Generative AI in Education Practice Test/Career Essentials in
Generative AI by Microsoft and LinkedIn Updated 2025/2026
New!! - 250 Questions
This exam covers the foundational concepts of Generative AI, including model architectures, training paradigms,
ethical considerations, and applications in educational contexts. It assesses deep understanding of generative
models, their capabilities, limitations, and implications for teaching and learning. It contains 250 multiple-choice
questions, each with four distractors and a fully worked rationale that explains why the keyed answer is correct.
Content is organized into 10 focused sections: Introduction to Generative AI, Large Language Models and
Transformers, Prompt Engineering and Fine-Tuning, Ethics and Responsible AI, Generative AI in Education: Use
Cases, Assessment and Feedback with AI, AI-Powered Content Creation, Data Privacy and Security in AI,
Implementing AI in Learning Environments, Future Trends in Generative AI. Targeted learning outcomes include:
Analyze the core principles of generative adversarial networks and variational autoencoders.; Evaluate the
ethical and societal implications of generative AI in education.; Apply generative AI techniques to create
educational content and assess their effectiveness.. Every item has been reviewed for clinical accuracy, current
guidelines, and clarity so that students can study with confidence and self-correct as they work through the bank.
Use it as a high-yield review immediately before the exam, or as a structured practice tool during the unit - the
rationales double as concise teaching notes. The recommended writing time is 3 hours, with a passing score of
90%. Aligned with This exam meets the standards of top US research universities, including Harvard, MIT, and
Section 1: Introduction to Generative AI (Questions 1-25)
1 In a generative adversarial network (GAN), the generator's objective is to
minimize log(1-D(G(z))). However, early in training, this loss saturates,
providing weak gradients. A common fix is to instead train the generator to
maximize log(D(G(z))). This modification is known as which of the
following?
A) Wasserstein loss with gradient penalty
B) Non-saturating heuristic loss
C) Least squares loss
D) Hinge loss
Answer: B
Rationale: The non-saturating heuristic loss (maximizing log(D(G(z))))
provides stronger gradients early in training, avoiding saturation. Wasserstein
loss uses a different critic and gradient penalty. Least squares loss and hinge
loss are alternative GAN objectives but not the specific modification described.
2 A variational autoencoder (VAE) is trained on a dataset of handwritten digits.
After training, you generate new samples by drawing z from a standard
Gaussian prior and decoding. However, the generated digits are blurry.
Which architectural modification is most likely to address this issue?
,A) Increase the latent dimension size
B) Use a more expressive decoder, such as a PixelCNN
C) Decrease the weight of the KL divergence term in the loss
D) Replace the Gaussian prior with a mixture of Gaussians
Answer: B
Rationale: Blurry outputs in VAEs often arise because the decoder averages
over plausible outputs. A more expressive decoder like PixelCNN can model
sharp distributions. Increasing latent dimensions may help but often leads to
posterior collapse. Adjusting KL weight or prior may affect latent structure, but
the primary cause of blurriness is the decoder's limited capacity.
3 In the context of autoregressive generative models, such as GPT, the
probability of a sequence is factorized as the product of conditional
probabilities over tokens. Which of the following best describes the primary
limitation of this approach for text generation?
A) It cannot model long-range dependencies due to the Markov assumption
B) It is computationally expensive to compute the exact likelihood of a
sequence
C) It suffers from exposure bias during training because the model is only
exposed to ground-truth tokens
D) It requires a fixed-length context window, limiting the model's ability to
handle arbitrary-length sequences
Answer: C
Rationale: Exposure bias occurs because during training the model conditions
on ground-truth tokens, but at inference it conditions on its own previous
predictions, leading to error accumulation. While transformers can model
long-range dependencies, the Markov assumption is not inherent. Likelihood
computation is efficient. Fixed-length context is a limitation but not the
primary one described.
4 A researcher wants to generate realistic educational assessment questions
using a diffusion model. Which of the following is a fundamental advantage
of diffusion models over GANs for this task?
A) Diffusion models are inherently easier to train due to a simpler loss
function
,B) Diffusion models provide a principled way to compute the exact
likelihood of generated samples
C) Diffusion models avoid mode collapse by directly modeling the data
distribution through iterative denoising
D) Diffusion models generate samples faster than GANs because they require
fewer forward passes
Answer: C
Rationale: Diffusion models mitigate mode collapse by learning a denoising
process that covers the entire data distribution. Their loss is not simpler;
training involves multiple steps. Exact likelihood is tractable only in
continuous-time formulations. GANs typically generate samples in one
forward pass, while diffusion models require many steps, making them slower.
5 When fine-tuning a large language model (LLM) for educational content
generation, you observe that the model frequently produces factually
incorrect information. Which of the following interventions is most directly
targeted at reducing factual inaccuracies?
A) Increase the temperature parameter during sampling
B) Apply reinforcement learning from human feedback (RLHF) with a
reward for factual correctness
C) Use top-k sampling with a small k value
D) Increase the model's context window size
Answer: B
Rationale: RLHF directly optimizes the model to align with human preferences,
including factual accuracy. Temperature and top-k sampling affect diversity
but not correctness. Increasing context window helps with longer dependencies
but does not inherently improve factual accuracy.
6 A generative model for producing personalized learning pathways is trained
using a sequence-to-sequence architecture with attention. The model's
performance degrades when the input sequence (student profile) is very long.
Which modification is most likely to help?
A) Use a convolutional neural network instead of attention
B) Add positional encoding and use a transformer with self-attention
C) Reduce the number of layers in the encoder
D) Apply gradient clipping during training
, Answer: B
Rationale: Transformers with self-attention and positional encoding are
designed to handle long sequences effectively. CNNs may not capture
long-range dependencies well. Reducing layers could worsen performance.
Gradient clipping addresses exploding gradients, not sequence length issues.
7 In the context of generative models for education, which of the following
best describes the concept of 'alignment'?
A) Ensuring the model's latent representations are aligned with human brain
activity
B) Adjusting the model's output distribution to match a target distribution of
correct educational content
C) Training the model to produce outputs that are consistent with human
values and instructional goals
D) Using adversarial training to align the generator and discriminator in a
GAN
Answer: C
Rationale: Alignment refers to the process of ensuring AI systems act in
accordance with human intentions and values. In education, this means
generating content that is pedagogically sound, unbiased, and safe. Option B
describes distribution matching, not alignment. Option A is a neuroscience
concept. Option D is specific to GANs.
8 A research team develops a generative model to create synthetic student
essays for training grading algorithms. Which of the following ethical
concerns is most directly amplified by using such synthetic data?
A) The model may inadvertently memorize and reproduce sensitive
information from the training data
B) The synthetic essays may lack the diversity of real student writing, leading
to biased grading
C) The model's training process consumes excessive energy, contributing to
climate change
D) The synthetic data may be used to cheat on assignments by generating
essays indistinguishable from human-written ones
Answer: B
Rationale: Synthetic data may not capture the full variability of real student
Generative AI by Microsoft and LinkedIn Updated 2025/2026
New!! - 250 Questions
This exam covers the foundational concepts of Generative AI, including model architectures, training paradigms,
ethical considerations, and applications in educational contexts. It assesses deep understanding of generative
models, their capabilities, limitations, and implications for teaching and learning. It contains 250 multiple-choice
questions, each with four distractors and a fully worked rationale that explains why the keyed answer is correct.
Content is organized into 10 focused sections: Introduction to Generative AI, Large Language Models and
Transformers, Prompt Engineering and Fine-Tuning, Ethics and Responsible AI, Generative AI in Education: Use
Cases, Assessment and Feedback with AI, AI-Powered Content Creation, Data Privacy and Security in AI,
Implementing AI in Learning Environments, Future Trends in Generative AI. Targeted learning outcomes include:
Analyze the core principles of generative adversarial networks and variational autoencoders.; Evaluate the
ethical and societal implications of generative AI in education.; Apply generative AI techniques to create
educational content and assess their effectiveness.. Every item has been reviewed for clinical accuracy, current
guidelines, and clarity so that students can study with confidence and self-correct as they work through the bank.
Use it as a high-yield review immediately before the exam, or as a structured practice tool during the unit - the
rationales double as concise teaching notes. The recommended writing time is 3 hours, with a passing score of
90%. Aligned with This exam meets the standards of top US research universities, including Harvard, MIT, and
Section 1: Introduction to Generative AI (Questions 1-25)
1 In a generative adversarial network (GAN), the generator's objective is to
minimize log(1-D(G(z))). However, early in training, this loss saturates,
providing weak gradients. A common fix is to instead train the generator to
maximize log(D(G(z))). This modification is known as which of the
following?
A) Wasserstein loss with gradient penalty
B) Non-saturating heuristic loss
C) Least squares loss
D) Hinge loss
Answer: B
Rationale: The non-saturating heuristic loss (maximizing log(D(G(z))))
provides stronger gradients early in training, avoiding saturation. Wasserstein
loss uses a different critic and gradient penalty. Least squares loss and hinge
loss are alternative GAN objectives but not the specific modification described.
2 A variational autoencoder (VAE) is trained on a dataset of handwritten digits.
After training, you generate new samples by drawing z from a standard
Gaussian prior and decoding. However, the generated digits are blurry.
Which architectural modification is most likely to address this issue?
,A) Increase the latent dimension size
B) Use a more expressive decoder, such as a PixelCNN
C) Decrease the weight of the KL divergence term in the loss
D) Replace the Gaussian prior with a mixture of Gaussians
Answer: B
Rationale: Blurry outputs in VAEs often arise because the decoder averages
over plausible outputs. A more expressive decoder like PixelCNN can model
sharp distributions. Increasing latent dimensions may help but often leads to
posterior collapse. Adjusting KL weight or prior may affect latent structure, but
the primary cause of blurriness is the decoder's limited capacity.
3 In the context of autoregressive generative models, such as GPT, the
probability of a sequence is factorized as the product of conditional
probabilities over tokens. Which of the following best describes the primary
limitation of this approach for text generation?
A) It cannot model long-range dependencies due to the Markov assumption
B) It is computationally expensive to compute the exact likelihood of a
sequence
C) It suffers from exposure bias during training because the model is only
exposed to ground-truth tokens
D) It requires a fixed-length context window, limiting the model's ability to
handle arbitrary-length sequences
Answer: C
Rationale: Exposure bias occurs because during training the model conditions
on ground-truth tokens, but at inference it conditions on its own previous
predictions, leading to error accumulation. While transformers can model
long-range dependencies, the Markov assumption is not inherent. Likelihood
computation is efficient. Fixed-length context is a limitation but not the
primary one described.
4 A researcher wants to generate realistic educational assessment questions
using a diffusion model. Which of the following is a fundamental advantage
of diffusion models over GANs for this task?
A) Diffusion models are inherently easier to train due to a simpler loss
function
,B) Diffusion models provide a principled way to compute the exact
likelihood of generated samples
C) Diffusion models avoid mode collapse by directly modeling the data
distribution through iterative denoising
D) Diffusion models generate samples faster than GANs because they require
fewer forward passes
Answer: C
Rationale: Diffusion models mitigate mode collapse by learning a denoising
process that covers the entire data distribution. Their loss is not simpler;
training involves multiple steps. Exact likelihood is tractable only in
continuous-time formulations. GANs typically generate samples in one
forward pass, while diffusion models require many steps, making them slower.
5 When fine-tuning a large language model (LLM) for educational content
generation, you observe that the model frequently produces factually
incorrect information. Which of the following interventions is most directly
targeted at reducing factual inaccuracies?
A) Increase the temperature parameter during sampling
B) Apply reinforcement learning from human feedback (RLHF) with a
reward for factual correctness
C) Use top-k sampling with a small k value
D) Increase the model's context window size
Answer: B
Rationale: RLHF directly optimizes the model to align with human preferences,
including factual accuracy. Temperature and top-k sampling affect diversity
but not correctness. Increasing context window helps with longer dependencies
but does not inherently improve factual accuracy.
6 A generative model for producing personalized learning pathways is trained
using a sequence-to-sequence architecture with attention. The model's
performance degrades when the input sequence (student profile) is very long.
Which modification is most likely to help?
A) Use a convolutional neural network instead of attention
B) Add positional encoding and use a transformer with self-attention
C) Reduce the number of layers in the encoder
D) Apply gradient clipping during training
, Answer: B
Rationale: Transformers with self-attention and positional encoding are
designed to handle long sequences effectively. CNNs may not capture
long-range dependencies well. Reducing layers could worsen performance.
Gradient clipping addresses exploding gradients, not sequence length issues.
7 In the context of generative models for education, which of the following
best describes the concept of 'alignment'?
A) Ensuring the model's latent representations are aligned with human brain
activity
B) Adjusting the model's output distribution to match a target distribution of
correct educational content
C) Training the model to produce outputs that are consistent with human
values and instructional goals
D) Using adversarial training to align the generator and discriminator in a
GAN
Answer: C
Rationale: Alignment refers to the process of ensuring AI systems act in
accordance with human intentions and values. In education, this means
generating content that is pedagogically sound, unbiased, and safe. Option B
describes distribution matching, not alignment. Option A is a neuroscience
concept. Option D is specific to GANs.
8 A research team develops a generative model to create synthetic student
essays for training grading algorithms. Which of the following ethical
concerns is most directly amplified by using such synthetic data?
A) The model may inadvertently memorize and reproduce sensitive
information from the training data
B) The synthetic essays may lack the diversity of real student writing, leading
to biased grading
C) The model's training process consumes excessive energy, contributing to
climate change
D) The synthetic data may be used to cheat on assignments by generating
essays indistinguishable from human-written ones
Answer: B
Rationale: Synthetic data may not capture the full variability of real student
