AIGP practice questions — 10 free

Which of the following is a method that helps to ensure AI integrity and that data is representative, accurate and unbiased?

• ATracking data lineage.
• BAssuming the data is accurate.Assuming the data is accurate supplies no verification step at all, so it cannot establish that the data is representative, accurate and unbiased.
• CPurchasing data from a vendor.Purchasing data from a vendor says nothing about how that data was collected or manipulated before use, so by itself it cannot establish accuracy or absence of bias.
• DReversioning the data internally.Reversioning the data internally is version management of a dataset; it tracks copies rather than the origin and handling of the data, so it does not confirm data quality.

Please use the following scenario to answer the next TWO questions. Amy is a data scientist who works for a health care company in France. She is asked to develop a plan to build a machine learning model that predicts patient satisfaction with doctors based on an analysis of the doctors’ notes. The doctors are from various countries, and the notes are in free-form text, compiled in a database in Germany. John is the AI governance lead for his EU-based company, XYZ. XYZ provides users with a virtual assistant. When exiting the system, users receive a notice which informs them they have interacted with an AI system and its responses may include hallucinations that should be reviewed by the user. John informs XYZ of his approval of this notice along with the way it is presented to him. Based on the information provided, why is XYZ not in compliance with the transparency requirements set out by the EU AI Act?

• AThe notice is not GDPR compliant.Nothing in the scenario points to a data protection defect in the notice; the issue is when users are told they are dealing with an AI system.
• BThe notice was not displayed in a timely manner.
• CThe notice is not accessible to vulnerable persons.The scenario gives no information about whether the notice is accessible to vulnerable persons, so that cannot be the reason for non-compliance.
• DThe notice is not adequate in terms of the information provided.The content is adequate: it states that the user interacted with an AI system and that its responses may include hallucinations to be reviewed.

You are the chief privacy officer of a medical research company that would like to collect and use sensitive data about cancer patients, such as their names, addresses, race and ethnic origin, medical histories, insurance claims, pharmaceutical prescriptions, eating and drinking habits and physical activity. The company will use this sensitive data to build an AI algorithm that will spot common attributes that will help predict if seemingly healthy people are more likely to get cancer. However, the company is unable to obtain consent from enough patients to sufficiently collect the minimum data to train its model. Which of the following solutions would most efficiently balance privacy concerns with the lack of available data during the testing phase?

• ADeploy the current model and recalibrate it over time with more data.Deploying an under-validated clinical model and recalibrating later transfers the risk onto patients, which is not acceptable in a medical setting.
• BExtend the model to multimodal ingestion with text and images.Adding image and text modalities neither solves the data shortage nor the consent problem; it increases the volume of sensitive data that must be collected.
• CUtilize synthetic data to offset the lack of patient data.
• DRefocus the algorithm to patients without cancer.Refocusing on patients without cancer abandons the stated business purpose and still leaves the data and consent constraints unresolved.

A company initially intended to use a large data set containing personal information to train an AI model. After consideration, the company determined that it can derive enough value from the data set without any personal information and permanently obfuscated all personal data elements before training the model. This is an example of applying which privacy-enhancing technique (PET)?

• AAnonymization.
• BPseudonymization.Pseudonymization is reversible because a key still allows re-identification, which contradicts the stem's statement that the personal data elements were permanently obfuscated.
• CDifferential privacy.Differential privacy protects individuals by adding calibrated noise during training or analysis, not by permanently removing the personal data elements before training.
• DFederated learning.Federated learning trains across distributed data without centralizing it, which is a different approach from obfuscating the personal data elements in one data set.

You asked a generative AI tool to recommend new restaurants to explore in Boston, Massachusetts that have a specialty Italian dish made in a traditional fashion without spinach and wine. The generative AI tool recommended five restaurants for you to visit. After looking up the restaurants, you discovered one restaurant did not exist and two others did not have the dish. This information provided by the generative AI tool is an example of what commonly called:

• APrompt injection.Prompt injection is an attack in which crafted input manipulates a model's instructions, whereas here an ordinary request returned fabricated restaurants and dishes.
• BModel collapse.Model collapse is degradation that occurs when a model is repeatedly trained on its own generated output, not a single fabricated answer to a user.
• CHallucination.
• DOverfitting.Overfitting is a training failure in which a model memorises its training data and generalises poorly; it does not describe confidently inventing a nonexistent restaurant.

Which of the following compliance related controls within an organization is most easily adapted to identify AI risks?

• APrivacy training.Privacy training raises awareness. It is not a structured risk identification process that can be repurposed.
• BPenetration testing.Penetration testing targets security vulnerabilities and does not map onto bias, fairness or societal impact risks.
• CTransfer risk assessments.Transfer risk assessments are scoped to cross-border data transfers and cover only a narrow slice of AI risk.
• DPrivacy impact assessments.

All of the following are examples of types of bias that can occur in AI systems EXCEPT:

• AImplicit biasImplicit bias is a standard type; the unconscious assumptions of data collectors and designers pass into the model.
• BExplicit bias
• CSampling biasSampling bias, where training data under-represents part of the population, is the classic AI bias type.
• DTemporal biasTemporal bias arises when the world changes and the model is not updated to match.

All of the following are potential benefits of using private over public LLMs EXCEPT:

• AReduction in time taken for data validation and verification.A private deployment can draw on internal data the enterprise has already governed, which reduces validation and verification effort, so this is a genuine benefit.
• BConfirmation of security and confidentiality.Keeping data inside the enterprise boundary allows security and confidentiality to be confirmed, which is one of the main reasons for choosing a private model.
• CReduction in possibility of hallucinated information.
• DApplication for specific use cases within the enterprise.A private model can be tuned to specific enterprise use cases, which is a real advantage of private over public deployment rather than the exception.

A team plans to ingest a customer's full profile into a churn model although only three fields are predictive. Which two requirements does this most directly conflict with?

• ABreach notification and record keeping.Neither of these is engaged by the facts described.
• BData minimization and privacy by design.
• CCross-border transfer and localization.Nothing in the scenario describes data leaving the jurisdiction.
• DConsent withdrawal and portability.Consent withdrawal and portability are data subject rights and are a different matter from the scope of fields ingested.

During testing, the team discovers the model fails badly on a rare but safety-critical input pattern. Which response follows the requirement to identify and manage issues and risks during training and testing?

• ANote it in the release notes and proceed, since the pattern is rare.A line in the release notes is not a treatment. The risk still ships.
• BLog the issue, assess its severity and likelihood, decide on treatment and record the decision before release.
• CRemove the rare pattern from the test set so the score improves.Removing the failing pattern from the test set manipulates the criterion while leaving the defect intact.
• DEscalate only if a customer reports it after launch.Waiting for a customer report leaves a safety-critical risk to be surfaced by actual harm.