SkillTestPro | AI-900 Practice Test With Answers

AI-900 Practice Test 1

AI-900 Practice Test

1 / 80

You are implementing a facial recognition system. Which aspect of Responsible AI is crucial to consider to prevent unintended bias against certain demographic groups?

Transparency

Reliability and Safety

Accountability

Privacy and Security

Fairness

2 / 80

Which Azure AI service would be used to build intelligent agents that can understand and respond to user queries in a natural conversational manner, often seen in chatbots?

Azure AI Vision

Azure AI Speech

Azure AI Language

Azure Bot Service

Azure AI Search

3 / 80

What is the primary purpose of a "label" in a supervised machine learning dataset?

It represents an input feature that describes the data.

It is the target variable that the model is trying to predict.

It is a unique identifier for each data record.

It is used to group similar data points together.

It indicates the source of the data.

4 / 80

Which type of AI workload involves teaching an AI model to understand, interpret, and generate human language?

Computer Vision

Anomaly Detection

Knowledge Mining

Natural Language Processing (NLP)

Conversational AI

5 / 80

You need to train a machine learning model but have limited data science expertise. Which Azure Machine Learning capability allows you to automate the process of model selection and hyperparameter tuning?

Azure Machine Learning Studio designer

Automated Machine Learning (AutoML)

Azure Machine Learning compute instances

Azure Machine Learning datasets

MLflow integration

6 / 80

A company wants to analyze customer reviews to determine whether the sentiment expressed is positive, negative, or neutral. Which Natural Language Processing (NLP) capability would be most suitable for this task?

Key Phrase Extraction

Entity Recognition

Sentiment Analysis

Language Modeling

Speech-to-Text

7 / 80

Which Azure AI service would you use to detect common objects, such as cars or people, within an image?

Azure AI Language

Azure AI Speech

Azure AI Vision

Azure OpenAI Service

Azure Machine Learning

8 / 80

You are building a machine learning model to predict the sales price of a house based on its features (e.g., number of bedrooms, square footage, location). Which type of machine learning problem is this?

Classification

Clustering

Regression

Anomaly Detection

Reinforcement Learning

9 / 80

Which of the following is a core principle of Responsible AI, focusing on ensuring that AI systems are fair and do not discriminate against individuals or groups?

Reliability and Safety

Privacy and Security

Transparency

Fairness

Accountability

10 / 80

What is the primary goal of Artificial Intelligence (AI)?

to develop software that can store vast amounts of data

to enable machines to perform tasks that typically require human intelligence

to create robots that can move and interact with the physical world

to analyze historical data to identify trends and patterns

to design complex algorithms for mathematical calculations

11 / 80

Which Responsible AI principle focuses on ensuring an AI system's decision-making process can be explained and understood by the people affected by it?

Transparency

Accountability

Privacy and security

Reliability and safety

Fairness

Inclusiveness

12 / 80

Which Responsible AI principle involves ensuring that AI systems empower everyone and engage people, regardless of physical ability, gender, sexual orientation, or other characteristics?

Privacy and security

Transparency

Reliability and safety

Accountability

Fairness

Inclusiveness

13 / 80

Which Responsible AI principle requires that people who design and deploy AI systems are held responsible for how those systems operate?

Accountability

Reliability and safety

Privacy and security

Transparency

Inclusiveness

Fairness

14 / 80

Which Responsible AI principle focuses on ensuring an AI system performs consistently as intended, handles unexpected situations gracefully, and does not cause harm?

Accountability

Reliability and safety

Privacy and security

Inclusiveness

Transparency

Fairness

15 / 80

Which Responsible AI principle requires that personal data used to train or operate an AI system is protected and handled in compliance with applicable regulations?

Fairness

Accountability

Transparency

Reliability and safety

Inclusiveness

Privacy and security

16 / 80

Which type of AI workload would be used to predict the likely sale price of a house based on features such as square footage, location, and number of bedrooms?

A computer vision workload

A clustering workload, which groups similar items rather than predicting a specific numeric value

A knowledge mining workload

A regression machine learning workload

A classification workload, which predicts a discrete category rather than a continuous numeric value

A natural language processing workload

17 / 80

Which type of AI workload would be used to group customers into segments based on similar purchasing behavior, without predefined category labels?

A computer vision workload

A natural language processing workload

A classification workload, which requires predefined category labels rather than discovering groups without labels

A regression workload, which predicts a continuous numeric value rather than grouping unlabeled items

An anomaly detection workload

A clustering machine learning workload

18 / 80

Which type of AI workload would be used to identify fraudulent credit card transactions that deviate significantly from a customer's normal spending pattern?

An anomaly detection workload

A natural language processing workload

A reinforcement learning workload

A regression workload, which predicts a continuous value rather than identifying unusual deviations from a pattern

A clustering workload, which groups similar items rather than specifically flagging unusual outliers

A computer vision workload

19 / 80

Which AI workload category would extract structured, searchable insights, such as key phrases and entities, from a large volume of unstructured documents like contracts?

Clustering, which groups similar items rather than extracting structured insights from a document's content

Regression, which predicts numeric values rather than extracting structured insights from unstructured documents

A generative AI workload alone, which creates new content rather than specifically extracting insights from existing documents

Computer vision alone, which primarily processes image data rather than the broader unstructured document text

Knowledge mining

An anomaly detection workload

20 / 80

Which consideration is most important when an organization deploys a facial recognition system for building access control, given the potential for the technology to be misused for surveillance?

Assuming legal and ethical considerations only apply to publicly deployed, not internal, AI systems

Assuming Responsible AI principles are irrelevant to a building access control use case specifically

Assuming accuracy is the only consideration, with no attention to privacy or potential misuse

Establishing appropriate governance, oversight, and Responsible AI principles such as accountability and privacy to guide the deployment

Deploying the system with no governance or oversight structure of any kind

Assuming facial recognition technology carries no meaningful risk of misuse whatsoever

21 / 80

Which machine learning training approach uses a dataset where each example is paired with a known, correct output label to train the model?

Self-supervised learning

Transfer learning

Reinforcement learning

Supervised learning

Semi-supervised learning

Unsupervised learning

22 / 80

Which machine learning training approach works with unlabeled data, identifying hidden patterns or structures, such as natural groupings, without predefined correct answers?

Reinforcement learning, which learns through trial-and-error rewards rather than discovering structure in unlabeled data

Semi-supervised learning, which uses a combination of labeled and unlabeled data rather than purely unlabeled data

Transfer learning

Unsupervised learning

Self-supervised learning

Supervised learning, which requires labeled data with known correct outputs

23 / 80

Which machine learning training approach involves an agent learning to make decisions by receiving rewards or penalties based on the outcomes of its actions within an environment?

Self-supervised learning

Transfer learning

Supervised learning

Reinforcement learning

Semi-supervised learning

Unsupervised learning

24 / 80

Which term describes the specific measurable characteristics or attributes, such as square footage or number of bedrooms, used as input to a machine learning model?

An epoch, which represents a training iteration rather than the input data's characteristics

A hyperparameter, which configures the training process rather than representing the input data's characteristics

A model, which is the trained artifact rather than the input characteristics themselves

Labels, which represent the known correct output rather than the input characteristics

An algorithm, which is the training method rather than the input characteristics

Features

25 / 80

Which term describes the known, correct output value associated with a specific set of features in a supervised learning training dataset?

An algorithm

A hyperparameter

A label

A feature, which represents an input characteristic rather than the known correct output

A model, which is the trained artifact rather than a specific data point's known output

An epoch

26 / 80

Which portion of a dataset is used exclusively to evaluate a trained model's performance on data it has never seen before, providing an unbiased assessment of generalization?

The test dataset (test set)

A synthetic dataset generated purely for demonstration purposes with no evaluation role

The validation dataset, which is typically used for tuning hyperparameters during training rather than the final unbiased evaluation

The entire original dataset with no split whatsoever

The training dataset, which is used to actually fit the model's parameters rather than provide an unbiased final evaluation

A single individual data point used repeatedly for both training and evaluation

27 / 80

Which term describes a machine learning model that performs very well on training data but performs poorly on new, unseen data, because it has learned the training data's noise rather than the underlying pattern?

Underfitting, which describes a model failing to capture the underlying pattern even in the training data itself

A classification model, which is a model type rather than a description of poor generalization

Overfitting

A regression model, which is a model type rather than a description of poor generalization

A clustering model, which is a model type rather than a description of poor generalization

A balanced, well-generalized model, which is the opposite of the described scenario

28 / 80

Which term describes a machine learning model that is too simple to capture the underlying pattern in the data, performing poorly on both training and new data?

A classification model

A balanced, well-generalized model, which is the opposite of the described scenario

Overfitting, which describes a model that performs well on training data but poorly on new data, rather than performing poorly on both

A regression model

A clustering model

Underfitting

29 / 80

Which evaluation metric for a classification model represents the proportion of all predictions that were correct?

R-squared, which is a regression metric rather than a classification metric

Mean absolute error, which is a regression metric rather than a classification metric

F1 score, which is a combined balance of precision and recall rather than simple overall correctness

Accuracy

Precision, which measures the proportion of positive predictions that were actually correct, rather than overall correctness across all predictions

Recall, which measures the proportion of actual positives that were correctly identified, rather than overall correctness

30 / 80

Which evaluation metric for a regression model measures the average magnitude of the errors between predicted and actual values, without considering their direction?

Precision, which is a classification metric rather than a regression error metric

F1 score, which is a classification metric rather than a regression error metric

Mean absolute error (MAE)

Accuracy, which is a classification metric rather than a regression error metric

Recall, which is a classification metric rather than a regression error metric

A confusion matrix, which visualizes classification results rather than measuring regression error magnitude

31 / 80

Which Azure Machine Learning feature allows a user to build, train, and evaluate a machine learning model using a visual, drag-and-drop interface without writing code?

The Azure Machine Learning SDK, which requires writing code

Azure Machine Learning notebooks, which require writing code rather than a no-code drag-and-drop interface

Azure Machine Learning designer

MLflow, which is used for experiment tracking rather than visual pipeline building

Automated machine learning (AutoML), which automatically selects and tunes algorithms rather than providing a manual drag-and-drop pipeline builder

Azure Machine Learning compute clusters, which provide compute resources rather than a visual pipeline building interface

32 / 80

Which Azure Machine Learning capability automatically tries multiple algorithms and hyperparameter combinations to find the best-performing model for a given dataset and task?

Automated machine learning (AutoML)

Azure Machine Learning notebooks, which require the data scientist to manually write and select algorithms

Azure Machine Learning designer, which is a manual, visual pipeline builder rather than an automated algorithm-selection process

MLflow, which tracks experiments rather than automatically trying algorithms

A datastore, which manages data connections rather than automating algorithm selection

A compute cluster, which provides compute resources rather than automating algorithm selection

33 / 80

Which Azure Machine Learning resource represents a scalable, managed compute environment used to run training jobs or host a code-first development environment?

A workspace alone, which is the overall container for these resources rather than the specific compute resource itself

An endpoint, which serves a deployed model rather than providing training compute resources

A model registry, which stores trained models rather than providing compute resources

A datastore, which manages data connections rather than providing compute resources

A compute cluster or compute instance

A pipeline, which orchestrates a sequence of steps rather than providing the underlying compute resources itself

34 / 80

Which Azure Machine Learning concept represents the top-level container that organizes all the resources, such as datasets, compute, and models, for a specific machine learning project?

A pipeline, which is one specific resource within a workspace

A compute cluster, which is one specific resource within a workspace rather than the overall top-level container

A workspace

A model registry entry, which is one specific resource within a workspace

A datastore, which is one specific resource within a workspace

An endpoint, which is one specific resource within a workspace

35 / 80

Which Azure Machine Learning feature allows a trained model to be deployed and made accessible for real-time predictions via a web service?

A workspace alone, without a specific deployed endpoint resource

A datastore, which manages data connections rather than serving model predictions

A pipeline, which orchestrates a sequence of processing steps rather than specifically serving real-time predictions

A compute cluster alone, without an associated deployed endpoint

A batch endpoint, which processes large volumes of data asynchronously rather than providing real-time, immediate predictions

An online endpoint (real-time endpoint)

36 / 80

Which Azure Machine Learning feature allows a trained model to process a large volume of data asynchronously, generating predictions for that entire batch at once rather than in real time?

An online endpoint (real-time endpoint), which provides immediate, synchronous predictions rather than asynchronous batch processing

A pipeline alone, without an associated deployed batch endpoint for serving predictions

A workspace alone

A datastore alone

A batch endpoint

A compute cluster alone, without an associated deployed batch endpoint

37 / 80

Which Azure Machine Learning feature allows a data scientist to define a repeatable, multi-step workflow, such as data preparation followed by training and then evaluation, that can be run consistently?

A workspace alone, which is the overall container rather than the specific repeatable workflow definition

A model registry alone, which stores trained models rather than defining the workflow that produced them

A compute cluster alone, which provides underlying resources rather than defining the workflow steps themselves

A pipeline

A datastore alone, which manages data connections rather than defining a broader workflow

An endpoint alone, which serves predictions rather than defining a multi-step training workflow

38 / 80

Which Azure Machine Learning feature centrally stores and versions trained models, making it easier to track, manage, and deploy specific model versions?

The model registry

A datastore, which manages data connections rather than versioning trained models themselves

A pipeline, which defines the training workflow rather than storing the resulting versioned models

An endpoint, which serves a specific deployed model rather than centrally storing and versioning all models

A workspace alone, which is the overall container rather than the specific model versioning and storage mechanism

A compute cluster, which provides compute resources rather than versioning trained models

39 / 80

Which Azure Machine Learning feature manages the connection details to an external data source, such as an Azure Blob Storage container, without requiring credentials to be hardcoded in scripts?

A workspace alone, which is the overall container rather than the specific data connection management mechanism

An endpoint, which serves predictions rather than managing data source connection details

A pipeline, which defines a workflow rather than managing data source connection details

A datastore

A compute cluster, which provides compute resources rather than managing data source connection details

A model registry, which stores trained models rather than managing data source connection details

40 / 80

Which Azure Machine Learning feature provides a visual, interactive way to understand how much each individual feature contributed to a specific model's prediction, supporting the Responsible AI principle of transparency?

An endpoint alone, which serves predictions rather than specifically explaining why those predictions were made

AutoML alone, which automates algorithm selection rather than explaining an already-trained model's specific predictions

A compute cluster alone, which provides compute resources rather than explaining model predictions

A pipeline alone, which defines the training workflow rather than explaining the resulting model's specific predictions

Model interpretability / explainability tools

A datastore alone, which manages data rather than explaining model predictions

41 / 80

Which Azure AI Vision capability would identify and label the specific objects present in a photograph, such as 'dog,' 'car,' and 'tree'?

Optical character recognition (OCR), which extracts text rather than identifying general objects

Image analysis (tagging)

Facial recognition, which identifies specific people rather than general objects

Object detection, which additionally provides bounding box coordinates for each identified object, rather than simply providing general descriptive tags

Spatial analysis, which analyzes movement and presence within a physical space rather than tagging static objects in a photo

Image classification alone, which assigns a single overall category to an entire image rather than multiple individual object tags

42 / 80

Which Azure AI Vision capability would identify multiple specific objects within an image and provide a bounding box indicating each object's exact location?

Optical character recognition (OCR)

Spatial analysis

Image classification alone

Facial recognition

Object detection

Image tagging alone, which provides general descriptive labels without specific bounding box location coordinates

43 / 80

Which Azure AI Vision capability assigns a single overall category label, such as 'indoor' or 'outdoor,' to an entire image based on its dominant content?

Image classification

Spatial analysis

Image tagging alone, which can apply multiple descriptive tags rather than a single, defined overall category

Object detection, which identifies multiple individual objects and their specific locations rather than a single overall category for the entire image

Facial recognition

Optical character recognition (OCR)

44 / 80

Which Azure AI Vision capability extracts printed or handwritten text from an image, such as a scanned document or a photo of a street sign?

Spatial analysis, which analyzes movement within physical space rather than extracting text from an image

Image tagging alone, which provides general descriptive labels rather than extracting the actual text content

Optical character recognition (OCR)

Facial recognition, which identifies faces rather than extracting text

Object detection, which identifies general objects rather than extracting the specific text characters within an image

Image classification, which assigns a general category rather than extracting specific text

45 / 80

Which Azure AI service specifically provides pre-built facial detection and recognition capabilities, such as identifying whether two photos likely show the same person?

Azure OpenAI Service, which is focused on generative language and image capabilities rather than specific facial recognition matching

Azure AI Vision (Face capabilities, with appropriate limited access approval)

Azure AI Search, which indexes and searches content rather than performing facial recognition

Azure AI Speech, which processes audio rather than facial images

Azure AI Content Safety, which detects harmful content rather than performing facial recognition

Azure AI Language, which processes text rather than facial images

46 / 80

Which Azure AI Vision capability would analyze video footage from a retail store to count the number of people present in a specific area at a given time?

Image classification alone

Facial recognition alone, which identifies specific individuals rather than simply counting people present in an area

Object detection alone, which identifies objects in a single static image rather than analyzing movement and presence over time within a defined physical space

Spatial analysis

Image tagging alone

Optical character recognition, which extracts text rather than analyzing people's movement and presence

47 / 80

Which Azure AI Vision capability would automatically generate a natural language description of an image's content, such as 'a dog running on the beach'?

Spatial analysis, which analyzes movement over time rather than generating a static image description

Facial recognition, which identifies specific people rather than generating a general scene description

Object detection alone, which identifies individual objects and bounding boxes rather than generating a fluent, natural language sentence describing the overall scene

Optical character recognition, which extracts text rather than generating a description of the image's visual content

Image description (captioning)

Image classification alone, which assigns a single category label rather than a fluent, descriptive sentence

48 / 80

Which Azure AI Vision feature would detect whether an uploaded image contains adult or violent content, supporting a content moderation use case?

Image classification alone, without a specific focus on flagging inappropriate content categories

Optical character recognition alone

Content moderation (via Azure AI Content Safety or Vision moderation capabilities)

Spatial analysis alone

Object detection alone, which identifies general objects rather than specifically flagging inappropriate content categories

Facial recognition alone

49 / 80

Which Azure AI Vision capability would be most appropriate for a manufacturing quality control scenario needing to detect specific defect types on a product, using a model trained on the organization's own labeled example images?

The pre-built, general-purpose image analysis (tagging) model, which is trained on broad, generic categories rather than an organization's specific custom defect types

Facial recognition alone

Optical character recognition alone, which extracts text rather than detecting visual product defects

The pre-built face detection model alone

Spatial analysis alone

A custom vision model (Azure AI Custom Vision, for custom image classification or object detection)

50 / 80

Which Azure AI Vision capability would extract structured data, such as a specific field value from a standardized form, rather than simply extracting all raw text?

Object detection alone

Facial recognition alone

Spatial analysis alone

Azure AI Document Intelligence (Form Recognizer) capabilities

Generic OCR alone, which extracts raw text without understanding the specific structured fields and their relationships within a form

Image classification alone

51 / 80

Which Azure AI Vision workload would be used to read and extract the total amount and vendor name from a photograph of a paper receipt?

Facial recognition alone

Image classification alone

Azure AI Document Intelligence (Form Recognizer) structured field extraction

Spatial analysis alone

Object detection alone

Generic image tagging alone, which provides broad descriptive labels rather than extracting specific structured field values

52 / 80

Which consideration is most important when deploying a facial recognition system for identity verification, given the varying accuracy of facial recognition across different demographic groups documented in research?

Assuming accuracy testing is only necessary once, with no ongoing monitoring after initial deployment

Deploying the system with no evaluation of its fairness across different groups

Ensuring the model has been evaluated for fairness across demographic groups and implementing appropriate human oversight before relying solely on the automated result

Assuming facial recognition accuracy is always identical across every demographic group with no variation whatsoever

Assuming Responsible AI fairness considerations are irrelevant to identity verification use cases specifically

Assuming human oversight is entirely unnecessary once a facial recognition system is deployed

53 / 80

Which Azure AI Language capability would determine whether a customer's product review expresses a positive, negative, neutral, or mixed opinion?

Translation, which converts text between languages rather than determining sentiment

Sentiment analysis

Key phrase extraction, which identifies main topics rather than determining the overall emotional tone or opinion

Language detection, which identifies the language rather than determining sentiment

Text summarization, which condenses text rather than determining sentiment

Named entity recognition, which identifies specific entities like names and places rather than determining sentiment

54 / 80

Which Azure AI Language capability would identify and extract the main topics or important concepts discussed within a long block of text, such as 'customer service' and 'billing issue'?

Named entity recognition, which identifies specific named entities like people and organizations rather than broader topical concepts

Key phrase extraction

Text summarization, which condenses the full text rather than specifically extracting key topical phrases

Language detection, which identifies the language rather than extracting topics

Translation

Sentiment analysis, which determines emotional tone rather than identifying the main topics or concepts discussed

55 / 80

Which Azure AI Language capability would identify and categorize specific real-world objects mentioned in text, such as people's names, organizations, dates, and locations?

Language detection

Named entity recognition (NER)

Key phrase extraction, which identifies general topical concepts rather than specific, categorized named entities

Sentiment analysis, which determines emotional tone rather than identifying and categorizing specific named entities

Text summarization

Translation

56 / 80

Which Azure AI Language capability automatically identifies which specific language a given piece of text is written in?

Sentiment analysis

Named entity recognition

Text summarization

Translation, which converts text between languages rather than simply identifying which language the original text is in

Key phrase extraction

Language detection

57 / 80

Which Azure AI Language capability would condense a lengthy news article down to a few key sentences, preserving the most important information?

Key phrase extraction, which extracts individual phrases rather than producing a coherent, condensed summary

Translation

Sentiment analysis, which determines emotional tone rather than condensing the overall length of a text

Language detection

Named entity recognition, which identifies specific entities rather than condensing overall text length

Text summarization

58 / 80

Which Azure AI Language capability would identify a specific, custom category, such as 'billing complaint' versus 'technical support request,' for incoming customer support tickets, based on a model trained on the organization's own labeled examples?

Named entity recognition alone

Key phrase extraction alone, which extracts general topics rather than assigning a specific, predefined custom category

Sentiment analysis alone, which determines emotional tone rather than a custom, organization-specific category

Generic pre-built key phrase extraction alone, without an organization-specific custom-trained classification model

Language detection alone

A custom text classification model (Azure AI Language custom classification)

59 / 80

Which Azure AI Speech capability converts spoken audio, such as a recorded phone call, into written text?

Speech translation, which converts spoken audio in one language into text or speech in a different language, rather than simply transcribing within the same language

Speech to text

Speaker recognition, which identifies who is speaking rather than transcribing the spoken words

Text to speech, which converts written text into spoken audio, the reverse operation

Voice assistants, which are a broader application built using multiple speech capabilities rather than the specific transcription capability itself

Custom neural voice, which creates a synthetic voice rather than transcribing audio

60 / 80

Which Azure AI Speech capability converts written text into natural-sounding spoken audio, such as for a voice assistant's response?

Speech translation, which converts spoken audio between languages rather than converting written text into audio

Speech to text, which converts spoken audio into written text, the reverse operation

Speaker recognition, which identifies a speaker's identity rather than generating spoken audio from text

Language detection, which identifies the language of a text or audio sample rather than generating audio

Text to speech

Sentiment analysis, which analyzes the emotional tone of text rather than converting it into audio

61 / 80

Which Azure AI Speech capability would translate a spoken conversation in real time from English into French, producing translated speech or text output?

Language detection alone, which identifies the language rather than performing the actual translation

Speaker recognition alone, which identifies who is speaking rather than translating the content

Text to speech alone, which generates audio from text rather than translating spoken input between languages

Speech translation

Speech to text alone, which transcribes audio within the same original language rather than translating into a different language

Custom neural voice alone, which creates a synthetic voice rather than performing translation

62 / 80

Which Azure AI Speech capability would verify a specific individual's identity based on the unique characteristics of their voice, such as for a voice-based banking authentication system?

Custom neural voice, which creates a synthetic voice rather than verifying an existing speaker's identity

Speech translation, which converts language rather than verifying identity

Text to speech, which generates audio rather than verifying an identity

Language detection, which identifies the spoken language rather than verifying the specific speaker's identity

Speaker recognition (speaker verification)

Speech to text, which transcribes spoken words rather than verifying the speaker's specific identity

63 / 80

Which term describes a type of AI model, often based on a transformer architecture, capable of generating new, original content such as text, images, or code in response to a prompt?

An anomaly detection model, which flags unusual existing data points rather than generating new original content

A clustering model, which groups existing data rather than generating new original content

A generative AI model

A regression model, which predicts a numeric value rather than generating new original content

A classification model, which assigns a category rather than generating new original content

An object detection model, which identifies existing objects in an image rather than generating new original content

64 / 80

Which term describes a large-scale generative AI model, trained on vast amounts of text data, capable of understanding and generating human-like language across a wide range of tasks?

A large language model (LLM)

An anomaly detection model

A simple rule-based chatbot, which follows predefined rules rather than learning language patterns from vast training data

A clustering model

A regression model, which is typically a much simpler, task-specific model rather than a large-scale, general-purpose language model

An object detection model

65 / 80

Which Azure service provides access to large language models, such as GPT-family models, allowing organizations to build custom generative AI applications?

Azure Machine Learning designer, which is a visual pipeline builder rather than a managed service providing access to pre-built large language models

Azure AI Vision, which is focused specifically on image analysis rather than providing broad access to large language models

Azure AI Document Intelligence, which is focused on structured document data extraction rather than providing broad access to large language models

Azure AI Search, which is focused on indexing and retrieval rather than providing the underlying generative language models themselves

Azure OpenAI Service

Azure AI Speech, which is focused specifically on audio processing rather than providing broad access to large language models

66 / 80

Which term describes the specific input text provided to a generative AI model to guide the type and nature of the content it generates?

A hyperparameter, which configures the training process rather than serving as the runtime instruction to a generative model

A model checkpoint, which is a saved state of a model during training rather than a runtime instruction

A label, which represents a known correct output in supervised learning rather than the guiding input to a generative model

A feature, which represents an input characteristic in a traditional predictive model rather than a natural language instruction to a generative model

A prompt

An epoch, which represents a training iteration rather than a runtime instruction

67 / 80

Which term describes the practice of carefully designing and refining a prompt to elicit a more accurate, relevant, or useful response from a generative AI model?

Hyperparameter tuning, which adjusts training configuration values rather than crafting the input instruction text

Feature engineering, which involves transforming input data for a traditional predictive model rather than crafting an instruction for a generative model

Data labeling, which is the process of assigning correct outputs to training examples rather than crafting an instruction to a generative model

Prompt engineering

Fine-tuning, which involves further training the underlying model itself on additional data, rather than simply crafting the input instruction text

Model training, which is the broader process of creating a model from scratch rather than crafting an instruction to an already-trained model

68 / 80

Which term describes a generative AI model's tendency to generate plausible-sounding but factually incorrect or fabricated information?

Prompt engineering, which is the practice of crafting effective input instructions rather than a description of a model's output error tendency

Hallucination

Fine-tuning, which is a training technique rather than a description of a model's output error tendency

Bias, which specifically refers to systematic unfairness toward certain groups rather than the general fabrication of plausible-sounding but incorrect information

Overfitting, which describes a traditional predictive model's poor generalization to new data rather than a generative model fabricating false information

Underfitting, which describes a traditional predictive model failing to capture patterns rather than a generative model fabricating false information

69 / 80

Which Azure AI Content Safety capability would automatically detect and flag harmful, offensive, or inappropriate content generated by a generative AI application before it is shown to a user?

Fine-tuning, which further trains a model on additional data rather than filtering generated output for harm

Prompt engineering, which is the practice of crafting effective input instructions rather than filtering the resulting output for harm

Model interpretability tools, which explain a traditional predictive model's specific numeric decision rather than filtering generative content for harm

AutoML, which automates traditional predictive model selection rather than filtering generative AI output

Hallucination detection alone, which specifically addresses factual inaccuracy rather than the broader category of harmful or offensive content

Content filtering / content safety classification

70 / 80

Which technique involves providing a generative AI model with relevant, up-to-date, or organization-specific information retrieved from an external data source at the time of the query, to improve the accuracy and relevance of its response?

Hyperparameter tuning, which adjusts training configuration values rather than retrieving external information at query time

Fine-tuning, which involves further training the underlying model's parameters on additional data, rather than retrieving external information at query time

Model interpretability, which explains an existing prediction rather than retrieving external information to improve a generative response

Retrieval augmented generation (RAG)

AutoML, which automates traditional predictive model selection rather than retrieving external information for a generative model

Prompt engineering alone, which focuses on crafting the instruction wording itself rather than specifically retrieving external data

71 / 80

Which Azure OpenAI Service capability allows a developer to further adjust a pre-trained large language model's behavior using a smaller, organization-specific dataset, tailoring it to a specific style or domain?

Content filtering, which screens output for harm rather than adjusting the model's underlying behavior through additional training

Prompt engineering alone, which crafts the input instruction at query time rather than further training the underlying model's parameters

Retrieval augmented generation (RAG) alone, which retrieves external context at query time rather than adjusting the model's underlying parameters

Model interpretability, which explains predictions rather than adjusting a model's behavior through additional training

Fine-tuning

AutoML, which is used for traditional predictive machine learning tasks rather than adjusting a large language model

72 / 80

Which Microsoft product embeds generative AI capabilities directly within familiar productivity applications, such as drafting a document in Word or summarizing an email thread in Outlook?

Azure Machine Learning designer, which is a data science tool for building traditional predictive models rather than an embedded productivity assistant

Azure AI Content Safety, which filters harmful content rather than embedding generative assistance within productivity applications

Azure AI Vision, which processes images rather than embedding generative assistance within productivity applications

Azure AI Document Intelligence, which extracts structured data from documents rather than embedding a broad generative productivity assistant

Microsoft Copilot

Azure AI Search, which indexes and retrieves content rather than embedding generative assistance within productivity applications

73 / 80

Which term describes a specific type of generative AI model that generates new, original images based on a text description provided as a prompt?

An object detection model, which identifies existing objects within an already-existing image rather than generating a brand-new image from text

A clustering model, which groups existing data points rather than generating new images

Optical character recognition, which extracts existing text from an image rather than generating a new image

An image classification model, which categorizes an already-existing image rather than generating a new one

A text-to-image generation model (such as DALL-E)

A regression model, which predicts a numeric value rather than generating a new image

74 / 80

Which consideration is most important when an organization deploys a customer-facing generative AI chatbot, given the risk that the model might generate an inappropriate, biased, or factually incorrect response?

Assuming Responsible AI principles apply only to traditional predictive models, not generative AI chatbots

Deploying the chatbot with no content filtering, monitoring, or safeguards of any kind

Assuming a generative AI chatbot will never produce an inappropriate, biased, or factually incorrect response under any circumstance

Assuming monitoring is only necessary once, with no ongoing oversight after initial deployment

Implementing content filtering, monitoring, and appropriate Responsible AI safeguards, along with clear disclosure that the user is interacting with an AI system

Assuming disclosure that users are interacting with an AI system is unnecessary

75 / 80

Which term describes the broader field of AI research and models focused on understanding and generating images, video, and other visual content using deep learning techniques, including generative approaches?

Speech processing, which is focused on audio rather than visual content

Natural language processing, which is focused on text and language rather than visual content

Computer vision (with generative AI extending its capabilities to include image generation)

Knowledge mining, which is focused on extracting insights from unstructured documents rather than the broader visual content field

Reinforcement learning, which is a training methodology rather than a content domain like vision

Anomaly detection, which is a specific analytical task rather than the broader visual content field

76 / 80

Which Azure OpenAI Service feature allows a developer to programmatically integrate a large language model's capabilities into their own custom application via API calls?

Microsoft Copilot alone, which is a pre-built end-user productivity application rather than a developer-facing API for building custom applications

Azure AI Content Safety alone, which filters content rather than providing programmatic access to a large language model's core generative capabilities

Azure AI Search alone, which indexes content rather than providing programmatic access to a large language model

Azure Machine Learning designer alone, which is a visual pipeline builder for traditional predictive models rather than an API for large language models

The Azure OpenAI Service REST API / SDK

Azure AI Vision alone, which processes images rather than providing programmatic access to a large language model

77 / 80

Which term describes providing a generative AI model with one or more example input-output pairs directly within the prompt, to guide it toward producing a response in a similar style or format?

Zero-shot prompting, which provides no examples at all, relying solely on the model's general pre-trained knowledge

Hyperparameter tuning, which adjusts training configuration values rather than crafting prompt content

Fine-tuning, which involves further training the model's underlying parameters rather than simply including examples within the prompt text itself

Retrieval augmented generation (RAG), which retrieves external documents rather than providing illustrative examples directly within the prompt

Few-shot prompting

Model interpretability, which explains predictions rather than guiding a generative model's output format

78 / 80

Which term describes providing a generative AI model with no specific examples, relying solely on its general pre-trained knowledge and the instruction itself to complete a task?

Zero-shot prompting

Retrieval augmented generation (RAG), which retrieves external documents to supplement the prompt rather than relying solely on the model's pre-trained knowledge

Hyperparameter tuning

Fine-tuning, which involves further training the model's underlying parameters rather than a specific prompting technique used at query time

Few-shot prompting, which specifically includes one or more illustrative examples directly within the prompt

Model interpretability

79 / 80

Which risk specifically associated with generative AI systems involves the model inadvertently reproducing copyrighted or proprietary content it was exposed to during training?

Underfitting, which is a traditional predictive model's specific failure-to-learn behavior rather than the reproduction of copyrighted training content

Clustering, which is a specific analytical task rather than a risk category

Overfitting, which is a traditional predictive model's specific poor-generalization behavior rather than the reproduction of copyrighted training content

Regression, which is a specific predictive task rather than a risk category

Intellectual property / copyright risk

Anomaly detection, which is a specific analytical task rather than a risk category

80 / 80

Which term describes a generative AI system's tendency to produce different outputs for functionally similar prompts, requiring careful evaluation to ensure the range of outputs remains appropriate and consistent with organizational standards?

Overfitting, which describes a traditional predictive model's specific training-versus-new-data performance gap rather than a generative model's inherent output variability

Underfitting, which describes a traditional predictive model's failure to learn patterns rather than a generative model's inherent output variability

Hyperparameter tuning, which is a training configuration technique rather than a description of an inherent output characteristic

Feature engineering, which is a traditional predictive modeling technique rather than a description of a generative model's inherent output characteristic

Output variability (non-determinism)

AutoML, which automates traditional predictive model selection rather than describing a generative model's inherent output characteristic

Your score is

The average score is 83%