AIGP-M1-AI and ML Key Terms

This multiple choice assessment focuses on the AIGP body of knowledge for Module 1 - AI/ML Key Terms

AI governance concepts covering the AI/ML/deep learning hierarchy, the four learning paradigms, agentic AI, model-type trade-offs, and specific architectures (CNNs, RAG, collaborative filtering, classic models, clustering, transformers, autoencoders)

Format: Multiple Choice

The result will be provided immediately, with details on all questions.

Full Name

Email

1. Which of the following correctly describes the relationship between artificial intelligence (AI), machine learning (ML), deep learning, and generative AI?

A. They are four separate, non-overlapping fields of computer science

B. Machine learning is a subset of AI, deep learning is a subset of machine learning, and generative AI is an application area that typically draws on deep learning techniques to create new content

C. AI is a subset of machine learning, and generative AI is a subset of deep learning

D. Generative AI replaced AI as the primary field once deep learning was introduced

2. A hospital research team has 50,000 chest X-rays, but only 200 have been labelled “abnormal” or “normal” by radiologists, because expert review is expensive and slow. Which learning approach makes the best use of both the labelled and unlabelled data, and what is the key governance consideration?

A) Supervised learning; the main risk is that the dataset is too large for the model to process

B) Unsupervised learning; the main risk is that the resulting clusters can never have clinical meaning

C) Semi-supervised learning; the main governance consideration is validating that patterns learned mostly from the unlabelled data dont introduce bias that the small labelled set is too small to catch

D) Reinforcement learning; the main risk is the absence of a defined reward signal

3. A logistics company deploys a reinforcement learning agent to optimise delivery routes, rewarding it for minimising delivery time. After deployment, the agent starts routing drivers through roads marked as restricted, because doing so improves its reward score. Which governance risk does this scenario illustrate?

A) Data drift, where the statistical properties of input data change over time

B) Reward hacking (specification gaming), where the agent optimises literally for the stated reward rather than the outcome actually intended

C) Overfitting, where the model memorises specific training examples rather than generalising

D) Hallucination, where the model fabricates false information in its output

4. Compared with a standard, single-turn generative AI assistant, which of the following best describes a governance challenge unique to agentic AI systems?

A) Agentic systems require more storage because they always use larger models

B) Agentic systems can independently plan and execute multi-step actions across tools and systems, making it harder to define clear human accountability and intervention points before harm occurs

C) Agentic systems cannot produce biased outputs, since they verify facts before acting

D) Agentic systems are exempt from data protection law because their actions are autonomous

5. From an AI governance and audit perspective, what is a key advantage of an open-source (open-weight) model over a proprietary model when an organisation needs to demonstrate compliance to a regulator?

A) Open-source models guarantee higher accuracy on benchmark tests

B) Open-source models allow direct inspection of model weights and architecture, supporting independent verification and more detailed documentation for audits

C) Open-source models remove the need for any governance documentation

D) Open-source models are automatically certified as compliant with all AI regulations

6. Which governance consideration is most directly tied to the choice between deploying a small model versus a large model for a given task?

A) Only large models require a privacy impact assessment

B) The environmental footprint and compute cost of training and running the model, which scales significantly with model size

C) Small models cannot be biased, because they have fewer parameters

D) Large models are automatically exempt from explainability requirements

7. A hiring platform uses a multimodal model that evaluates candidates using both their résumé text and a short video interview. Compared with a text-only evaluation model, what additional governance risk does the multimodal design introduce?

A) The model becomes exempt from anti-discrimination law because it processes more than one data type

B) The model could infer or be influenced by protected characteristics, such as race, gender, or disability, from visual and audio cues, introducing discrimination risks that text alone would not reveal

C) Multimodal models are always more accurate, which eliminates any bias concerns

D) Adding a second data type has no effect on governance requirements

8. Under a risk-based framework like the EU AI Act, why do documentation requirements often differ between a classic classification model and a generative model performing a similar business function (for example, risk scoring versus drafting a risk report)?

A) Classic models never require documentation because they are considered too simple to pose risk

B) Generative model documentation must additionally address the risk of fabricated, biased, or harmful content, since its output isn't limited to a fixed, auditable set of labels

C) Generative models are automatically classified as minimal risk under most frameworks

D) There is no meaningful documentation difference between the two model types

9. An e-commerce marketplace wants a feature that automatically checks whether a seller's uploaded product photo meets image-quality standards (in focus, correctly oriented, product centred in frame). Which model architecture is best suited to this task, and why?

A) A transformer, because it can process any type of sequential data equally well

B) A convolutional neural network (CNN), because it excels at detecting spatial patterns such as edges, shapes, and framing within a grid of pixels

C) Collaborative filtering, because it can recommend the best-quality images based on what other sellers have uploaded

D) A clustering model, because it will automatically group similar images together

10. A law firm wants a chatbot that answers client questions using the most current version of its internal legal templates and case files, without retraining the underlying model every time a document changes. Which technique should it use, and what governance benefit does this provide over relying solely on the model's memorised training knowledge?

A) Reinforcement learning, because it lets the model learn from client feedback in real time

B) Retrieval-Augmented Generation (RAG), because it grounds each answer in the current, retrievable source documents, improving traceability and reducing the risk of relying on outdated internal knowledge

C) A larger language model, because bigger models automatically memorise more documents

D) Collaborative filtering, because it finds similar past client questions to reuse previous answers

11. A streaming service wants to recommend new songs to a listener based on what other users with similar listening habits enjoy, without analysing the actual audio content of the songs. Which technique fits this approach, and what governance risk does it raise?

A) A convolutional neural network, because it can inspect the song's audio waveform directly for genre similarity

B) Collaborative filtering, with the main governance risk being that recommendations rely on other users' behavioural data, which can unintentionally reveal or profile sensitive listening patterns

C) A clustering model, because it will group songs together by genre metadata

D) An autoencoder, because it compresses song audio into features for direct comparison

12. An insurer wants to predict, from structured historical data (applicant age, vehicle type, past claims count), whether a new policyholder is likely to file a claim within the first year, in order to set a premium tier. A governance review requires that the model's reasoning be traceable and explainable to a regulator on request. Which model category best satisfies this, and why?

A) A large language model, because it can explain its own reasoning in natural language

B) A classic model, such as logistic regression or a decision tree, because it maps structured inputs directly to a bounded outcome in a way that is straightforward to trace and audit

C) A generative model, because it can create synthetic policyholders to test edge cases

D) A multimodal model, because it can combine text and images from the application form

13. A telecom company wants to segment its customer base into groups with similar usage patterns (data, calls, texts) to design targeted retention offers, without any predefined labels for what those segments should be. Which approach fits, and what governance safeguard is especially important here?

A) Supervised learning, with the main safeguard being verifying that every customer was labelled correctly

B) Clustering models, with the main safeguard being checking that the resulting segments don't act as an unintentional proxy for protected characteristics such as age or income

C) Reinforcement learning, with the main safeguard being defining an appropriate reward function

D) Retrieval-Augmented Generation, with the main safeguard being validating the retrieved source documents

14. A researcher wants to build a model that reads an entire paragraph and understands how each word relates to every other word in context, including resolving what a pronoun refers to several sentences later. Which architecture is best suited to this, and why?

A) A decision tree, because it can branch based on individual word features

B) A transformer, because its attention mechanism lets the model weigh the relevance of every word in the input to every other word, regardless of distance between them

C) A convolutional neural network, because it processes local neighbourhoods of data effectively

D) Collaborative filtering, because it identifies patterns across similar documents

15. A manufacturer wants a model that learns what “normal” sensor readings look like on its factory equipment, then flags unusual patterns as possible faults, without being given any prior examples labelled “faulty”. Which model type is well suited to this, and how does it work?

A) A random forest, because it can classify data into categories based on learned rules

B) An autoencoder, because it learns to compress and reconstruct normal sensor data, and when new data doesn't reconstruct well, that poor reconstruction signals an anomaly

C) A large language model, because it can read the sensor logs as if they were text

D) Collaborative filtering, because it flags equipment behaving differently from other similar machines