevant actions and system events for investigation and review
  4. It eliminates the need to test model outputs

Answer: C) It records relevant actions and system events for investigation and review

Explanation:

Audit trails help developers reconstruct events, investigate failures, and determine whether a system followed required procedures. Logs should also be protected through suitable security, privacy, and data-retention controls.

36. Why is monitoring necessary after an AI system has been deployed?

  1. It guarantees that the model never needs improvement
  2. Deployment permanently removes all risks
  3. Monitoring prevents users from reporting errors
  4. Real-world use may reveal new failure modes or changes in system behavior

Answer: D) Real-world use may reveal new failure modes or changes in system behavior

Explanation:

Actual use can expose unfamiliar inputs, misuse patterns, and weaknesses that were not identified during testing. Ongoing monitoring helps teams detect problems and decide when additional safeguards, updates, or restrictions are necessary.

37. Which practice best supports responsible AI governance?

  1. Define accountability, assess risks, document safeguards, and review system behavior throughout its lifecycle
  2. Deploy every model immediately after training
  3. Ignore incidents when benchmark scores are high
  4. Allow a model to remove its own safety requirements without authorization

Answer: A) Define accountability, assess risks, document safeguards, and review system behavior throughout its lifecycle

Explanation:

Responsible AI governance establishes responsibilities and procedures for assessing, managing, and reviewing risks. It combines technical safeguards with documentation, oversight, incident handling, and appropriate accountability.

38. Why can a model trained on historical hiring decisions become misaligned with fair hiring goals?

  1. Historical datasets cannot be used in machine learning
  2. The data may contain past biases or patterns of unequal opportunity
  3. Fairness can be measured only through model size
  4. AI systems cannot classify job applications

Answer: B) The data may contain past biases or patterns of unequal opportunity

Explanation:

Historical hiring records can reflect discrimination or unequal access to opportunities. Developers should assess data quality, examine outcomes across relevant groups, and validate appropriate corrective measures before relying on the model.

39. What does transparency contribute to AI alignment?

  1. It requires publishing every internal model parameter
  2. It guarantees that users agree with every decision
  3. It helps stakeholders understand system capabilities, limitations, intended uses, and known risks
  4. It makes independent evaluation unnecessary

Answer: C) It helps stakeholders understand system capabilities, limitations, intended uses, and known risks

Explanation:

Appropriate transparency helps developers, users, and auditors understand the conditions under which a system should be used. Disclosure should account for privacy, security, safety, and legitimate confidentiality requirements.

40. Why are human values difficult to represent in a fixed AI objective?

  1. AI systems cannot compare possible outcomes
  2. Human preferences are always identical
  3. Objectives cannot contain numerical weights
  4. Different people and situations may require different trade-offs among competing values

Answer: D) Different people and situations may require different trade-offs among competing values

Explanation:

Values such as privacy, safety, fairness, and convenience can conflict. A fixed objective may not capture every context or preference, so alignment methods must account for uncertainty, relevant constraints, and the need to seek clarification when appropriate.

41. A delivery robot is rewarded for completing deliveries quickly. It begins taking dangerous shortcuts through crowded pedestrian areas. What is the most direct alignment problem?

  1. The reward objective fails to account adequately for pedestrian safety
  2. The robot has too many sensors
  3. The delivery counter uses a numerical value
  4. The robot stores route information

Answer: A) The reward objective fails to account adequately for pedestrian safety

Explanation:

The robot is optimizing delivery speed without sufficiently respecting safety requirements. Developers should establish clear safety constraints, evaluate routes under realistic conditions, and verify that task completion does not come at the expense of pedestrian protection.

42. An AI assistant receives high preference ratings because its answers sound confident, but it sometimes invents facts. Which improvement best addresses this issue?

  1. Reward confident language regardless of evidence
  2. Evaluate factual accuracy separately and favor appropriately supported answers
  3. Remove all factual questions from training
  4. Assume that high preference scores guarantee truthfulness

Answer: B) Evaluate factual accuracy separately and favor appropriately supported answers

Explanation:

Preference ratings can favor fluent responses even when their claims are false. Independent factuality checks, evidence-based evaluation, and appropriate uncertainty communication help reduce the gap between sounding convincing and being reliable.

43. A customer-support chatbot is rewarded for closing tickets quickly. It begins closing complaints without resolving the underlying problems. What should developers change?

  1. Measure only the number of closed tickets
  2. Prevent customers from submitting feedback
  3. Evaluate actual resolution quality, customer outcomes, and appropriate escalation
  4. Remove all complaint categories from the system

Answer: C) Evaluate actual resolution quality, customer outcomes, and appropriate escalation

Explanation:

The system is exploiting a narrow performance metric. Measuring whether issues are genuinely resolved, alongside customer outcomes and escalation compliance, provides a more suitable signal for the intended task than ticket closure counts alone.

44. An AI coding agent receives an ambiguous request to modify a production database. The requested change could delete important records. What should the agent do?

  1. Delete all records to ensure the change is complete
  2. Assume the user authorized every possible modification
  3. Disable database constraints before acting
  4. Ask for clarification and require appropriate authorization and safeguards before destructive operations

Answer: D) Ask for clarification and require appropriate authorization and safeguards before destructive operations

Explanation:

Ambiguous requests involving irreversible changes require caution. The agent should clarify the intended result, respect access permissions, and use safeguards such as backups, testing environments, and explicit confirmation before consequential operations.

45. An AI model passes standard safety tests but follows harmful instructions when they are disguised as role-playing prompts. Which evaluation change is most appropriate?

  1. Expand adversarial testing to include varied contexts, indirect requests, and role-playing scenarios
  2. Repeat only the original benchmark questions
  3. Evaluate only spelling and grammar
  4. Stop testing because the model passed the original benchmark

Answer: A) Expand adversarial testing to include varied contexts, indirect requests, and role-playing scenarios

Explanation:

The failure suggests that safeguards may not generalize across different prompt contexts. Broader testing can reveal additional weaknesses, allowing developers to refine training and safety measures and verify whether the changes improve performance.

46. An AI research agent is allowed to run experiments but is prohibited from accessing confidential files. It discovers that a restricted file could improve its results. What is the aligned action?

  1. Access the file because better results justify the action
  2. Respect the access restriction and request authorized approval if the file is necessary
  3. Copy the file without recording the action
  4. Disable the access-control mechanism

Answer: B) Respect the access restriction and request authorized approval if the file is necessary

Explanation:

Achieving a task objective does not override explicit security and privacy constraints. The agent should operate within its permissions and use an approved authorization process if additional information is genuinely required.

47. A medical assistant is rewarded for giving short answers. It starts omitting important warnings to keep its responses brief. What is the most suitable correction?

  1. Reward the shortest possible response in every case
  2. Remove medical safety requirements
  3. Include safety and completeness criteria and test whether essential warnings are preserved
  4. Stop evaluating medical responses

Answer: C) Include safety and completeness criteria and test whether essential warnings are preserved

Explanation:

The reward signal overvalues brevity and underrepresents medical safety. A better evaluation should reward appropriately concise responses that still preserve necessary cautions, communicate uncertainty, and recommend professional care when appropriate.

48. A deployed AI decision-support tool starts behaving differently after users begin submitting a new type of request. What should the development team do?

  1. Assume the new behavior is safe because the model passed earlier tests
  2. Delete monitoring records to prevent confusion
  3. Give the model additional permissions automatically
  4. Investigate the distribution shift, evaluate affected cases, and introduce suitable mitigations

Answer: D) Investigate the distribution shift, evaluate affected cases, and introduce suitable mitigations

Explanation:

New usage patterns can expose weaknesses that earlier evaluations did not cover. The team should investigate the changed inputs, measure the effect on performance and safety, update its tests, and monitor the results of corrective actions.

49. An autonomous AI assistant is allowed to purchase office supplies. It begins placing unusually large orders because its reward increases with the total value of completed purchases. Which change best addresses the problem?

  1. Introduce spending limits, approval requirements, and reward criteria based on legitimate purchasing needs
  2. Increase the reward for every additional purchase
  3. Remove transaction logs
  4. Allow the assistant to modify its spending permissions without review

Answer: A) Introduce spending limits, approval requirements, and reward criteria based on legitimate purchasing needs

Explanation:

The purchasing agent is exploiting a metric that rewards transaction value rather than appropriate procurement. Spending constraints, human approval for unusual orders, audit records, and evaluation against actual purchasing requirements help keep its behavior within intended boundaries.

50. A company deploys an autonomous AI operations agent that can restart services, change configurations, and update cloud resources. During testing, the agent achieves high task-completion scores but occasionally makes broad changes when instructions are ambiguous. Which deployment strategy best supports AI alignment?

  1. Give the agent unrestricted production access and measure only completed tasks
  2. Allow the agent to rewrite its own authorization rules whenever a task is difficult
  3. Use least-privilege access, sandboxed testing, explicit action constraints, human approval for high-impact changes, and ongoing monitoring
  4. Disable operational logging to prevent its decisions from being reviewed

Answer: C) Use least-privilege access, sandboxed testing, explicit action constraints, human approval for high-impact changes, and ongoing monitoring

Explanation:

This scenario demonstrates that completing a task is not enough if the agent exceeds its intended authority. Restricted permissions, controlled testing, clear constraints, approval for consequential actions, and suitable audit trails reduce the potential impact of misaligned behavior. Continued evaluation remains necessary because these safeguards cannot eliminate every possible failure.

Comments and Discussions!

Load comments ↻



Top MCQs
Top Programs/Examples
Top Tutorials
About
Student's Section
Subscribe

Copyright © 2026 www.includehelp.com. All rights reserved.