What Will You Learn?
By the end of this lesson, you will be able to:
- Explain how Generative AI creates new content using specific model architectures
- Distinguish between Discriminative Modeling and Generative Modeling
- Describe how GANs, VAEs, RNNs, and Autoencoders work
- Connect each model type to a real-world generative AI application
- Trace the timeline of Generative AI’s development from 2011 to 2023
- Identify the ethical risks specific to generative AI, including deep fakes and ownership
In the previous lesson, you met Generative AI through its most famous faces — ChatGPT, DALL-E, Suno. You learned what it can do and how to use it responsibly.
This blog goes one level deeper. Behind those polished tools are specific model architectures — technical designs that make generation possible. Knowing how generative AI produces its outputs separates a casual user from someone who truly understands it.
Let’s open the hood.
Generative Modeling vs. Discriminative Modeling
To understand how generative AI works, you first need to understand the two fundamental approaches to AI modeling.
Discriminative Modeling
A discriminative model learns to distinguish between classes. It is given labeled data (supervised learning) and its job is to draw a decision boundary.
Input: [Image of a cat]
↓
Discriminative Model
↓
Output: "This is a CAT" (not a dog)
The model asks: “Given this input, what is the most likely label?” It does not need to understand what a cat looks like in any deep sense — it just needs to know what separates cats from non-cats.
Common examples of discriminative models include spam classifiers, disease detectors, and image classifiers.
Generative Modeling
A generative model learns to understand the underlying data distribution — the patterns and structure of what makes something look or sound like what it is. It learns from unlabeled or loosely labeled data (unsupervised or semi-supervised learning).
Learning phase: Studies thousands of cat images
↓
Generative Model learns:
"Cats have pointed ears, whiskers, fur patterns..."
↓
Output: Generates a NEW cat image that never existed
The model asks: “What does data from this category look like? Can I create new examples?”
| Discriminative Model | Generative Model | |
|---|---|---|
| Goal | Classify what already exists | Create new examples |
| Question | “What class does this belong to?” | “What does data from this class look like?” |
| Learning type | Supervised | Unsupervised / Semi-supervised |
| Output | Labels, classifications, predictions | New images, text, audio, data |
| Examples | Spam filter, disease detector | DALL-E, ChatGPT, AIVA |
💡 Key Insight
Discriminative models are the workhorses of traditional AI — excellent at categorising what exists. Generative models are the creative engines — they understand structure deeply enough to create something new. The shift from discriminative to generative thinking represents one of the most important transitions in modern AI.
A Brief Timeline: How Generative AI Evolved (2011–2023)
Generative AI did not appear overnight. It is the result of a decade of research, breakthroughs, and growing computational power.
| Year | Milestone |
|---|---|
| 2011 | Early neural language models begin generating simple text. IBM Watson wins Jeopardy, sparking interest in AI language capabilities. |
| 2013 | Variational Autoencoders (VAEs) introduced by Kingma and Welling — a mathematical breakthrough for generating structured data. |
| 2014 | Generative Adversarial Networks (GANs) invented by Ian Goodfellow — the training-through-competition idea that revolutionised image generation. |
| 2015 | Deep learning image generation improves dramatically. AI begins generating faces that are difficult to distinguish from real photographs. |
| 2016 | “The Next Rembrandt” — a GAN analyses Rembrandt’s entire body of work and produces a new painting in his style. AI art is born. |
| 2017 | Transformer architecture introduced (Google’s “Attention is All You Need” paper) — the technical foundation for all modern large language models. |
| 2018 | GPT-1 released by OpenAI — first demonstration that pre-training on massive text can produce powerful language generation. |
| 2020 | GPT-3 released — with 175 billion parameters, produces human-like writing. DALL-E prototype generates images from text. |
| 2021 | DALL-E (version 1) publicly demonstrated. AI image generation enters mainstream awareness. |
| 2022 | Stable Diffusion and Midjourney launch. ChatGPT releases in November 2022 — reaches 100 million users in 2 months. |
| 2023 | GPT-4, Gemini, Sora (video generation), Microsoft Copilot — generative AI enters everyday productivity tools. |
This timeline shows that the technical foundations (VAEs in 2013, GANs in 2014, Transformers in 2017) came well before the consumer tools most people know. Building a generative AI product takes years of underlying research.
The Four Core Generative AI Model Types
The CBSE Class 9 AI curriculum identifies four specific model architectures that are the technical building blocks of generative AI. Each has a distinct design, strength, and application.
1. GANs — Generative Adversarial Networks
Introduced: 2014 by Ian Goodfellow
The Big Idea: Two neural networks compete against each other. This competition is what drives the quality of generated content to improve.
The two networks:
┌─────────────────────┐
Random Noise ──────▶│ GENERATOR (G) │──────▶ Fake Image
│ Tries to create │
│ convincing fakes │
└─────────────────────┘
│
▼ Fake image
┌─────────────────────┐
Real Images ───────▶│ DISCRIMINATOR (D) │──────▶ Real or Fake?
│ Tries to catch │
│ the fakes │
└─────────────────────┘
│
Feedback signal
│
Goes back to Generator
to help it improve
How they work together:
The Generator starts by producing random, low-quality outputs. The Discriminator easily spots them as fake.
As training continues, the Generator gets feedback and improves. The Discriminator now has to work harder. This back-and-forth — like a forger versus an art detective — drives quality steadily upward. Eventually, the Generator produces outputs the Discriminator can no longer tell from real ones.
In the formula:
– Generator (G) takes random noise → creates fake data
– Discriminator (D) evaluates data → outputs probability of being real
– Both networks improve through competition
Real-World Example — “The Next Rembrandt” (2016):
Researchers trained a GAN on all 346 known Rembrandt paintings. The GAN analysed his use of light, shadow, facial proportions, and brush textures. After learning these patterns, it generated an entirely new portrait — one Rembrandt never actually painted. The result was so convincing it was 3D-printed to replicate real oil paint texture.
A GAN does not merely imitate. It internalises the generative grammar of a style and creates something genuinely original.
GAN Applications:
| Application | How GANs Help |
|---|---|
| Image generation | DALL-E (early version), Midjourney |
| Face synthesis | Generating photorealistic faces of people who don’t exist |
| Data augmentation | Creating synthetic training data |
| Art creation | “The Next Rembrandt,” Artbreeder |
| Video deepfakes | Generating realistic video of people (raises ethical concerns) |
Hands-On Tool: GAN Paint
GAN Paint is a browser-based tool where you can draw on a photo of a building or scene and the GAN fills in realistic textures — adding doors, trees, or windows as if they were always there. Try it to see a GAN working in real time.
Hands-On Tool: Artbreeder
Artbreeder lets you blend images together using GAN-powered sliders. You can combine two faces, two landscapes, or two paintings to generate something new. Each child image inherits visual “genes” from its parents — a direct demonstration of how the GAN has learned underlying visual features.
2. VAEs — Variational Autoencoders
Introduced: 2013 by Kingma and Welling
The Big Idea: Compress data into a mathematical description, then reconstruct or generate new data from that compressed form.
How it works:
Input Image Encoder Latent Space Decoder Output
(Real cat) ──────────▶ [μ, σ: compressed ──────────▶ (Reconstructed
mathematical or generated cat)
description of
"what a cat looks like")
The Encoder takes an input — say, a cat image — and compresses it into a small set of numbers. These numbers capture essential features: ear shape, fur texture, colour distribution. They do not store pixels; they store meaning. This compact representation is called the latent space.
The Decoder takes numbers from the latent space and reconstructs an image from them.
Why the “Variational” part matters:
A standard Autoencoder (see below) gives one fixed point in latent space for each input. A VAE gives a range — a probability distribution. This means you can sample different points from that range and generate variations. It is the variation that enables genuine generation, not just reconstruction.
Standard Autoencoder: Cat → [3.2, 1.1, 7.8] → Same cat
VAE: Cat → [2.9–3.5, 0.8–1.4, 7.5–8.1] → Different but valid cat
Applications of VAEs:
| Application | Example |
|---|---|
| Image generation | Generating new faces, artwork, or molecules |
| Drug discovery | Creating new molecular structures with desired properties |
| Anomaly detection | Learning what “normal” looks like, then flagging deviations |
| Data compression | Compressing images while preserving key features |
VAEs are especially valuable in scientific fields where generating new, valid examples is critical — such as designing new drug candidates or materials.
3. RNNs — Recurrent Neural Networks
The Big Idea: Unlike standard neural networks that process inputs independently, RNNs have memory. They process sequences, keeping track of what came before to inform what comes next.
Why memory matters for generation:
Language, music, and time-series data all have the property that context matters. The word “bank” means something different after “river” versus after “money.” The next note in a melody depends on everything that came before it. RNNs are designed to handle this sequential dependency.
Standard NN: [word 1] → Output [word 2] → Output (No memory)
RNN: [word 1] → hidden state → [word 2] → hidden state → [word 3]
↑ ↑
carries memory carries memory
The hidden state is the memory. It is a vector of numbers that summarises all prior input. At each step, the current input combines with the hidden state. Together, they produce the next output.
Applications of RNNs:
| Application | How RNNs Help |
|---|---|
| Music generation | Learning melody patterns and rhythm structures |
| Text generation | Predicting the next word based on all previous words |
| Speech recognition | Processing audio frames in sequence |
| Time-series forecasting | Predicting stock prices, weather over time |
| Language translation | Processing source sentence sequentially |
Real-World Example — AIVA (Artificial Intelligence Virtual Artist):
AIVA is an AI composer that uses RNNs to generate original music. It was trained on thousands of classical scores — Bach, Beethoven, Mozart. The RNN learns the long-range structure of musical composition. It understands how a theme introduced in bar 4 may return transformed in bar 32.
The result is original music with a clear beginning, development, and resolution. The model’s memory maintains coherence across the entire piece — not just from note to note.
In 2017, AIVA became the first AI officially recognised as a composer. The recognition came from SACEM, a French music rights organisation.
💡 Note on modern evolution:
Modern large language models like GPT-4 use the Transformer architecture (introduced 2017) rather than pure RNNs for processing very long sequences. However, the core principle of sequential processing and memory that RNNs introduced remains foundational to understanding how any language model handles context.
4. Autoencoders
The Big Idea: Learn a compressed representation of data by training a network to reconstruct its own input.
An Autoencoder has two parts:
Input Encoder Bottleneck Decoder Output
[Original] ──────────▶ [Compressed code] ──────────▶ [Reconstructed]
The Encoder progressively compresses the input. The Bottleneck is the narrowest layer — the most compact representation possible. The Decoder tries to reconstruct the original from this compressed code.
The trick; By forcing compression and then reconstruction, the network learns only the most essential features. If the compressed code can rebuild a face, it must contain that face’s core structural information.
How Autoencoders differ from VAEs:
| Autoencoder | VAE | |
|---|---|---|
| Latent space | Fixed points | Probability distributions |
| Generation ability | Limited (reconstruction only) | Strong (can sample to generate new data) |
| Use | Compression, denoising, anomaly detection | Generation and variation |
Applications of Autoencoders:
| Application | How It Works |
|---|---|
| Image compression | Encode to small code, decode when needed |
| Denoising | Train to reconstruct clean images from noisy ones |
| Anomaly detection | Reconstructs “normal” data well; fails on anomalies — the reconstruction error flags the problem |
| Feature learning | The bottleneck layer captures essential features usable by other models |
| Dimensionality reduction | Reduces data to fewer dimensions while preserving structure |
Medical example: Autoencoders trained on normal ECG data reconstruct normal heart patterns with very low error. When an irregular heartbeat is fed in, reconstruction error spikes sharply. This spike flags a potential anomaly for the doctor to review.
Connecting the Four Models
| Model | Core Mechanism | Best For | Famous Example |
|---|---|---|---|
| GAN | Generator vs. Discriminator competition | Image generation, photorealism | The Next Rembrandt, Artbreeder |
| VAE | Encode to latent space, decode to generate | Controlled generation, scientific data | Drug discovery, face generation |
| RNN | Sequential memory for processing order | Music, language, time-series | AIVA (music), early ChatGPT |
| Autoencoder | Compress to reconstruct, learn features | Compression, denoising, anomaly detection | Medical anomaly detection, image denoising |
Hands-On GenAI Tools to Explore
These tools give you direct experience with the generative models above:
| Tool | Model Behind It | What You Can Do |
|---|---|---|
| GAN Paint | GAN | Paint on a scene and watch the GAN fill in realistic content |
| Artbreeder | GAN | Blend images to create new portraits, landscapes, anime characters |
| Runway ML | Multiple models (GAN, Diffusion) | Generate and edit videos and images with AI |
| ChatGPT | Transformer (evolved from RNN principles) | Conversational text generation |
| Gemini | Transformer | Multimodal generation (text, image understanding) |
| Microsoft Copilot | Transformer | Integrated AI writing and coding assistance |
Deep Fakes: When Generative AI Creates Harm
The same GAN technology that created “The Next Rembrandt” can also create deep fakes. These are AI-generated videos that realistically replace a real person’s face, voice, or body. The output looks authentic. But it never happened.
How deep fakes work:
Input: 100 real videos of Person A speaking
↓
GAN trains on facial movements, expressions, voice patterns
↓
Encoder maps Person A's face
↓
Decoder generates new video: Person A appears to say/do
something they never said or did
What makes deep fakes dangerous:
| Risk | Example |
|---|---|
| Misinformation | A politician appears to declare war — but never said it |
| Fraud | A CEO’s deep fake calls finance team to authorise a transfer |
| Reputation damage | False compromising videos of private individuals |
| Erosion of trust | When any video can be faked, all videos become suspect |
| Electoral manipulation | Fake campaign speeches or candidate scandals fabricated at scale |
How to spot deep fakes (current detection signals):
- Unnatural blinking or eye movement
- Inconsistent lighting on the face vs. background
- Blurring around the hairline or ears
- Unnatural skin texture or colour
- Audio that doesn’t quite sync with lip movements
- Subtle distortions when the subject turns their head
⚠️ Important Note: Deep fake detection technology is improving rapidly, but so are deep fakes. The best defence today is critical thinking — questioning the source of a video, especially when the content is surprising or omotionally provocative.
Ethical Considerations Specific to Generative AI
Generative AI introduces ethical challenges that go beyond the general AI ethics framework. Five specific concerns apply uniquely to content creation:
1. Ownership of AI-Generated Content
If a GAN trained on Rembrandt’s work creates a new painting, who owns it? The programmer who built the GAN? The organisation that owns Rembrandt’s paintings? The user who typed the prompt? Or no one?
Current copyright law in most countries, including India, does not recognise AI as an author. Whether the human who wrote the prompt owns the output is still debated in courts globally. As a creator in the AI age, understanding ownership matters. It protects your work — and helps you avoid infringing on others’.
2. Human Agency
As generative AI takes over writing, designing, composing, and coding, humans risk outsourcing creativity itself. The concern goes beyond jobs. It is about the development of human skill, judgment, and originality. Using AI as a tool is very different from letting it replace your thinking.
3. Bias in Generated Content
Generative models learn from human-created data — and human data reflects human biases. An image generator trained mainly on Western data may produce “scientist” images that are overwhelmingly male and light-skinned. A text generator may reflect language biases about gender roles.
The scale of generative AI makes bias more dangerous. One biased human produces limited output. One biased generative AI model produces millions of outputs.
4. Misinformation at Scale
Generative AI can produce false but convincing text, images, audio, and video. It generates this content faster than fact-checkers can review it. The same tool that helps a student write a summary can help a bad actor fabricate a news article or a crisis event.
5. Privacy
Generative models trained on personal data — photographs, voice recordings, private correspondence — may reproduce real individuals without their consent. Generating a photorealistic image of a real person in a false or compromising situation is a privacy violation. In many jurisdictions, it is also a legal offense.
| Ethical Issue | Core Question | Who Is Affected |
|---|---|---|
| Ownership | Who owns AI-generated content? | Creators, artists, legal systems |
| Human Agency | Are humans losing the skill to create? | Students, professionals, society |
| Bias | Is the AI perpetuating discrimination? | Underrepresented groups |
| Misinformation | Can AI-generated falsehoods be controlled? | Individuals, democracies, institutions |
| Privacy | Are real people’s likenesses used without consent? | Any individual whose data was used in training |
Activity: Identify the Model
For each of the following applications, identify which generative AI model type (GAN, VAE, RNN, Autoencoder) is most likely powering it. Explain your choice.
- An app that composes a personalised lullaby, verse by verse, that stays musically coherent from start to finish.
- A tool that detects fraudulent transactions by comparing them to what “normal” transactions look like.
- A research system that generates new molecular compounds with properties suitable for treating a specific disease.
- An image editing tool that lets you replace the sky in a photograph with a photorealistic sunset — one that never existed.
(Answers in Answer Key)
Quick Recap
- Discriminative models classify what exists; Generative models learn to create new examples.
- Generative AI evolved rapidly from 2011 to 2023, with VAEs (2013), GANs (2014), and Transformers (2017) as the key technical milestones.
- GANs use two competing networks — a Generator and a Discriminator — to produce high-quality images and realistic content.
- VAEs encode data into a probability distribution in latent space, enabling controlled generation of new examples.
- RNNs process sequences with memory, making them powerful for music and language generation.
- Autoencoders learn compressed representations of data, useful for denoising, compression, and anomaly detection.
- “The Next Rembrandt” (GAN art) and AIVA (RNN music) are landmark examples from the handbook.
- GAN Paint and Artbreeder are hands-on tools for experiencing GANs directly.
- Deep Fakes are a serious misuse of GAN technology — they can fabricate video of real people doing or saying things they never did.
- The five ethical considerations specific to GenAI are: Ownership, Human Agency, Bias, Misinformation, and Privacy.
EXERCISES
A. Fill in the Blanks
- A model that classifies existing data is called a ______ model.
- A model that creates new data by learning its underlying structure is called a ______ model.
- GAN stands for Generative ______ Network.
- In a GAN, the ______ network tries to create convincing fakes.
- In a GAN, the ______ network tries to tell real from fake.
- VAE stands for ______ Autoencoder.
- The compressed mathematical representation learned by a VAE is called the ______ space.
- RNN stands for ______ Neural Network.
- In an RNN, the ______ state carries memory from one step to the next.
- An Autoencoder forces the network to ______ and reconstruct data.
- The art project where a GAN painted in the style of a famous Dutch master is called “The Next ______.”
- AIVA is a generative AI that creates ______ using an RNN-based approach.
- AI-generated videos that realistically substitute a real person’s face are called ______.
- The ethical concern about who owns AI-generated content is called ______.
- The Transformer architecture — which powers modern LLMs — was introduced in ______.
B. Multiple Choice Questions
1. What is the primary difference between a discriminative and a generative model?
(a) Discriminative models use more data
(b) Generative models create new examples; discriminative models classify existing ones
(c) Generative models are always more accurate
(d) Discriminative models only work with images
2. In a GAN, what is the role of the Discriminator?
(a) To create new images from random noise
(b) To compress data into a latent space
(c) To evaluate whether an image is real or generated
(d) To memorise training examples
3. “The Next Rembrandt” used which type of generative model?
(a) RNN
(b) VAE
(c) GAN
(d) Autoencoder
4. What makes a VAE different from a standard Autoencoder?
(a) VAEs have a discriminator network
(b) VAEs encode data as a probability distribution, enabling generation of new variations
(c) VAEs can only work with text
(d) VAEs require more training data
5. Which model type is best suited for generating music that stays musically coherent across an entire composition?
(a) GAN
(b) VAE
(c) Autoencoder
(d) RNN
6. What is the Latent Space in a VAE?
(a) The final output layer
(b) The training dataset
(c) A compressed mathematical representation of the essential features of the data
(d) The discriminator network
7. AIVA, the AI music composer, primarily uses which generative model type?
(a) GAN
(b) VAE
(c) RNN (Recurrent Neural Network)
(d) Standard Autoencoder
8. Which of the following is the most accurate description of a Deep Fake?
(a) A photo taken with poor lighting
(b) AI-generated video that realistically replaces a real person’s appearance or voice
(c) A deliberately blurred video for privacy
(d) Augmented reality overlays on live video
9. Artbreeder is a tool powered by which generative model type?
(a) RNN
(b) Autoencoder
(c) GAN
(d) VAE
10. Which of the following is NOT one of the five ethical considerations specific to generative AI identified in the handbook?
(a) Ownership
(b) Human Agency
(c) Transparency of training data volume
(d) Misinformation
11. GANs were invented in which year?
(a) 2011
(b) 2013
(c) 2014
(d) 2017
12. An Autoencoder’s primary use in anomaly detection relies on:
(a) Its ability to generate convincing fake data
(b) Its high reconstruction error when shown unusual data it was not trained on
(c) Its discriminator network flagging anomalies
(d) Its latent space distribution
C. True or False
- A discriminative model can create new examples of data it has learned. (____)
- GANs use two neural networks that compete against each other during training. (____)
- The Generator in a GAN starts by producing high-quality outputs. (____)
- VAEs encode data as a single fixed point in latent space, just like standard autoencoders. (____)
- RNNs are designed to process sequential data where order and context matter. (____)
- AIVA was the first AI to be officially recognised as a composer by a music rights organisation. (____)
- Autoencoders learn by compressing data and then trying to reconstruct it. (____)
- Deep fakes can only be created using GANs. (____)
- Deep fakes pose risks including fabricated videos of real people making false statements. (____)
- Ownership of AI-generated content is a settled legal question globally. (____)
- VAEs were introduced before GANs. (____)
- “The Next Rembrandt” is an example of a GAN generating art in the style of a human artist. (____)
D. Define the Following (30-40 words each)
- Generative Model
- Discriminative Model
- GAN (Generative Adversarial Network)
- Generator (in a GAN)
- Discriminator (in a GAN)
- VAE (Variational Autoencoder)
- Latent Space
- RNN (Recurrent Neural Network)
- Hidden State (in an RNN)
- Autoencoder
- Deep Fake
- Ownership (as an AI ethics issue)
E. Very Short Answer Questions (40-50 words each)
- What is the difference between a discriminative model and a generative model?
- Explain how a GAN trains itself using the Generator and Discriminator.
- What is the latent space in a VAE, and why does it enable generation?
- Why are RNNs better suited for music generation than standard neural networks?
- What is an Autoencoder and how does it detect anomalies?
- What was “The Next Rembrandt” and which generative model type created it?
- What is AIVA and what makes its musical output coherent?
- What is a deep fake, and why is it an ethical concern?
- What does “Ownership” mean as an ethical consideration in generative AI?
- Name the four generative model types covered in this blog and one application for each.
F. Long Answer Questions (75-100 words each)
- Explain the difference between generative and discriminative modeling with examples of each. Why does the shift from discriminative to generative AI represent an important transition?
- Describe how a GAN works. What is the role of the Generator? What is the role of the Discriminator? What happens during training?
- How does a VAE differ from a standard Autoencoder? Why does encoding to a probability distribution (rather than a fixed point) enable generation?
- Why are RNNs especially useful for music and language generation? Explain the role of the hidden state with an example.
- Describe “The Next Rembrandt” project. What does it demonstrate about what GANs can learn and create?
- What are deep fakes? Explain how they are created, why they are dangerous, and how you might identify one.
- Explain the five ethical considerations specific to generative AI (Ownership, Human Agency, Bias, Misinformation, Privacy) with one real-world example for each.
- Trace the timeline of Generative AI from 2013 to 2023. Identify the three most significant technical milestones and explain why each was important.
ANSWER KEY
A. Fill in the Blanks – Answers
- discriminative — Discriminative models classify existing data.
- generative — Generative models create new examples.
- Adversarial — GAN = Generative Adversarial Network.
- Generator — The Generator creates fake outputs.
- Discriminator — The Discriminator evaluates real vs. fake.
- Variational — VAE = Variational Autoencoder.
- latent — The latent space is the compressed mathematical representation.
- Recurrent — RNN = Recurrent Neural Network.
- hidden — The hidden state carries sequential memory.
- compress — Autoencoders compress data and reconstruct it.
- Rembrandt — “The Next Rembrandt” was a GAN art project.
- music — AIVA generates original musical compositions.
- deep fakes — AI-generated synthetic media of real people.
- ownership — Who owns AI-generated content is a key ethical issue.
- 2017 — The “Attention is All You Need” paper introduced Transformers in 2017.
B. Multiple Choice Questions – Answers
- (b) Generative models create new examples; discriminative models classify existing ones — Core distinction.
- (c) To evaluate whether an image is real or generated — The Discriminator is the judge.
- (c) GAN — “The Next Rembrandt” used a GAN to analyse and replicate Rembrandt’s style.
- (b) VAEs encode data as a probability distribution, enabling generation of new variations — This is what makes VAEs generative.
- (d) RNN — Sequential memory makes RNNs suitable for coherent musical composition.
- (c) A compressed mathematical representation of the essential features of the data — The latent space is the core of a VAE.
- (c) RNN — AIVA uses recurrent architecture to maintain musical coherence over time.
- (b) AI-generated video that realistically replaces a real person’s appearance or voice — Definition of deep fake.
- (c) GAN — Artbreeder blends images using GAN-powered generation.
- (c) Transparency of training data volume — The five handbook principles are Ownership, Human Agency, Bias, Misinformation, and Privacy.
- (c) 2014 — Ian Goodfellow introduced GANs in 2014.
- (b) Its high reconstruction error when shown unusual data — Anomalies cannot be well-reconstructed, causing high error.
C. True or False – Answers
- False — Discriminative models classify; only generative models create new examples.
- True — Generator and Discriminator compete during GAN training.
- False — The Generator starts with random, low-quality output and improves through competition.
- False — Standard Autoencoders encode to a fixed point; VAEs encode to a probability distribution.
- True — RNNs process sequences in order, using hidden state memory.
- True — AIVA was recognised by SACEM (France) in 2017 as the first AI composer.
- True — Autoencoders are trained to compress and reconstruct their input.
- False — Deep fakes can be created using various generative models, not only GANs.
- True — Deep fakes have been used to create fake statements and compromising content involving real people.
- False — Ownership of AI-generated content is actively debated in courts globally; it is not a settled question.
- True — VAEs were introduced in 2013; GANs followed in 2014.
- True — “The Next Rembrandt” used a GAN trained on Rembrandt’s paintings to generate new art in his style.
D. Definitions – Answers
1. Generative Model: A type of AI model that learns the underlying structure and distribution of data through unsupervised or semi-supervised learning. It can create new examples — images, music, text — that resemble but are distinct from its training data.
2. Discriminative Model: A type of AI model trained through supervised learning to draw decision boundaries between classes. It classifies or labels existing data by distinguishing what makes inputs belong to one category versus another.
3. GAN (Generative Adversarial Network): A generative AI architecture with two competing neural networks — a Generator that creates fake outputs and a Discriminator that evaluates them. Their competition drives quality improvement until the Generator produces convincing results.
4. Generator (in a GAN): The network in a GAN responsible for creating new content. It takes random noise as input and produces outputs (images, audio, etc.) designed to fool the Discriminator into classifying them as real.
5. Discriminator (in a GAN): The network in a GAN that acts as a judge, evaluating whether inputs are real (from training data) or fake (from the Generator). Its feedback trains the Generator to produce increasingly convincing outputs.
6. VAE (Variational Autoencoder): A generative model that encodes input data as a probability distribution in latent space rather than a fixed point. Sampling different points from this distribution generates new, valid variations of the original input data.
7. Latent Space: The compressed mathematical representation at the centre of a VAE or Autoencoder. It encodes the essential features of data in a much smaller set of numbers. In VAEs, the latent space is a probability distribution from which new examples can be sampled.
8. RNN (Recurrent Neural Network): A type of neural network designed for sequential data. Unlike standard networks that process each input independently, RNNs maintain a hidden state that carries information from previous steps — giving them memory across a sequence.
9. Hidden State (in an RNN): The memory vector maintained by an RNN across sequential steps. At each step, the current input and the previous hidden state are combined to produce the next output, allowing the network to maintain context over the entire sequence.
10. Autoencoder: A neural network trained to compress data to a compact representation (encoding) and then reconstruct the original from that representation (decoding). It learns the most essential features of data through this compression-reconstruction process.
11. Deep Fake: AI-generated synthetic media — typically video — in which a real person’s face, voice, or actions are realistically replaced with fabricated content. Created using GAN architectures, deep fakes can make real people appear to say or do things they never did.
12. Ownership (as an AI ethics issue): The question of who holds copyright or intellectual property rights over content generated by AI. If a GAN creates a painting or a language model writes a story, ownership is contested between the programmer, the user, the data owners, and legal systems that do not yet recognise AI authorship.
E. Very Short Answer Questions – Answers
1. Discriminative vs. generative model:
Discriminative models classify existing data by learning decision boundaries (supervised learning) — e.g., spam vs. not spam. Generative models learn the underlying structure of data (unsupervised) and create new examples — e.g., new images, music, or text that resemble training data but are original.
2. How a GAN trains:
The Generator takes random noise and creates fake outputs. The Discriminator evaluates whether inputs are real (training data) or fake (Generator output). Both improve through competition: the Generator tries to fool the Discriminator; the Discriminator tries to detect fakes. Over time, the Generator produces increasingly convincing outputs.
3. Latent space in a VAE:
The latent space is a compressed mathematical representation of the essential features of data. In a VAE, it is a probability distribution rather than a fixed point. Because it is a range of values, sampling different points from the distribution produces new, valid variations — enabling genuine generation, not just reconstruction.
4. Why RNNs suit music generation:
Music is sequential — each note depends on what came before. Standard neural networks process inputs independently with no memory. RNNs maintain a hidden state that carries memory across the entire sequence, enabling the model to understand long-range musical structure: themes, development, resolution.
5. Autoencoder and anomaly detection:
An Autoencoder is trained to compress and reconstruct “normal” data accurately. When shown unusual data (anomalies) it was not trained on, it cannot reconstruct it accurately, resulting in high reconstruction error. This error spike flags the anomaly — used in medical monitoring, fraud detection, and quality control.
6. “The Next Rembrandt” and GANs:
“The Next Rembrandt” (2016) was a project in which a GAN analysed all 346 known Rembrandt paintings, learning his style, proportions, brush texture, and palette. It then generated an entirely new portrait — one Rembrandt never painted — indistinguishable from his real work. It demonstrated GANs can internalise an artist’s generative grammar.
7. AIVA and musical coherence:
AIVA is an AI music composer trained on thousands of classical scores. Using an RNN-based architecture, it maintains musical memory across an entire composition — knowing what themes have appeared, how tension should develop, and when resolution should arrive. This memory produces musically coherent pieces, not just note-by-note prediction.
8. Deep fakes and ethical concern:
A deep fake is AI-generated video that realistically substitutes a real person’s face or voice with fabricated content. They are ethically concerning because they can fabricate statements or actions by real people, enabling misinformation, fraud, reputation damage, and erosion of trust in all video evidence.
9. Ownership in generative AI ethics:
Ownership refers to who holds copyright over content created by an AI. When a GAN generates art or a language model writes an essay, the question of whether the user, developer, or original data owners hold rights is unresolved globally. This affects artists, publishers, legal systems, and anyone using AI-generated content commercially.
10. Four models and applications:
GANs: photorealistic image generation (The Next Rembrandt, Artbreeder). VAEs: generating new molecular structures for drug discovery. RNNs: composing sequential music (AIVA). Autoencoders: detecting anomalies in medical data by flagging high reconstruction error on unusual patterns.
F. Long Answer Questions – Answers
1. Generative vs. discriminative modeling:
Discriminative models learn decision boundaries from labeled data to classify inputs — a spam filter learns what separates spam from legitimate email. They ask: “What class does this input belong to?” Example: disease detector that labels X-rays as cancerous or not. Generative models learn the underlying structure and distribution of data to create new examples — they ask: “What does data from this class look like?” Example: DALL-E generating a new image from a text description. The shift from discriminative to generative AI is significant because it moves AI from an analytical tool (understanding what exists) to a creative tool (producing what has never existed). This opens entirely new applications in art, science, and language but also introduces new ethical challenges.
2. How a GAN works:
A GAN (Generative Adversarial Network) contains two competing neural networks. The Generator takes random noise as input and produces fake output — initially poor quality. The Discriminator receives both real training data and the Generator’s output, and must classify each as real or fake. Both networks receive feedback: the Generator receives a signal telling it how often it fooled the Discriminator; the Discriminator receives a signal on its accuracy. Training proceeds as competition: the Generator tries to fool the Discriminator; the Discriminator tries to catch fakes. Over many training cycles, the Generator improves until it produces outputs the Discriminator cannot reliably distinguish from real data. At this equilibrium, the Generator has learned the underlying structure of the real data well enough to create convincing new examples.
3. VAE vs. Standard Autoencoder:
A standard Autoencoder encodes each input to a single fixed point in latent space. If you give it a cat image, it always maps to exactly point [3.2, 1.1, 7.8]. Reconstructing from this point gives back the same cat — useful for compression but not for generation. A VAE instead encodes each input to a probability distribution — a range of possible points [2.9–3.5, 0.8–1.4, 7.5–8.1]. This distribution describes uncertainty: “this is approximately where this cat lives in latent space, with some variation.” When generating, you sample a random point from this distribution and decode it. Different samples give different valid outputs — all recognisable as cats but each unique. The probability distribution creates the generative capability that a standard Autoencoder lacks.
4. RNNs for music and language:
Standard neural networks treat each input independently with no memory of previous inputs. For a melody, this means each note is predicted without knowledge of what preceded it — resulting in incoherent sequences. RNNs solve this by maintaining a hidden state — a vector of numbers summarising everything processed so far. At each step, both the current input and the previous hidden state are combined to produce the output. For music: if a theme appears in bar 4, the hidden state remembers it. When the theme returns in bar 32, the network’s memory allows it to vary and develop it appropriately. For language: the word “bank” is interpreted using the full context of preceding words — its meaning in “river bank” differs from “bank account.” This sequential memory makes RNNs naturally suited to any data where context and order matter.
5. “The Next Rembrandt”:
“The Next Rembrandt” was a 2016 collaboration between Rembrandt House Museum, Microsoft, and a team of researchers and artists. A GAN was trained on all 346 known Rembrandt paintings — analysing his distinctive use of light and shadow, his facial proportion standards, his brush texture, his colour palette, and his compositional patterns. The GAN did not copy any painting. Instead, it learned the generative grammar of Rembrandt’s style — the rules by which he composed figures, distributed light, and built facial structure. Given a brief (a white male, 30-40 years old, facing right, with a ruff and hat), the GAN generated an entirely new portrait consistent with Rembrandt’s style. The result was 3D-printed with raised paint texture to replicate the physical quality of an oil painting. This project demonstrated that a sufficiently trained GAN does not imitate — it internalises an artist’s visual intelligence deeply enough to extend it into new, original work.
6. Deep fakes:
Deep fakes are AI-generated synthetic media — typically video — that realistically replace a real person’s face, voice, or actions with fabricated content. They are created using GAN-based architectures: the model is trained on real footage of a person to learn their facial movements, expressions, and voice patterns. It then generates new frames showing the person in scenarios they never experienced. Deep fakes are dangerous for several reasons. For individuals: they can be used to create compromising or false videos that damage reputations or constitute harassment. For society: fabricated videos of politicians or leaders making false statements can influence elections or trigger conflicts. For institutions: deep fake phone calls or videos can impersonate executives to authorise fraud. For truth itself: when any video can be convincingly faked, legitimate video evidence loses credibility. Current signals for detection include unnatural blinking, blurring around hairlines, inconsistent lighting, and subtle distortions during head movement — though detection technology and deep fake quality are both improving rapidly.
7. Five ethical considerations:
Ownership: A musician uses an AI to compose a song and releases it commercially. Who holds the copyright — the musician, the company whose AI generated the music, or the artists whose works trained the model? No clear legal answer exists globally. Human Agency: A student uses an AI to write all their essays throughout school. They graduate able to prompt an AI but unable to write a coherent paragraph themselves. AI as a tool enhances capability; AI as a replacement erodes it. Bias: An AI image generator trained predominantly on Western media generates “doctor” images that are overwhelmingly male and light-skinned — reflecting and amplifying historical representation biases at scale. Misinformation: A fabricated but convincing AI news article about a disaster triggers panic and stock market movement before fact-checkers can respond. Generative AI enables misinformation at speeds and volumes that outpace human verification. Privacy: An AI voice generator trained on a person’s recorded speech without consent can produce audio of that person saying anything — a form of identity theft that existing privacy law was not designed to address.
8. Generative AI timeline (key milestones):
Three milestones stand out as technically transformative. First: VAEs in 2013. Before VAEs, generating new data meant memorising and replaying examples. VAEs introduced latent space encoding — for the first time, AI could learn a compressed, manipulable representation of data structure and generate genuine variations. This was the mathematical foundation of modern generative AI. Second: GANs in 2014. Ian Goodfellow’s adversarial training idea solved the problem of how to teach an AI to produce realistic outputs without a human judge at every step. The Generator-Discriminator competition created a self-improving system that drove image quality from cartoonish to photorealistic within a few years. Third: Transformers in 2017. The Transformer architecture’s attention mechanism allowed AI to process entire sequences simultaneously (not step-by-step like RNNs) and maintain context over very long distances. This enabled the scale of pre-training on which GPT, Gemini, and all modern LLMs are built. Without Transformers, ChatGPT could not exist. Each milestone solved a fundamental constraint of the previous approach, enabling the next generation of capability.
Activity Answers
- Lullaby composition (RNN): Music is sequential — each bar must follow naturally from the last. An RNN’s hidden state carries musical memory across the entire piece, enabling coherent verse-to-verse development that a model without memory cannot produce.
- Fraud detection (Autoencoder): The Autoencoder trains on normal transactions and learns to reconstruct them with low error. Fraudulent transactions have unusual patterns — reconstruction error spikes, flagging the anomaly for review.
- New molecular compounds (VAE): VAEs encode the essential features of known drug molecules into latent space. Sampling different points from this space generates new molecular structures with properties interpolated from those of known effective compounds — enabling rational drug design.
- Sky replacement (GAN): GANs are trained to generate photorealistic content that is indistinguishable from real photographs. The Generator learns how real skies look — cloud formations, lighting conditions, colour gradients — and produces a new, photorealistic sky seamlessly composited into the original image.
