Font size
WorksheetsGenerative AI/Fairness and Ethics
Total questions: 52
Worksheet time: 26mins
What is the fundamental difference between a generative AI model and a predictive AI model?
Predictive models must always be trained on image data, while generative models only use text documents.
Generative models create new content, such as images or text, while predictive models classify or forecast outcomes.
Generative models use unsupervised learning exclusively, while predictive models always use supervised learning techniques.
Predictive models eliminate all data bias, while generative models are known to introduce ethical risks into the system.
Which Azure AI Service provides secure, enterprise-grade access to large-scale foundation models like GPT-4 and DALL-E 3?
Azure Machine Learning Studio, which requires custom deployment of open-source models for use.
Azure Cognitive Services, which primarily offers pre-built models for tasks like speech recognition.
Azure OpenAI Service, which ensures models run within a secure, isolated Azure environment.
Azure Data Factory, which is used mainly for data migration and complex ETL pipelines.
What is the primary purpose of a Large Language Model (LLM) during its initial training phase?
To accurately identify and label specific objects within a large volume of complex image data sets.
To analyze the preceding text sequence and predict the most statistically probable next word or token.
To detect and eliminate all forms of bias and ethical risk from the foundational training data set.
To segment customer data into distinct clusters for use in unsupervised learning tasks.
A model that processes and generates content across two different modalities, such as accepting an image input and generating a text description, is classified as what type of AI model?
Multimodal AI, which handles multiple data types like text and images in a unified system.
Discriminative AI, which focuses on providing a single label or classification output.
Regression AI, which predicts a continuous numerical value as its final output.
Temporal AI, which specializes in time-series forecasting and sequence prediction.
Which of the following is NOT typically considered a generative AI use case?
Automatically synthesizing a detailed marketing email draft from a short bulleted list.
Generating new, high-resolution product images for an e-commerce website.
Creating new Python code snippets based on a natural language instruction prompt.
Calculating the exact price forecast for a stock over the next 30 days based on historical data.
Which capability within Azure AI enables a large language model (LLM) to reference external, authoritative enterprise data when formulating a response?
Data grounding, often implemented via Azure AI Search to retrieve relevant documents.
Tokenization, the process of breaking down text into smaller processing units.
Parameter Efficient Fine-Tuning (PEFT), used for faster model adaptation.
Content filtering, used only for detecting and preventing harmful outputs.
A developer implements a safety mechanism to scan and block harmful text generated by the LLM before the user sees it. Which part of Azure AI Content Safety performs this function?
The Prompt Safety Filter, which is primarily responsible for analyzing user input only.
The Completion Safety Filter, which analyzes the model's output before delivery.
The Responsible AI Dashboard, which provides post-deployment analysis and reporting.
The Custom Translator service, which manages language localization.
Which Microsoft service allows organizations to build, deploy, and govern their own custom machine learning and AI models, in addition to foundation models?
Azure AI Search, which provides document indexing and vector storage capabilities.
Azure Cosmos DB, a global, multi-model NoSQL database service.
Azure Machine Learning, which provides an end-to-end platform for the MLOps lifecycle.
Azure AI Content Safety, which is only used for content moderation filtering.
What term describes the technique of feeding the LLM a few examples of input-output pairs in the prompt to guide its response format and style?
Zero-shot learning, where no examples are provided in the prompt context.
Full fine-tuning, which requires training the entire model on a large dataset.
Reinforcement learning, used to train models through environmental rewards.
Few-shot learning (or In-Context Learning), using examples provided in the prompt.
Which Microsoft technology integrates AI capabilities directly into low-code platforms like Power Apps and Power Automate?
Microsoft Fabric, which focuses on data analytics and data warehousing capabilities.
AI Builder, which provides pre-built and custom AI models for the Power Platform.
Azure Kubernetes Service (AKS), used for large-scale container orchestration.
Visual Studio Code, which is the primary developer environment for coding.
A customer is using Azure OpenAI and wants to ensure the model responds with the strict persona and tone defined for their brand. Which configuration detail should they use?
Adjusting the model's temperature setting to a higher value for more creative responses.
Increasing the number of tokens allowed in the model's context window size.
Defining a detailed instruction set in the system message of the API request.
Requiring the model to only use proprietary data for grounding its output.
When using Azure OpenAI, which responsibility related to ethical deployment belongs specifically to the customer (user), rather than to Microsoft?
Ensuring the foundational model (e.g., GPT-4) was pre-trained on diverse and unbiased data.
Providing physical security and governance for the Azure data centers globally.
Managing the underlying compute hardware and GPU allocation for the service.
Defining and implementing human oversight and content policy enforcement for the specific application.
A company wants to use a generative AI model to automatically summarize long legal contracts. Which type of generative task is this?
Classification, which predicts a single label based on input features.
Text summarization, a form of natural language generation.
Image generation, which synthesizes visual output from a text prompt.
Anomaly detection, which identifies outliers in the contract data.
Which feature ensures that an LLM connected to enterprise data will only discuss information found within those documents, effectively preventing the model from drawing on its general training knowledge?
Fine-Tuning the model weights on the small set of enterprise documents.
Using zero-shot learning by simply asking the question in the prompt.
Implementing tokenization algorithms to compress the input data.
Data grounding via Retrieval-Augmented Generation (RAG).
If a user submits a prompt that violates the defined content policy (e.g., a hate speech prompt), which component of Azure AI Content Safety immediately intervenes?
The Completion Safety Filter, which only monitors the model's generated response output.
The Prompt Safety Filter, which analyzes the user's input before it reaches the model.
The Responsible AI Dashboard, which is a reporting and analysis tool only.
The Text to Speech Service, which converts text to auditory output.
Which of the following is one of Microsoft's six core Responsible AI principles?
Efficiency, which focuses on running models with the lowest possible server cost.
Scalability, which focuses on increasing deployment capacity on demand.
Fairness, which ensures AI systems treat all people equitably and without prejudice.
Optimization, which focuses on achieving the highest accuracy metric.
The Responsible AI principle of Accountability primarily emphasizes which requirement?
The AI model must be able to self-correct its own errors in production without human intervention.
People and organizations must be responsible for the AI's impact, requiring human governance and oversight.
The AI system must provide a clear record of every single decision it makes for audit purposes.
The AI system must only use data that is publicly available and free of all security restrictions.
An AI system that fails unexpectedly when exposed to novel inputs outside of its training data, causing a critical failure, is primarily violating which Responsible AI principle?
Transparency, because the user cannot see the model's internal decision process.
Fairness, because the failure disproportionately affects one demographic group.
Privacy and Security, because the failure exposes sensitive user data.
Reliability and Safety, because the system does not behave as intended or safely.
What is the main goal of the Responsible AI principle of Inclusiveness?
Designing AI systems that consider the diverse range of human abilities, languages, and backgrounds.
Ensuring the core data used to train the model is sourced from publicly available datasets only.
Reducing the number of required features to simplify the model for low-resource environments.
Requiring that all AI model outputs are translated into every official language before deployment.
Which Responsible AI principle requires that AI systems be accompanied by clear documentation of their capabilities, limitations, and decision-making processes?
Transparency, to ensure users understand how the AI works and its limitations.
Accountability, to ensure that human review is required for all outputs.
Privacy and Security, to guarantee that data transmission is encrypted.
Fairness, to ensure equitable outcomes for all demographic groups.
Which type of bias occurs when the data collected to train an AI model is not representative of the real-world population it will serve, leading to poor performance for certain groups?
Representational bias, where certain groups are underrepresented or stereotyped in the training data.
Historical bias, where data reflects past societal prejudices and inequities.
Measurement bias, where the way data is collected or labeled is flawed or inconsistent.
Algorithmic bias, where the model design itself introduces systematic errors.
An AI model used for resume screening was trained on 20 years of historical hiring data which favored male applicants. This resulted in the model unfairly penalizing female applicants. This is an example of which bias type?
Measurement bias, caused by inconsistent feature labeling in the input data.
Representational bias, caused by the underrepresentation of a group in the collected samples.
An AI system used to predict house prices consistently undervalues properties owned by a specific minority group due to flawed, inconsistent labeling of property features during data collection. This is an example of which bias type?
Historical bias, where past prejudices in real estate records skewed the data.
Measurement bias, where the process of collecting and labeling data is flawed or inconsistent.
Representational bias, where the model fails to include diverse geographic areas in training.
Algorithmic bias, where the model design itself is inherently flawed and introduces errors.
The principle of Privacy and Security in Responsible AI emphasizes which key action?
Ensuring that all AI systems are deployed exclusively on private, on-premises cloud servers.
Requiring AI systems to self-correct any internal errors and inconsistencies automatically.
Ensuring that personal data is protected, encrypted, and used only with explicit consent.
Mandating that all AI-generated content is watermarked to confirm its synthetic origin.
Which statement best describes a consequence of an AI system violating the Fairness principle?
The system fails to meet its required low latency threshold during peak user activity.
The system generates offensive or hateful text output to the end user.
The system consistently produces inaccurate results when dealing with time-series data.
The system provides a statistically worse quality of service for a particular demographic group.
Which of the following is an example of Allocative Harm caused by an AI system?
A language model generating toxic language when given a hostile prompt by a user.
A computer vision system failing to detect objects in dimly lit environments.
A chatbot providing an inaccurate weather forecast due to bad input data.
An AI loan approval system denying credit unfairly based on the applicant's gender.
What is the primary method an organization should use to ensure Accountability in an AI deployment?
Allowing the AI system to make all final decisions automatically without human intervention.
Establishing human-in-the-loop (HITL) procedures and clearly assigning ownership and oversight roles.
Implementing a Generative Adversarial Network (GAN) to check the output quality.
Removing all logging and monitoring systems to protect the user's privacy.
Which of the following is a common practice used to reduce Representational Bias in an AI model's training?
Using complex proprietary algorithms that automatically correct all known data flaws.
Deploying the model to only a single, geographically limited region initially.
Actively seeking out and collecting more diverse and representative data samples for training.
Removing all feature columns related to demographics from the training dataset entirely.
Which two of the following are the primary components of Microsoft's Responsible AI governance strategy?
Core principles (Fairness, Reliability, Transparency, etc.)
Server optimization guides for maximum throughput
Mandatory use of open-source models only
Tools and documentation (e.g., Azure AI Content Safety)
What is the foundational element that underpins Microsoft's entire Responsible AI governance strategy?
A set of six core principles (e.g., Fairness, Accountability, Transparency).
The exclusive use of Azure Machine Learning for all AI development projects.
A mandate for continuous, unsupervised self-correction by all deployed AI models.
The requirement that all AI models must be trained exclusively on synthetic data.
Which practice best supports the Transparency principle in a generative AI application?
Using obfuscation techniques to hide the model's input features from the end user.
Limiting the amount of data logging to protect the model's proprietary data integrity.
Clearly labeling AI-generated content (e.g., images, text) as synthetic to the user.
Requiring the model to achieve a minimum accuracy score of 99% before deployment.
You are deploying a chatbot for healthcare. Which Privacy and Security action is mandatory?
Ensuring all sensitive patient data is encrypted both during transmission and while stored at rest.
Implementing a RAG system to ground the model in external, verifiable clinical studies only.
Using the Completion Safety Filter to prevent the model from generating toxic language.
Documenting the model's training data sources for the Transparency principle.
A financial AI model provides less accurate fraud detection scores for transactions originating from a certain country. This is a failure of the Fairness principle, likely resulting from which bias source?
Historical bias, due to outdated regulations reflecting past political decisions.
Measurement bias, due to errors in how the transaction volume was aggregated.
Representational bias, due to the underrepresentation of that country's transaction patterns in the training data.
Algorithmic bias, due to the model's use of a complex deep learning architecture.
An AI application generates marketing copy in English only, failing to serve customers who use Spanish as their primary language. This violates which Responsible AI principle?
Reliability and Safety, as the model could fail to execute commands in Spanish.
Accountability, as no one is monitoring the Spanish language market.
Transparency, as the lack of Spanish support is not explicitly documented.
Inclusiveness, as the design fails to consider linguistic diversity and access.
To ensure Accountability in a high-stakes AI decision (e.g., granting parole), an organization must establish which mechanism?
A complex tokenization algorithm to process the input prompt quickly.
A human review and final approval process for all model recommendations.
An open-source model deployment to allow for full public auditing access.
An automatic retraining schedule to constantly update the model weights.
A developer is building a system where a user provides a voice command, and the system responds with a generated image. Which two core Azure AI services will they need?
Azure AI Translator and Azure AI Content Safety
Azure Bot Service and Azure Machine Learning
Azure AI Search and Azure Cosmos DB
Azure AI Speech and Azure OpenAI Service (for DALL-E)
A company wants to use an LLM for customer service but needs to ensure the model never shares pricing information that is currently being negotiated. Which Azure service configuration should they use?
Setting the model's temperature parameter to a very low value (e.g., 0.1).
Using the Completion Safety Filter to block all numerically generated content.
Implementing a carefully crafted system message to restrict the model's knowledge boundary.
Forcing the model to use only publicly available data for grounding its responses.
Which practice best adheres to the Reliability and Safety principle during the deployment phase of an LLM?
Limiting the number of users who can access the LLM at any given time.
Performing red teaming and adversarial testing to discover and fix potential misuse vulnerabilities.
Only using models trained on entirely synthetic and non-biased data sets.
Documenting all data sources used to train the model for audit purposes.
What is the primary function of Azure AI Search when integrated with Azure OpenAI Service?
It manages user authentication and authorization for the LLM endpoints.
It is used to generate the LLM's initial training data and token vocabulary.
It provides the ability to store, index, and retrieve vector embeddings for RAG/grounding.
It automatically translates the LLM's output into multiple foreign languages.
Which two of the following are explicit goals of Azure AI Content Safety?
Blocking prompts that contain hate speech or sexually explicit material.
Preventing the LLM from generating violent or self-harm-related text outputs.
Calculating the model's fairness metrics against a given demographic data set.
Automatically translating the user's input into the LLM's native training language.
A university uses an AI system to grade essays, but the system performs significantly worse when grading non-native English speakers' work. This is a violation of Fairness due to which specific type of harm?
Denigration harm, caused by the use of abusive language in the output.
Quality of Service harm, caused by unequal performance across a demographic group.
Allocative harm, caused by the unfair withholding of a resource (a passing grade).
Historical harm, caused by reflecting outdated societal prejudices in the data.
Which of the following is a key reason why Transparency is essential for high-stakes AI systems (e.g., medical diagnosis)?
It allows regulatory bodies to check if the model used the most cost-effective resources.
It ensures the model can be fine-tuned quickly with minimal effort and data.
It reduces the need for human oversight since the AI is already documented.
It allows human experts to understand the AI's reasoning, verify its outputs, and catch errors.
Which step is most crucial for establishing Privacy and Security when integrating an LLM with a sensitive financial database?
Storing all of the LLM's generative outputs in a publicly accessible log file.
Implementing robust access controls (RBAC) and data masking to limit the LLM's data view.
Using open-source models exclusively to allow external security audits globally.
Forcing the LLM to use only zero-shot learning for all financial prediction tasks.
A developer wants to enable a user to type a request and receive a completely new, unique poem. Which fundamental AI capability does this scenario rely on?
Prediction, which involves classifying the user's request into a fixed category.
Generation, which involves creating a novel text sequence (the poem).
Clustering, which involves grouping similar poems based on their content.
Recognition, which involves identifying the user's intent from the text.
Which action is the most effective way to address a violation of the Historical Bias principle in a deployed AI model?
Immediately deleting the model from production and starting the entire project over.
Ignoring the historical bias if the model's overall accuracy metric is high enough.
Carefully re-examining and re-labeling the historical training data to mitigate biased samples and then retraining the model.
Applying a post-processing filter to the model's final output to manually override biased decisions.
Which of the following best demonstrates the principle of Inclusiveness in an AI application's design?
The application offers multiple input methods, such as voice commands and keyboard typing.
The application restricts the user interface to only the English language for simplicity.
The application requires users to have a high-speed, low-latency internet connection.
The application only provides output in text format, avoiding visual or auditory cues.
The goal of Red Teaming in a generative AI project is to proactively test for which key issue?
The model's latency and performance metrics under high load conditions.
The model's compliance with all external open-source licensing agreements.
The model's ability to achieve maximum accuracy on its validation dataset.
The system's vulnerability to adversarial prompts, jailbreaks, and unsafe outputs.
Which Azure service would you use to create a knowledge base (vector index) from thousands of internal company documents to support a Retrieval-Augmented Generation (RAG) system?
Azure Cosmos DB, for storing key-value pairs of user data.
Azure Kubernetes Service, for running containerized microservices.
Azure AI Search, for indexing, vectorizing, and retrieving documents.
Azure Data Lake Storage, for archiving large, cold storage data files.
In the context of generative AI, what is the term for a response generated by the LLM that is factually incorrect or completely fabricated?
Overfitting, which is when the model performs poorly on training data.
Algorithmic bias, which is an error in the model's core code structure.
Data leakage, which is when sensitive training data is exposed.
Hallucination, which is the generation of non-factual or nonsensical content.
An organization decides to use only open-source, non-proprietary models for its generative AI development. Which Responsible AI principle does this decision primarily support?
Reliability and Safety, by using community-validated model checkpoints.
Transparency, by allowing external stakeholders to inspect the model's internal architecture and weights.
Fairness, by ensuring the model is easily accessible to all developers globally.
Accountability, by eliminating the need for internal human oversight.
Which Azure service must be used to set up the system message and safety filters for a GPT-4 deployment?
Azure Machine Learning Studio, which manages all custom model training.
Azure Cosmos DB, which stores the system message as a configuration file.
Azure OpenAI Service, which hosts the deployment and its governance settings.
Azure AI Search, which indexes the system message for quick retrieval.
Extra Credit:
Available for outstanding performance
Mandatory for all students
Given for poor attendance
Not allowed in this course
