AIGP
Free AIGP practice questions
10 questions from our AIGP practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.
Question 1 of 10, Understanding the Foundations of AI Governance
A 40-person startup builds an AI tool for scheduling hospital staff. It cannot afford a large governance function. Which approach to AI governance is MOST appropriate?
- AScale the program to its size and risk, assign duties to existing roles and focus controls on the highest-risk uses Correct answer
- BCopy a multinational bank's full AI governance framework so that the company is prepared for every possible regulator inquiry
- COutsource all governance decisions to its largest hospital customer because that customer bears the risk of using the tool
- DPostpone AI governance until the company has grown, since governance frameworks apply only to large enterprises and public bodies
Answer and explanation: Governance should be proportionate to the organization's size, resources and the risk of its AI uses, which usually means assigning responsibilities to existing roles and prioritising high-risk applications. Delaying governance ignores that obligations and harms do not depend on company size, especially in healthcare. Copying a large bank's framework wastes scarce resources on controls that do not fit. Customers can share some responsibilities, but the developer remains responsible for the system it builds and cannot outsource its own governance.
Question 2 of 10, Understanding the Foundations of AI Governance
A health insurer wants to test its pricing model for racial bias, but it does not collect race data from members. The privacy team objects to collecting it. Which statement BEST describes this situation?
- AThe insurer can confirm the model is unbiased simply because it never used race as an input feature during training or scoring
- BFairness testing is only legally relevant for employment decisions, so the insurer does not need to consider it for pricing models
- CFairness and privacy are in tension, so the insurer must weigh them, for example by gathering limited data under strict safeguards Correct answer
- DPrivacy always overrides fairness, so the insurer should not test for bias at all if doing so would require demographic data of any kind
Answer and explanation: Detecting bias often requires demographic data whose collection raises privacy concerns, so organizations must balance the two principles, for instance with limited, separately stored, access-controlled data or inference methods; the EU AI Act even permits special-category processing for bias detection under strict conditions. Privacy does not automatically override fairness. Excluding race does not prove the absence of bias because proxies exist. Fairness obligations extend to insurance, credit and many other domains.
Question 3 of 10, Understanding How Laws, Standards and Frameworks Apply to AI
Under the EU AI Act, which obligation applies to ALL providers of general-purpose AI models, including those without systemic risk?
- AKeeping technical documentation and putting in place a policy to comply with EU copyright law Correct answer
- BConducting adversarial testing and reporting serious incidents to the AI Office within a deadline
- CObtaining a notified body certificate before the model may be made available in the EU market
- DRegistering every downstream AI system built on the model in the EU high-risk system database
Answer and explanation: Article 53 requires all GPAI model providers to draw up and maintain technical documentation, provide information to downstream providers, put in place a copyright compliance policy (including respecting text-and-data-mining opt-outs) and publish a summary of training content. Adversarial testing, serious-incident reporting and cybersecurity duties are additional obligations for models with systemic risk under Article 55. There is no notified-body certificate for GPAI models, and downstream systems are registered only if they are high-risk.
Question 4 of 10, Understanding How Laws, Standards and Frameworks Apply to AI
A provider is training a general-purpose AI model and estimates that the cumulative compute used for training will exceed 10^25 floating-point operations. What is the significance of this threshold under the EU AI Act?
- AThe model is presumed to have systemic risk, and the provider must notify the Commission Correct answer
- BThe model is automatically prohibited, because the Act bans training models above this level of compute in the EU
- CThe model is exempt from GPAI obligations, because very large models are regulated only by the high-risk rules
- DThe threshold matters only for open-source models; proprietary models are classified only by their user numbers
Answer and explanation: A GPAI model is presumed to have high-impact capabilities, and so to be a model with systemic risk, when the cumulative training compute exceeds 10^25 FLOPs; the provider must notify the Commission and then meets the additional Article 55 obligations, though it may present arguments against the classification. The Act does not ban such models. Large models are not exempt from GPAI rules. The threshold applies to open and proprietary models alike, and the Commission can also designate models using other criteria.
Question 5 of 10, Understanding How to Govern AI Development
A team is designing human oversight for a high-risk AI system used to prioritise emergency calls in the EU. Which design feature BEST meets the AI Act's human oversight expectations?
- AOperators may override the system only after obtaining written approval from the provider's support team
- BOperators can understand the system's limits, are warned about automation bias, and can override or stop it Correct answer
- COperators are not shown the system's confidence levels, so they are not tempted to second-guess its outputs
- DOperators can view outputs only after decisions are executed, to avoid slowing response times in an emergency
Answer and explanation: Article 14 requires high-risk systems to be designed so overseers can properly understand capacities and limitations, remain aware of automation bias, correctly interpret outputs, decide not to use or override the output, and interrupt the system through a stop procedure. Seeing outputs only after execution prevents oversight. Hiding confidence information hinders interpretation. Requiring provider approval to override defeats real-time oversight.
Question 6 of 10, Understanding How to Govern AI Development
A provider wants to test its high-risk AI system for employment screening in real-world conditions before placing it on the EU market, outside a regulatory sandbox. What does the AI Act require?
- AApproval from every member state's data protection authority, followed by publication of all test results online
- BTesting is not allowed at all for high-risk systems until they have been registered and CE-marked by the provider
- CA testing plan submitted to the market surveillance authority, participants' informed consent and other safeguards Correct answer
- DNothing, because testing before market placement is excluded from the AI Act as research and development activity
Answer and explanation: Article 60 permits testing of Annex III high-risk systems in real-world conditions outside sandboxes, subject to conditions including a real-world testing plan submitted to and not objected to by the market surveillance authority, registration, limited duration, informed consent of participants and the ability to reverse or disregard outputs. The R&D exclusion does not cover real-world testing with people. Approval from every DPA and publication of results are not required. Testing before CE marking is exactly what the article allows.
Question 7 of 10, Understanding How to Govern AI Development
A user exercises the right to erasure, and the company must consider whether a model trained on that user's data is affected. Which statement about "machine unlearning" is MOST accurate?
- AMethods to remove a record's influence without retraining exist but are immature, so retraining may be needed Correct answer
- BDeleting the record from the source database automatically removes its influence from every model trained on it
- CErasure rights never extend to models, so the company need only delete the record from its databases
- DUnlearning is a legal procedure in which a court orders a model to forget specific personal information
Answer and explanation: Machine unlearning research aims to remove the influence of specific data without retraining from scratch, but methods are still maturing and verification is difficult, so organizations may need periodic retraining or other measures, and should design traceability for this. Deleting source records does not change a trained model. Whether erasure affects a model depends on whether personal data can be extracted from it, so it cannot be ruled out. Unlearning is a technical, not a judicial, process.
Question 8 of 10, Understanding How to Govern AI Deployment and Use
An EU company's customer chatbot, built on a US vendor's API, leaks other customers' order details in responses. The leak affects 3,000 customers and is likely to pose a risk to their rights. What must the company do under the GDPR as controller?
- ANotify the supervisory authority within 72 hours of awareness and assess whether to inform customers Correct answer
- BNotify only the US vendor, which as model provider is responsible for reporting the breach to authorities
- CNotify the European AI Office within 15 days, since breaches caused by AI fall under the AI Act instead
- DDo nothing unless customers complain, because data exposed by a chatbot does not count as a personal data breach
Answer and explanation: Under GDPR Article 33, the controller must notify the competent supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of a personal data breach likely to result in a risk, and under Article 34 must inform affected individuals if the risk is high. The vendor as processor must notify the controller but does not take over the controller's duty. The AI Office does not handle GDPR breaches. Disclosure of personal data through a chatbot is a personal data breach.
Question 9 of 10, Understanding How to Govern AI Deployment and Use
A software company integrates a third-party general-purpose AI model into a product it sells to EU employers for ranking job applicants. Which statement about its role under the EU AI Act is MOST accurate?
- AIt is the provider of a high-risk system, using information from the GPAI provider to comply Correct answer
- BIt is only a deployer, because the underlying model was developed and trained by another company outside the EU
- CIt has no obligations, because the GPAI model provider is responsible for all downstream systems
- DIt is an importer, because it brings a model from another company into its own commercial product
Answer and explanation: A company that builds an AI system on top of a GPAI model and places it on the market under its own name for a high-risk purpose (employment) is the provider of that high-risk system; GPAI providers must supply downstream providers with information to help them comply. It is not merely a deployer since it places the system on the market. The GPAI provider's duties do not cover the downstream system's compliance. Importer status concerns placing a non-EU provider's system on the EU market.
Question 10 of 10, Understanding How to Govern AI Deployment and Use
A deployer of a high-risk AI system in the EU wants to change how the system is used so that it now decides eligibility for a different public benefit than the one in its intended purpose. What is the governance and legal implication?
- AChanging the intended purpose can make the deployer a provider, with provider obligations Correct answer
- BNo implication, because deployers may use high-risk systems for any purpose once they have bought them
- CThe change is allowed if the deployer informs users afterwards in its annual report to the public
- DThe original provider automatically becomes responsible for the new purpose chosen by the deployer
Answer and explanation: Under Article 25, modifying the intended purpose of a high-risk system (or of a system so that it becomes high-risk) makes the deployer a provider with the corresponding obligations, including conformity assessment; good governance also requires a new impact assessment for the new use. Deployers must use systems according to their instructions. Informing users later does not satisfy these duties. The original provider is not responsible for a purpose it did not intend.