Step 1 - Introduction to AI

Step 2 - Understanding Machine Learning

Step 3 - Basics of Deep Learning

Step 4 - Introduction to Generative AI

Step 5 - What is Large Language Model (LLM)?

Step 6 - Prompt Engineering: The Art of Communicating with AI

Step 7 - Ethical Considerations in Generative AI

Step 8 - Challenges and Limitations in Generative AI

Step 1 - Introduction to AI

Let’s start the part 1 — Artificial Intelligence!

Artificial Intelligence (AI) — From a Kid’s Perspective

Let’s first have the simplest understanding of AI. Imagine you
have lost your dog, and you need to find him.

Here are some of the capabilities you need to find your dog:

You should be able to Identify your dog.

If you see any animals, you should be able to identify if it’s a dog
or not. If it’s a dog, you need to further identify if it’s your dog.

You should be able to make a strategy to find your dog.

You need to be able to make a strategy to find your dog. For

example:

· First search in our house.

· If you don’t find him, then search in play area where you usually
go with your dog.

· If you don’t find him yet, ask your friends.

· And so on….

You should be able to act according to situation.

For example, if it’s raining, and you know that your dog does not
prefer to get wet, you will focus your search on shaded places.

Now, imagine someone told you — “I have probably seen your

dog in garden”.

You (Actually Your Brain) know what to do.

· You know where garden is and how to go there.

· You will not confuse a cat or a tree with a dog.

· The moment you see a dog you will try to identify if it’s your dog
or not.

You could search your dog because you have all

these intelligences.

What if somehow, we could give all these intelligence to a

robot so that next time you lose your dog, your robot could find
him.
Imagine the robot can move and capture videos. But that’s not
enough. To find your dog, we need to enable this robot to think
like you and act like you.

For example:

• We enable the robot to identify your room. But it should be

able to recognize the room even if your bed is moved to
another wall, or blanket is changed. It needs
INTELLIGENCE to identify room even with new changes.

• We enable the robot to identify a dog and distinguish your

specific dog.
 • We enable the robot to understand human language and
instructions.

• We enable the robot to come up with a strategy and act as

per new situations. For example, search only in shaded
places if it’s raining.

In summary, to find your dog, the robot needs HUMAN LIKE

INTELLIGENCE.

If we could do that, next time you lose your dog, your robot friend
might just find him using its artificial intelligence.

This is Artificial Intelligence (AI) — Human like intelligence,

created in a robot (or a machine or computer) by human.

What is AI?
Artificial intelligence is when machines/computers mimic the
way humans think and make decisions.

AI enables computers to think as we human think.

In simple words — AI is when we enable computers to Think.

AI enables computers to understand, analyze data, and make
decisions without constant human guidance. These intelligent
machines use algorithms, which are step-by-step instructions, to
process information and improve their performance over time.

Real-world Examples of AI Applications

You’ve probably used AI even without knowing it! Voice

assistants such as Siri and Alexa or those helpful chatbots when
you’re on websites or generative AI tools such as ChatGPT and
Google’s Bard — they all use AI technology to make things easier
for you.

Let’s take a peek into some of the common usages of AI in our

daily life:
Virtual Assistants

Virtual assistants such as Siri or Alexa uses AI to understand our

questions and commands. They can answer questions, play your
favourite tunes, and even control your smart home devices.

Social Media Algorithms

Ever notice how Netflix suggests shows you might enjoy? Or how
Facebook’s suggested feed seems to know exactly what you want
to see — that’s AI at play!

Netflix usages AI to analyse your watching habits

to offer personalized recommendations. Similarly, other social
media platforms use AI to personalize your experience, showing
you content that matches your interests.

Online Shopping Recommendations

Have you ever wondered how online stores suggest products you
might buy?

When shopping online, AI algorithms examine your preferences,

your past choices and those of similar shoppers to recommend
items tailored just for you.

Predictive Text and Autocorrect

When your smartphone suggests the next word you want to type,
that’s AI predicting what you might say next.

Healthcare Diagnostics

AI helps doctors analyse medical images such as X-rays and MRIs

more quickly and accurately. This speeds up diagnosis and
improves the chances of successful treatment.

Language Translation Services

When we plan a trip abroad and use language translation

services, such as Google Translate, it usages AI
algorithms. These AI-powered language translation services
help bridge language barriers, making communication easier in
different parts of the world.
Fraud Detection in Banking

Now-a-days AI keeps a watchful eye on bank transactions. If

it spots something fishy, for example an unusual purchase, it can
alert you or even block the transaction to protect your account.

These examples show that AI isn’t confined to labs or the distant

future. It’s an integral part of our daily lives, working quietly
behind the scenes to make our life better.

AI vs. Human Intelligence

At one side, AI enables computers to become intelligent with
numbers and rules, doing super quick math with perfect
accuracy. On the other side, we human have brain and we are
also driven by emotions, creativity, and the ability to adjust to
all sorts of situations. Our brain is always evolving, adapting and
thinking new things.
It’s similar to comparing a super-fast calculator to a vibrant, ever-
evolving masterpiece!

Here are some major differences between AI and Human

Intelligence:

Learning Style:

• AI: Learns from loads of examples and

data. It crunches numbers and patterns to become a pro at
specific tasks.

• Humans: We learn by talking, experiencing, and

thinking. Our brains soak up a mix of things — from how to
ride a bike to why the sky turns pink at sunset.
Thinking Speed:

• AI: Fast, similar to a superhero at tasks it knows well. Show

it a task it’s trained on, and boom, it’s done in a flash.

• Humans: We might take a bit more time. But we are super

good at figuring out complex stuff. We are good in complex
thinking and creativity.

Memory Skills:

• AI: Remembers facts and figures but not with memories and
feelings. It’s a robot recalling programmed info rather
than cherishing a moment.

• Humans: We remember events, emotions, and lots of

details. From first dates to the lyrics of our favourite songs.
Our memories are collection of good and bad experiences.

Feeling Emotions:

• AI: Doesn’t feel joy, sorrow, or anything. It sticks to rules and

patterns.

• Humans: We’re an emotional rollercoaster — happiness,

sadness, and everything else. Our feelings shape who we
are and how we react.
Flexibility Factor:

• AI: Sticks to what it’s taught and might struggle in new

situations. It’s smart but rigid.

• Humans: We’re amazing in adapting new things. We

humans always figure out how to come out of any scenario
and solve any problem.

Creating Cool Stuff:

• AI: Can create things within its set limits. It may be

considered as an artist with a specific canvas and color
palette.

• Humans: We’re the masters of making things up — new

ideas, art, solutions. Our creativity knows no bounds.

Understanding the Big Picture:

• AI: Knows what it’s learned but might miss tricky situations,
for example reading between the lines, understanding inside
jokes or cultural nuances.
 • Humans: We understand everything — jokes, feelings, and
culture. Our brains is a complete packages that have a bit of
everything!

Decision Making Capabilities:

• AI: Decides based on its training and programming. It

follows the rules.

• Humans: We blend logic, feelings, and what’s right to make

decisions.

Types of AI
Artificial Intelligence is divided based on two main categorization
— based on capabilities and based on functionally of AI.

The following image illustrates these types of AI:

Types of AI — Based on Capability

Based on capability, there are 3 types of AI — Narrow AI, General

AI and Super AI.

1. Narrow AI

Narrow AI, also known as Weak AI, refers to artificial intelligence

systems that are designed and trained for a specific
task or a narrow set of tasks.

Have you seen a computer playing chess? That’s Narrow AI at

work. It’s superb in playing chess but won’t be as good at, say,
translating or in speech recognition.

Another good example of narrow AI virtual assistants such as Siri

or Alexa. Siri/Alexa is good in speech recognition but operates
with a limited pre-defined range of functions.

Other examples of narrow AI include:

• Self-driving cars

• Google search

• Conversational bots

• Email spam filters

• Netflix’s recommendations etc.

2 important point on Narrow AI:

 • Narrow AI is focused on performing a single task extremely
well.

• But it cannot perform beyond its field or limitations.

Almost all the AI-based systems built till this date fall under
the category of Weak AI.

2. General AI

General AI, also known as Strong AI or artificial general

intelligence (AGI), can understand and learn any intellectual
task that a human being can.

It refers to artificial intelligence that:

• Possesses the ability to understand, learn, and apply

knowledge across a wide range of task

• at a level equivalent to human intelligence.

Currently, there is no such system exist which can come

under general AI and can perform any task as perfect as a
human.

Creating Strong AI system poses significant scientific and

technical challenges.
Researchers and developers continue to make advancements in
various AI fields, but achieving true General AI, which mirrors the
broad capabilities of human intelligence, is a complex and
ongoing endeavour.

3. Super AI

Super AI represents a degree of intelligence in systems

where machines have the potential to exceed human
intelligence, outperforming humans in tasks
and exhibiting cognitive abilities.

Super AI is still a hypothetical concept of Artificial

Intelligence. Development of such systems in real is still world
changing task.

We have only seen Super AI systems/characters in movies such

as I,Robot, Terminator, The Matrix, Blade Runner etc.
A scene from movie I,Robot showing VIKI (Virtual Interactive
Kinetic Intelligence)

For example, in movie “I, Robot,” we get a glimpse of a future

world where Super AI plays a pivotal role. The central AI system in
the film is named VIKI, which goes beyond typical AI
capabilities. VIKI’s intelligence evolves into a form of Super
AI, where it surpasses its initial programming and starts making
decisions to “protect” humanity in a controversial way.
A Quick Comparison of Narrow AI, Strong AI and Super AI

Narrow AI (Weak AI):

• What it is: Similar to a specialist, good at one specific task.

• Example: Siri or Alexa — great at understanding and

responding to voice commands but not much beyond that.

• Analogy: Imagine a superhero with a

superpower dedicated to a particular task. For example a
hero who excels only in solving puzzles.

Strong AI (General AI):

• What it is: Similar to a human super hero, who can

understand, learn, and perform various tasks.
 • Example: Currently more theoretical, no real-world
examples yet.

• Analogy: Imagine a superhero with a whole array of

superpowers, able to adapt and excel in different situations.

Super AI:

• What it is: Similar to

an ultimate superhero, surpasses human intelligence and
can do pretty much anything better than humans.

• Example: Still theoretical, no real-world examples.

• Analogy: Imagine a superhero with the combined abilities of

all superheroes, making them unmatched and capable of
handling any situation with ease.

Types of AI — Based on Functionality

Based on functionality, there are 4 types of AI

— Reactive Machines, Limited Memory, Theory of Mind and Self
Awareness.

1. Reactive Machines

Reactive machines are AI systems that have no memory. These

systems operate solely based on the present data, taking into
account only the current situation. They can perform a
narrowed range of pre-defined tasks.

In a nutshell, Reactive machines are:

· AI systems which do not store memories or past experiences for

future actions.

· It only focus on current scenarios and react on it as per possible

best action.

Garry Kasparov playing against Deep Blue, image source

britannica.com

One of the examples of reactive AI is Deep Blue, IBM’s chess-

playing AI program, which defeated world champion, Garry
Kasparov in the late 1990s. Deep Blue had ability to identify its
own and its opponent’s pieces on the chessboard to make
predictions, but it didn’t have the memory to use past mistakes
to inform future decisions.

2. Limited Memory

As the name indicates, Limited Memory AI can

take informed and improved decisions by looking at its past
experiences stored in a temporary memory.

This AI doesn’t remember everything forever, but it uses its short-

term memory to learn from the past and make better decisions
for the future.

A good example of Limited Memory AI is Self-driving cars. The

AI system in self-driving car utilizes recent past data to make real-
time decisions. For instance, they employ sensors to recognize
pedestrians, steep roads, traffic signals, and
more, enhancing their ability to make safer driving choices.
This proactive approach contributes to preventing potential
accidents.

Another example is recommendation systems. Platforms such

as Netflix or Amazon use Limited Memory AI to suggest
movies, products, or content based on a user’s past
preferences and behaviours.
3. Theory of Mind

The initial two categories of AI — Reactive Machines and Limited

Memory, presently exist.

Next 2 types of AI — Theory of Mind and Self-aware AI,

however, are theoretical types that could be developed in the
future. As of now, there is no real-world examples of these types
are available.

Theory of Mind is supposed to have capability to understand the

human emotions, people, beliefs, and be able to interact
socially same as humans.

4. Self-Aware AI

This is similar to Super AI — We should pray that we don’t reach

the state of AI, where machines have their
own consciousness and become self-aware.

Self-aware AI systems will be super intelligent, and will have their

own consciousness, sentiments, and self-awareness. They will
be smarter than human mind.

As shown in movie “I, Robot,”, an AI system named VIKI becomes

self-aware and starts making decisions to “protect” humanity in a
controversial way.
Similar to Theory of Mind, Self-aware AI also does not exist in
reality. Many experts, for example Elon Musk and Stephen
Hawkings have consistently warned us about the evolution of AI.

Stephen Hawking stated that:

“The development of full artificial intelligence could spell the end

of the human race…. It would take off on its own, and re-design
itself at an ever-increasing rate. Humans, who are limited by slow
biological evolution, couldn’t compete and would be
superseded.”.

Summary
With this first step, we’ve taken the first step on a journey to
understand Generative AI. We learnt what AI is and explored
its fundamental concepts. We learnt types of AI based on
different categories and also understood how AI is different from
human intelligence.
Step 2 - Understanding Machine Learning

Machine Learning (ML) — From a Kid’s Perspective

In ourstep, while understanding AI, we talked about enabling the
robot to identify a dog. Imagine we want to enable the robot to
identify several animals.

To do so, we will show him pictures of various dogs, cats, bunnies

and other animals and label each picture with the name of the
animal. We train the robot to identify animals based on size,
colour, body shape, sound etc.

Once the training is completed, the robot will be able to identify

these animals we trained him for.
All dogs do not look alike. However, once robot has seen many
pictures of dogs, it can identify any dog even if it does not
exactly look like a specific picture. We need to show lots of
pictures of dog to the robot. More pictures it sees, more efficient
it will be.

This is Machine Learning — Teaching a robot (or any machine)

by giving lots of example pictures (or any other information).

To summarize, Machine Learning is:

• A subset of Artificial Intelligence.

• Which enables machines (or computers) to learn from

data and make decisions.
Types of Machine Learning
Machine learning can be broadly categorized into three main
types:

1. Supervised learning

2. Unsupervised learning

3. Reinforcement learning.

Each type serves different purposes and involves different

approaches to learning from data. Let’s have a close look into all
these types.

Supervised Learning
Let’s take the same example where we enabled the robot to
identify an animal.

When we trained our robot by showing pictures of animals, we

labelled each picture with the name of the animal. So, we acted
as a teacher to him. We first told him how does a dog or a cat
look like and then only he was able to identify them.

In Machine Learning we call this Supervised Learning.

Below image summarizes important points on supervised

learning.
Real-life Examples of Supervised Learning

Supervised learning is widely used in various real-life applications

where the algorithm is trained on labelled data to make
predictions or classifications. Here are some examples:

Email Spam Filtering

Classifying emails as spam or not spam based on

features derived from the content, sender information, and
other relevant attributes.

Image Classification

Identifying objects or patterns within images, such as classifying

animals, recognizing handwritten digits, or detecting objects in
self-driving cars.
Facial Recognition

Identifying and verifying individuals based on facial features, used

in security systems or for unlocking devices.

Financial Fraud Detection

Identifying potentially fraudulent transactions by analyzing

patterns and anomalies in financial data.

Speech Recognition

Converting spoken language into text, as seen in voice assistants

such as Siri or Google Assistant.

Unsupervised Learning
Let’s understand this from a kid’s school example. When kids go
to their class first day, they meet lots of classmates. At first all
classmates are same to them. But with time, they themselves
categorized them in different groups:

• They find some classmates very good and want to be friend

with them.

• They find some rude or irritating and want to avoid them.

• They find some very good in sports and want to be in the

same team as they are.

• And so on…
When kids categorized their classmates, nobody told them
how to do that. They did that without anyone’s help. — This is
how unsupervised learning works.

Let’s take a proper machine learning example. Imagine we

showed lots of pictures of dogs, cats, bunnies etc. without any
label to our robot and told him — “I’m not going to tell you which
one is which. Go explore and figure it out”.

The robot starts to look at these animals, noticing things such as

their fur, size, and how they move. It doesn’t know their names
yet, but it’s trying to find patterns and differences on its own.

After exploring, the robot might notice that:

• Some animals have long ears (bunnies)

• Some animals have soft fur and a tail (cats)

• Some animals have wagging tails (dogs)

It figures out these categories without you telling it directly.

In the end, the robot might not know the names of the animals,
but it can say that “These animals are similar in some ways,
and those are different in other ways.” — This is Unsupervised
Learning.

Below image summarizes important points on unsupervised

learning.
Real-life Examples of Unsupervised Learning
Unsupervised learning is used in various real-life scenarios where
the data is not labelled, and the algorithm needs to discover
patterns, structures, or relationships within the data. Here are
some examples:

Clustering Customer Segmentation

Businesses use unsupervised learning, specifically clustering

algorithms like k-means, to segment customers based on their
purchasing behavior. This helps in targeted marketing and
personalized services.

Anomaly Detection in Cybersecurity

Unsupervised learning is employed to identify unusual patterns or
behaviors in network traffic. Any deviation from the normal
behavior can be flagged as a potential security threat.

Recommendation Systems

Unsupervised learning is used in recommendation systems. By

identifying patterns in user behavior, these systems can suggest
products, movies, or content that a user might like.

Reinforcement Learning
Imagine teaching a dog a new trick — you reward it with a treat
when it does the trick correctly and give no treat when it
doesn’t. Over time, the dog learns to perform the trick to get
more treats.

Similarly, Reinforcement Learning is:

• Training a computer to make decisions

• By rewarding good choices and punishing bad ones

• Just as you might train a dog with treats for learning tricks
In reinforcement learning, there’s an agent (for example a robot or
computer program) that interacts with an environment. Let’s take
an example of teaching a computer program to play a game, for
example chess.

• In this case, computer program is agent and chess game is

the environment.

• The computer program can make different moves in the

game, such as moving a chess piece.

• After each move, it receives feedback (reward or penalty)

based on the outcome of the game.

• If the program wins the game, it receives a positive reward.

 • If it loses the game, it receives a negative reward, or a
‘penalty.

• Through trial and error, the program learns which moves

lead to the best rewards, helping it figure out the
best sequence of moves that leads to winning the game.

Reinforcement learning is powerful because it allows machines

to learn from their experiences and make decisions in complex,
uncertain environments — similar to how we learn from trial and
error in the real world.

Below image summarizes important points on reinforcement

learning.
Real-life Examples of Reinforcement Learning
Game playing is one of the main use-case of reinforcement
learning.

AlphaGo, developed by DeepMind, is a computer program that

uses reinforcement learning to play the board game Go at a
superhuman level. It defeated world champions and
demonstrated the power of reinforcement learning in mastering
complex games.

Another example is Self-driving cars. Reinforcement learning is

used in the development of self-driving cars. Agent learns how to
navigate traffic, make decisions at intersections, and respond to
various driving conditions through continuous learning from
simulated and real-world experiences.

Reinforcement learning is also used in algorithmic trading to

make decisions on buying or selling financial instruments. The
agent learns optimal trading strategies based on historical market
data and real-time market conditions.

Summary

Machine Learning is a subset of AI where we enable computers to

learn from examples and experiences. We don’t explicitly
program but let the machine learn from data and figure things out
on its own.
Step 3 – Deep Learning

What is Deep Learning?

Can the machine learn the way we human (human brain) learn
things? — This was the idea behind innovation of Deep Learning.

Deep learning is a subset of Machine Learning (ML is again a

subset of AI). At its core, deep learning is based on Artificial
Neural Network (ANN), which is a computational models
inspired by the structure and functioning of the human brain.

Sounds a bit confusing? Let’s simplify it in layman’s terms!

First, let’s understand few important concepts.

Biological Neural Network in Human Brain

A neuron is the human brain’s most fundamental cell. A human
brain has many billions of neurons, which interact and
communicate with one another, forming a neural network.

These neurons take in many inputs, from what we see and hear to
how we feel to everything in-between, and then send messages to
other neurons, which react in turn. Working neural networks are
what enable humans to think, and more importantly, learn.
Artificial Neural Network (ANN)
Artificial neural network is a computational network
designed based on biological neural networks in human brain.

Human brain has neurons interconnected to each other.

Similarly, artificial neural networks also have neurons that are
linked to each other. These neurons are known as nodes.

Let’s try to simplify ANN!

Picture making a big, 3D structure with pipes of different shapes

and sizes. Each pipe can connect to lots of other pipes and has a
switch that can be opened or closed. This gives you so many ways
to connect the pipes, making it seem a bit tricky, right?

Now, let’s attach this pipe thing to a water tap. The pipes, which
are of different-size, let the water move at different speeds. If we
close the switches, the water won’t move.

The water represents data going through the brain, and

the pipes represent the brain’s parts called neurons.
Architecture of an artificial neural network
Artificial Neural Network primarily consists of three layers —
Input Layer, Output Layer and Hidden Layers.

Imagine an Artificial Neural Network similar to a sandwich with

three layers.

The first layer, called the Input Layer, represent the bottom slice
of bread. It takes in information.

The second layer, called the Hidden Layers, represent the

yummy filling in the middle. It thinks and figures things out.

The third layer, called the Output Layer, represent the top slice of
bread. It gives us the final result.

In a nutshell:

Input Layer
 • This is where information goes into the artificial neural
network.

• It’s the starting point, where the network receives the data it
needs to work on.

Output Layer

• This is where the network gives the final result or answer.

• It’s the endpoint, where the network tells us what it has

learned or decided.

Hidden Layers

• These layers are in between the input and output layers.

• Neurons in these layers process information and help the

network learn patterns and make decisions.

How does Artificial Neural Network Work?

Imagine a group of kids trying to recognize a panda by sharing
their observations.

• Each kid focuses on specific features such as black-and-

white fur, round face, and distinct eyes.

• Individually, they might not fully understand what a

panda looks like,
 • But by combining their insights, they create a collective
understanding.

In the world of artificial neural networks, these kids represent

neurons.

• In artificial neural network, individual “neurons” (similar to

kids in our example) specialize in recognizing specific
aspects.

• When combined, they contribute to recognizing the overall

concept (panda).

• The network refines its understanding through

repeated exposure, similar to kids refining their panda
recognition skills over time.

Input Layer (Observation):

Each kid observes one aspect, such as fur colour or face shape,
forming the input layer of our network.

Hidden Layers (Processing):

The kids pass their observations to each other, mimicking the

hidden layers of a neural network. As they share information, they
collectively build a more comprehensive understanding of the
panda’s features.

Output Layer (Recognition):

Finally, they reach a conclusion by combining all the details. If the

majority agrees that the observed characteristics match those of
a panda, they output “panda.” This output layer corresponds to
the network’s final decision.

Scoring Approach:

To refine their recognition skills, the kids keep track of their

accuracy.

• If they correctly identify a panda, they gain points;

• otherwise, they learn from their mistakes.

• Similarly, in neural networks, a scoring approach

helps adjust the network’s parameters to enhance accuracy
over time.
This teamwork illustrates how artificial neural networks process
information layer by layer, learning from various features and
refining their understanding through a scoring mechanism.

Deep Neural Networks

A deep neural network (DNN) is an artificial neural network
(ANN) with multiple layers between the input and output layers.

Here “Deep” means it has multiple layers between the input and
output, making it capable of learning complex patterns.

Important Points about Deep Learning

Now, let’s summarize some important points on Deep Learning!

Subset of ML

Deep learning is the subset of machine learning, which is in turn

subset of AI.

Inspired by the Brain

Deep learning is based on artificial neural networks which is

inspired by how our brains work.

Artificial Neural Networks (ANN)

ANN is a computational network which mimics biological neural

networks in human brain.
Deep Neural Networks

The adjective “deep” refers to the use of multiple layers in the

network. It uses deep neural networks with more than one hidden
layer.

These layers process information, allowing the system to learn

complex patterns.

Learning from Data

The system learns by being shown lots of examples and adjusting

connections between neurons based on the differences between
predictions and correct answers.

Handling Complex Problems

Deep learning is particularly effective for solving complex

problems where traditional approaches may struggle.

Machine Learning Vs Deep Learning

Let’s break down the major differences between machine
learning and deep learning:
Summary
In this step, we’ve understood what deep learning is and how it
works. Deep learning, aptly named for its multi-layered neural
networks, which is similar to a human brain neural network with
multiple layers of thinking, each level contributing to a deeper
understanding of the information it processes.
Step 4 –

What is Generative AI?

Generative AI is:

• A type of artificial intelligence

• that can create new things, for example artwork, music, or

even realistic images.

• without being explicitly told what to create

While traditional AI focuses on specific tasks or solving a

problem, Generative AI is distinguished by its ability
to exhibit creativity similar to human creativity. Generative AI is
capable of generating new, unique content, ideas, or solutions as
we human do.

Let’s understand it better with an example!

Imagine I asked you to draw an animal you have never seen
before. You need to use your imagination and draw a brand-new
animal the world has never seen.

Since we human have imaginative power and creativity, you will

be able to do that. Maybe you will draw an animal that has the
body of a lion, face of a cow and the wings of a butterfly.
Now, what if a computer program could create new things all
by itself! It can create new things, for example artwork, music, or
even realistic images, without being explicitly told what to create.

The computer program has been given lots of pictures of lions,

cow and butterflies. Now, with this knowledge, it can draw a
completely new animal, say a “lion-cow-butterfly”
combination. It doesn’t copy any existing image; instead, it
uses its understanding of what makes lion, cow, and butterfly
unique to create something entirely new something as below.
Image generated using fotor.com

This is Generative AI — A machine (or computer) which has

imagination and creativity to draw pictures, tell stories, or
even make up new games without anyone showing it how.

Where Does Generative AI Fits into AI Hierarchy?

Generative AI is a subset of Deep Learning. Below diagram shows
the relation between AI, Machine Learning, Deep Learning and
Generative AI.

Generative AI leverages machine learning techniques,

particularly deep learning and neural networks.

The main differentiator of Generative AI is the ability to

generate new content.

AI, machine learning and even deep learning is mostly limited to

predictive models. These are mainly used to observe and classify
patterns in content or predict a new pattern or content. For
example, a classic machine learning use-case is to identify image
of a cat out of several given images or classify animals in different
clusters based on various properties.
Generative AI is a breakthrough, because it has the ability to
do something only humans were supposed to do — create an
image of a cat or create an image of a totally new animal from
it’s creativity.

The following image shows the evolution of AI with time. The

evolution of AI from traditional rule-based systems to Generative
AI has been driven by advancements in learning algorithms,
computational power, and access to vast amounts of data.

Generative Models
Generative AI uses different types of machine learning models,
called Generative Models.

The generative models:

 • learns the underlying set of data and generates new data the
closely mimics the original data

• are mainly used to create new content, such as images,

text, or even music which looks exactly the same as what
might be created by humans

• Usages unsupervised learning approach

Most common generative models are:

• Variational Autoencoders (VAEs),

• Generative Adversarial Networks (GANs)

• Limited Boltzmann Machines (RBMs)

• Transformer-based Language Models

In the next chapter, we will learn more about generative models.

Usages of Generative AI in Real-life

Here are some examples of how generative AI is being used to
create real-life applications:
Text Generation
Most of us have used ChatGPT which is based on Generative
AI. Similar to ChatGPT Generative AI based tools can be used to
generate new content such as articles, reports, poetry, stories or
any other text-based content.

One of the most common uses of generative AI is to build Virtual

Assistants and Chatbots. Generative models are used to build
advance chatbots which can interact with users mimicking
human interaction.
Image Generation
Generative AI tools are used to generate new pictures even
creative ones using various generative models. These models can
learn from large sets of images and generate new unique images
based on trained data. These models can even generate images
with creativity based on input prompts similar to content
generated by humans. There are various ways this can be used in
real-life applications such as image-to-image translation, text-to-
image translation, photograph editing, face generation, image
quality enhancement etc.

One of the most common tools which usages generative AI to

create realistic images and art is DALL-E, developed by OpenAI. It
is a text-to-image model, which usages deep learning to generate
digital images from natural language descriptions.

Video Generation
Generative models can be used to create whole videos from
scratch. It stitches together scenes, characters, and actions to
make a story. These videos can be used for entertainment,
advertisements, or even training simulations. Video game
development is one field which is heavily using generative AI.
Some generative models can be used to create new videos by
learning from existing videos. This can be used for video
prediction if an existing video such as security clip is damaged.

Voice Generation
Generative AI can also mimic voices or generate a whole new
voice! It can learn how people talk by analysing audio data, and
then generate voice in same style or create entirely new voices.

This is useful for making virtual assistants or audiobooks sound

more natural.

Healthcare Applications
Generative AI models can be used to generate synthetic data
samples that resemble real data. This can be very useful in
medical field, where sometimes collecting real-world data is
expensive or limited. For example, generative AI can be used to
generating synthetic patient data for research purposes.

Drug Discovery
Generative AI is being used in drug discovery to generate new
molecular structures with desired properties. This helps
accelerate the process of drug development by exploring vast
chemical spaces and identifying promising drug candidates.
Gaming
Generative AI has truly changed the world of gaming. It is
increasingly being used in the gaming industry to accelerate
game production and create unique experiences.

It helps game developers make games more exciting and

immersive by creating entire worlds, characters, and stories.

Generative AI can also be used to make virtual worlds more

realistic. It can be used to create unique creatures and
characters, finetune each character’s personality and traits,
making the game feel alive and full of surprises.

Art Generation
This is one major usage that distinguish generative AI from regular
AI. Generative AI has the capability of creative thinking like we
human do. Various generative models are used in generative
artistic artifacts such as paintings, poetries, stories, and other
multimedia-based arts.

Software Development
Generative AI has totally changed the way we write code and
build software. With Generative AI tools such as GitHub Copilot,
ChatGPT, AlphaCode, we can write code much faster with fine
details.
Generative AI tools can assist developers by generating code
snippets, enhancing software testing efficiency by identifying
more defects, and suggesting optimal solutions to coding
challenges. This results in faster development cycles and higher
code quality, ultimately leading to improved software products
and enhanced user experiences.

Finance
Financial institutions are using generative AI to analyse market
trends, forecast stock movements with a high accuracy rate, and
refine trading strategies. The technology also helps us having
better risk assessment, fraud detection, and portfolio
optimization, leading to increased efficiency, reduced costs,
more profitability and better investment choices.

Example of Some Popular Generative AI Tools

In previous section, we talked about various use-cases of
generative AI. Now, let’s have a look into some of popular
generative AI tools available currently.

ChatGPT

ChatGPT is a conversational AI developed by OpenAI. It is

designed to engage in natural language conversations with users,
providing responses that are contextually relevant and coherent.
ChatGPT works by processing input text and generating
responses based on the patterns and relationships it has learned
from vast amounts of training data. It usages deep learning
techniques, specifically transformers, which allow it to
understand and generate human-like text.

GPT (Generative Pre-trained Transformer)

GPT is a transformer-based large language model, developed by

OpenAI. This is the engine behind ChatGPT.

The free version of ChatGPT is based on GPT 3.5, while the more
advanced GPT-4 based version, is provided to paid subscribers
under the commercial name “ChatGPT Plus”.

AlphaCode

AlphaCode is a transformer-based language mode, developed

by DeepMind. It is an AI-powered coding engine
that generates computer programs. AlphaCode is more complex
than many existing language models, with 41.4 billion
parameters.

The tool leverages deep learning algorithms to analyse huge

amounts of code and learn from patterns, enabling it to generate
optimized code solutions. It supports a wide range of
programming languages, including C#, Ruby, Python, Java, C++,
and more.
GitHub Copilot

GitHub Copilot is an AI-powered code completion tool developed

by GitHub in collaboration with OpenAI. It integrates directly into
code editors like Visual Studio Code and provides real-time
suggestions and completions for code as developers write.

It’s designed to assist developers by generating code snippets,

suggesting entire lines or blocks of code, and providing
contextual documentation. GitHub Copilot supports multiple
programming languages such as Python, JavaScript, Java, C++,
and more.

Bard

Bard is a conversational Generative AI chatbot developed by

Google, as a direct response to the swift rise of OpenAI’s
ChatGPT. Bard was initially based on LaMDA, a transformer-
based model. Later it got upgraded to other models such
as PaLM and Gemini.

Microsoft Copilot

Microsoft Copilot was initially launched by Microsoft in 2023 as

an AI-powered assistant that can help to browse the web. Later it
got rebranded to Microsoft Copilot.
Microsoft Copilot can be used to request summaries of articles,
books, news etc., general text and images, reformat text, update
images etc.

DALL-E

Developed by OpenAI, DALL-E (other versions are DALL-E2

and DALL-E3) is one of the best generative AI tools to generate
images. It uses deep learning algorithms to generate images from
texts.

StyleGAN

StyleGAN, developed by NVIDIA, is a generative model of type

GAN (Generative Adversarial Network), which is used to generate
high-quality synthetic images.

StyleGAN is extremely good in creation of realistic images of

human faces and other visual content. It can generate images of
human faces with a high degree of control over specific visual
features such as facial attributes, pose, and background.

Below are some images generated by StyleGAN that looks like a

real person. There is an interesting site https://this-person-does-
not-exist.com which demonstrates how StyleGAN can be used to
generate human faces which actually don’t exists.
Summary
In this step, we got a clear idea on what generative AI is and how it
is different from other AI types. We also touched upon various
real-life use-cases of generative AI and some popular generative
AI tools available.
Step 5 – Large Language Model

Where does Large Language Model fits into Generative

AI?
Let’s take an example of ChatGPT to understand it clearly. Out of
many capabilities of ChatGPT, one is to understand human
language (questions asked in plain English). It can also generate
response which we human can understand. This capability of
ChatGPT, to communicate with humans, is powered by — Large
Language Models.

In other words, we can say — A generative AI system which

needs to generate human-like text needs Large language
models.

Let’s break down it further in layman’s terms!

What is Language Model?

Let’s first understand what a language model is.

Language model is:

• a type of machine learning model

• which uses various statistical and probabilistic techniques

• to predict probability of a given sequence of words in a

sentence or phrase.
In simple words, language model is designed to predict next most
suitable word to fill in a blank space in a sentence or
phrase, based on the context of the given sentence/phrase.

Let’s take an example to understand better!

When we use messaging apps in phone, it helps us by predicting
the next word when we type in a message. For example, as soon
as we type “how,” the phone might suggest words like “are” or
“is” because it knows that those words often come after “how” in
sentences.

Similarly, if we type “I am going to,” the phone might predict

words like “store,” “park,” “office”, or “beach” because those are
common words that comes after “going” in everyday language.
This prediction is made based on the context of what we have
typed so far and the patterns it has learned from analyzing lots of
text.

Large Language Model (LLM)

A large language model (LLM) is a language model which is:

• a type of machine learning model

• that is trained on a large dataset of text

• and uses advanced neural network architectures

• to generate or predict human-like text.

Coming back to our earlier example, it is the language model that

helps AI tools to predict upcoming words in a sentence.
Below image summarizes important points about large language
model.

The most unique and powerful point about large language models
is their ability to generate human-like text, summarize, and
predict content based on vast amounts of data. LLMs
can process and analyze vast amounts of text data, making them
highly proficient in language processing tasks such as text
generation, summarization, translation, and sentiment analysis.

Natural Language Processing (NLP)

Natural Language Processing is an important concept very
much linked with LLM.
Natural Language Processing (NLP) is a subset of AI, which
focuses on the interaction between computers and humans
through natural language (say English).

• NLP refers to the process of enabling computers to

understand human language and communicate with us in
the same language.

• NLP uses algorithms to analyze, understand, and generate

human language.

• It also helps computers understand the context, and

sentiment behind words and sentences.

Let’s take another example to understand NLP better. Virtual

assistant, for example Siri, can understand and respond to our
commands using NLP.

Imagine you ask Siri, “Set an alarm for 7 AM tomorrow.”

• Siri’s NLP algorithms analyze the sentence, breaking it down

into individual words and understanding their meanings,
grammar, and context.

• The NLP algorithm will be able to understand the user’s

intent, which is to set an alarm.

• Further, Siri does the action specified in the command,

setting an alarm for 7 AM the following day on your device.
 • Finally, Siri will give a response in your language.

Natural Language Processing is the backbone for tasks such as

responding to human (e.g. ChatGPT), language translation,
search engines etc.

Natural Language Processing (NLP) and Large Language

Model (LLM)
Large Language Models may be considered as an evolution of
Natural Language Processing models. In other words, we can
say that a large language model is any model designed for NLP
tasks having focussed on understanding and generating human-
like text.

While NLP includes a broad range of models and techniques for

processing human language, LLMs focus on understanding and
generating human-like text. LLMs are specially designed to
predict the probability of a word or sentence based on the words
that come before it, allowing them to generate coherent and
contextually relevant text.

From machine learning technique point of view, natural language

processing uses a wide range of techniques, ranging from rule-
based methods to machine learning and deep learning
approaches.
On the other hand, large language model mainly uses deep
learning techniques to understand patterns and context in text
data to predict probability of next word in the sequence. LLMs are
designed based on artificial neural network architecture. Most of
the large language models are based on transformer-based
models.

How is Large Language Model related with Generative AI?

Large Language Model (LLM) are a subset of Generative

AI. While generative AI can generate many types of content such
as text, image, video, code, music etc., LLM is focussed on
generating text only.

Where/How Large Language Models are used?

Large Language Models (LLMs) are used in various AI applications

across different industries. Here are some major examples:

Virtual Assistants

LLMs models are the engine that power virtual assistants for
example Siri, Alexa, or Google Assistant. It’s the LLM models that
analyze the human command and interpret the meaning out of it,
helping these virtual assistants to perform several actions on
user’s behalf.

Chatbots
ChatGPT is not a new word anymore. Most of us have used it or
similar AI conversational chatbots. These chatbots uses large
language models to understand human questions and response
in a way that mimic human-language.

Language Translation

Large language models play an important role in language

translation done by AI tools such as Google Translate.
These models are trained on huge amount of multilingual text
data, which enable them to capture the subtle distinctions,
variations, context, and complexity of different languages.

When we asked translation tools to translate a sentence, it uses

the LLM algorithms to analyze the input text in one language and
generate an accurate and contextually appropriate translation in
the target language.

By considering the relationships between words and phrases in

both languages bidirectionally, LLMs can produce translations
that preserve the meaning and tone of the original text.

Text Generation

Now-a-days large language models are used in many

applications to generate human-like text. These models are so
sophisticated that they can generate coherent and contextually
relevant text based on a given prompt or input. LLM models can
be used to compose stories, generating product descriptions,
write emails and many more.

Summarization

Large language models are very useful for doing document

summarization. Using natural language processing capabilities,
LLM models can summarize lengthy documents or articles into
concise summaries while preserving the key information and
main points. Using techniques such as attention
mechanisms and contextual understanding, LLMs can determine
the most salient information to include in the summary, ensuring
that it captures the essence of the original text.

Sentiment Analysis

Sentiment analysis is a process to determine the sentiment or

emotional tone expressed in a text. Large language models can
be used to analyze huge amounts of text data, understand the
context, nuances, and tone of language, and identify sentiment
polarity (positive, negative, or neutral).

Many organizations now-a-days use large language models to

identify sentiments in text data coming from social media posts,
product reviews, customer feedback, news articles etc.

Content Recommendations
Large language models (LLMs) are being increasingly used by
platforms such as Netflix, YouTube, Amazon etc., for content
recommendations to provide users with more personalized and
relevant suggestions. These models capture the relationships
between words, phrases, and topics, allowing them to
understand the meaning and context of content. When it comes
to content recommendations, LLMs analyze a user’s interactions
with content, such as articles they’ve read, products they’ve
bought, or videos they’ve watched. Based on this data, LLMs can
predict what other content a user might be interested in and
suggest relevant options.

Some Popular Examples of Large Language Models

Here are some of the popular applications which uses large
language models.

GPT (Generative Pre-trained Transformers)

Generative Pre-trained Transformer is probably the most popular

large language model, which is used in ChatGPT. After the
introduction of transformer architecture in 2017, OpenAI released
GPT-1 as their first transformer based large language model in
2018. GPT-1 was initially trained on BookCorpus, a dataset
consists over 7000 self-published books.
Subsequently, OpenAI released more advanced version of GPT as
GPT-2, GPT-3, GPT-3.5 and GPT-4. All these are transformer-
based large language models. GPT-4 is a multimodal model,
which means it can take images as well as text as input.

BERT (Bidirectional Encoder Representations from

Transformers)

Introduced by Google in 2018, BERT is a transformer-based large

language model. BERT represents a significant advancement in
the field of large language model and natural language
processing. It’s a bidirectional transformer model which allows it
to process words in parallel, making it more efficient compared to
traditional sequential models like recurrent neural networks
(RNNs).

LaMDA (Language Model for Dialogue Applications)

LaMDA is conversational large language model, developed by

Google, which is also a transformer-based model. After the
sudden rise of ChatGPT, Google announced it’s own
conversational AI chatbot called “Bard”. Bard is powered by
LaMDA.

Later, Google introduced PaLM (Pathways Language Model), as

the successor of LaMDA. Further, in 2024, Google rebranded Bard
with the new name “Gemini”. Gemini is powered by large
language model (LLM) of the same name. Gemini multimodal
large language model is the successor to LaMDA and PaLM.

LLaMA (Large Language Model Meta AI)

LLaMA (Large Language Model Meta AI) is a set of large language

models (LLMs), introduced by Meta AI. LLaMA is an auto-
regressive language model, is built on the transformer
architecture.
Step 6 - Prompt Engineering

Introduction
Have you seen the movie I, Robot? If yes, you will immediately
understand the below image. For people who have not seen the
movie, let me brief it.

In the movie, detective Spooner (played by Will Smith) is trying to

investigate the death of his friend (and a scientist) Dr. Lanning.
Before his death, Dr. Lanning has created his holographic image
powered with AI which is supposed to help Spooner in finding
answers.

However, sometimes when Spooner ask a question, the

holographic image says, “I’m sorry! My responses are limited. You
must ask the right questions.”.

A scene from the movie I, Robot

What’s the relevance of this scene in our discussion?
Have you Interacted with AI tools such as ChatGPT, but didn’t get
the answer you were looking for? Or did you ever feel that the
answer provided by ChatGPT is not up to the mark?

If you don’t get a proper/expected response from an AI system,

for example ChatGPT, your first reaction would be that — The AI
system is not good enough!

However, the real problem could be that you don’t know how
to ask the right question or how to give right set of commands.

While interacting with ChatGPT, we need to know how to ask the

right questions and give precise instructions to it — That’s exactly
is Prompt Engineering!

In today’s world, where we see AI based systems everywhere,

prompt engineering has emerged as a game-changing technique
and is required to unlock the full potential of AI.
What exactly is a Prompt?
Prompts are the inputs or questions user gives to AI systems to
get a specific response.

For example, if you want ChatGPT to write a story on animals for

kids, you can use a prompt “Tell me a story which includes
animal characters. The story is targeted for kids.” as shown
below.

The AI systems (in this example ChatGPT) uses the prompt to

generate response (in this example a text response, a kid’s story
on animals). Depending on the type of AI system, the response
could be either text, or image or video or something else.

Prompts can be:

• sentences in plain English (or any other human language),

 • or code snippets

• or commands

• or any other combination of texts and code.

The generative AI program uses the prompt to understand what

kind of content you want it to create, and then it generates new
content based on that starting point.

Let’s take one more example. Below image shows a prompt used
by DALL-E to generate an image — “An astronaut riding a horse in
photorealistic style.”

A screenshot of DALL-E Web Page

The more specific and detailed your prompt is, the better the
AI can understand what you want it to create.
What is Prompt Engineering?
Prompt engineering:

• includes designing and optimizing prompts

• in a strategic manner

• to generate more accurate and desired response from AI

systems.

Instead of asking a general question, prompt engineering

involves providing specific instructions or context to get
better results.

For example, instead of asking ChatGPT a generic question such

as “Tell me about dogs,” you can use prompt engineering to get
more focused results. For instance, you can ask, “What are the
top 5 breeds of dogs known for their intelligence?”

By doing so, you’re guiding ChatGPT to give you a list of intelligent

dog breeds.

To summarize:

• With prompt engineering, you can tailor your questions,

making them more specific and structured.

• This way, AI systems (such as ChatGPT) can better

understand your intent and provide more accurate and
relevant answers.
How to use Prompt Engineering to get better results?
Let’s go deeper and understand how prompt engineering can help
us to get better results. To make it simple to understand, take
example of ChatGPT.

There are 3 important concepts in prompt engineering

— Specificity, Contextualization and Fine-tuning.

Specificity

Specificity in prompt engineering means being clear and

detailed in the instructions you give to the AI. Instead of asking a
broad question, you give specific details about what you want the
AI to do or talk about.

Let’s understand with below examples:

Non-specific Prompt: “Tell me about cars.”

Specific Prompt: “Can you describe the features of electric cars

compared to traditional gasoline cars?”

Being specific helps the AI understand exactly what you’re asking

for, so it can give you a better answer.

Contextualization

Contextualization in prompt engineering means giving the AI

model clear details and information about the situation or
task it’s being asked to do. It’s similar to providing a background
story or setting the scene for the AI. This helps the AI system
understand what it’s supposed to do and who it’s supposed to do
it for.

For example, if you want the AI to write a story about a birthday

party, you would provide contextualization by telling it things such
as who the birthday person is, where the party is happening, and
what kind of party it is (e.g., surprise party or themed party). This
helps the AI create a story that fits the context you’ve provided.

Let’s take another example:

Non-contextualized Prompt: “Write a review of this product.”

Contextualized Prompt: “Write a review of this product focusing

on its performance for outdoor activities.”

The contextualized prompt ensures that the generated review is

tailored to the specific use case and audience, improving its
relevance and usefulness.

Fine-tuning

Fine-tuning in prompt engineering involves iteratively adjusting

and refining the prompt based on the AI system’s output. It is an
ongoing process to optimize the prompts and guide AI system to
generate desired outcomes.
Fine-tuning is a process of trial and error. We keep adjusting your
prompt until you get the response you want.

Let’s understand it with an example.

Imagine you’re asking ChatGPT to write a short story about a dog.

Initial prompt: “Write a story about a dog.”

After getting the response, you might notice it’s too general or not
exactly what you wanted. This is where fine-tuning comes in. You
can adjust your prompt to give ChatGPT more guidance.

For example:

Initial prompt: “Write a story about a dog.”

Fine-tuned prompt: “Write a heartwarming story about a golden

retriever named Max who helps a little girl overcome her fear of
swimming.”

Fine-tuning is an iterative process. If the AI system’s response

still isn’t quite right, you can keep adjusting the prompt until you
get the desired outcome.

Some Examples of Good and Bad Prompts

Sure, here are some examples of good and bad prompts you can
try with ChatGPT.

Bad Prompt: “Write a short story”

Good prompt (being more specific): “Write a short story about a
detective solving a mysterious murder case.”

Explanation: The second prompt provides clear instructions and

sets the context for the desired output, guiding the ChatGPT to
generate a story focused on the specified theme and characters.

Bad Prompt: “Explain photosynthesis”.

Good prompt (providing detail information): “Explain the

process of photosynthesis in plants, including the role of
chlorophyll and sunlight.”

Explanation: The second prompt specifies the topic and includes

key details, helping the ChatGPT understand the specific
information required and produce a coherent and informative
response.

Bad Prompt: “What should I do today?”

Good Prompt: “Suggest some fun outdoor activities for a sunny

day.”

Explanation: The first prompt is too general and open-ended.

While the second one provides specific points and contexts.

How to Write Effective Prompts?

Here are some important points to keep in mind to write a clear

and effective prompt.
Be Clear and Specific

Ensure that your prompt clearly communicates the task or

question you want the AI system to address. Avoid ambiguity or
overly complex language that could confuse the AI.

Provide Context

Give enough context for the AI system to understand the problem

or topic it’s addressing. This helps the AI system generate more
relevant and useful responses.

Use Examples

If applicable, provide examples to illustrate what you’re asking

for. Examples can help the AI system understand the desired
output and provide more accurate responses.

Ask Specific Questions

Instead of vague prompts, ask specific questions that guide the AI

system toward the desired outcome.

Include Constraints

If there are any constraints or requirements for the response (e.g.,

word count limits, specific formats), make sure to include them in
the prompt. This helps the AI system generate responses that
meet your criteria.

Test and Iterate

Experiment with different prompts and observe how the AI system
responds. Adjust your prompts based on the results to improve
their effectiveness over time.

Focus on Clarity Over Creativity

While creativity can be beneficial in some cases, prioritize clarity

and effectiveness in your prompts. Clear and straightforward
prompts are more likely to produce the desired outcomes.

Step 7 - Ethical Considerations in Generative AI

What is ethical AI?

Ethical AI refers to use of artificial intelligence in a fair,
transparent, and responsible way. It involves treating everyone
equally, being clear about how AI decisions are made, and taking
responsibility for any errors. Ethical AI also includes protecting
people’s privacy, ensuring safety and reliability, and making sure
AI is accessible to all. It’s about using AI for good while
minimizing harm.

Key Principles of Ethical AI

There are some major principles involved in ethical AI. Let’s take
a look into them.
Fairness and Bias Mitigation

Ethical AI makes sure to prevent and mitigate bias in AI systems,

ensuring that they treat all individuals fairly and without
discrimination based on characteristics such as race, gender,
ethnicity, or socioeconomic status.

Transparency and Explainability

Explainable AI refers to the set of processes and methods that

allows human users to understand and trust the response
generated by AI systems. Ethical AI ensures that there is
transparency and explainability in AI systems. This enables users
to understand how AI-driven decisions are made.

AI transparency works hand in hand with explainable AI. AI

transparency helps ensure that all stakeholders can clearly
understand the workings of an AI system, including how it makes
decisions and processes data.

While explainability focuses on providing understandable

reasons for the decisions made by an AI system, transparency
involves being open about data handling, the model’s limitations,
potential biases, and the context of its usage.

Privacy and Data Protection

Ethical AI ensures the protection of individuals’ privacy and

personal data. It makes sure that AI systems collect, use, and
store data in a responsible and respectful manner, with
appropriate safeguards in place to prevent misuse or
unauthorized access.

Safety and Reliability

Ethical AI focuses on building AI systems that are safe, reliable,

and trustworthy, minimizing the risk of harm to individuals,
communities, and society at large. This includes ensuring
robustness against adversarial attacks and unforeseen
circumstances.

Inclusivity and Accessibility

Ethical AI promotes inclusivity and accessibility, ensuring that AI

technologies are designed to serve the needs of diverse
populations and that they do not exacerbate existing inequalities
or marginalize certain groups.

Ethical Concerns and Challenges with Generative AI

Generative AI can achieve remarkable tasks, like support drug

discovery and cancer diagnostics, create beautiful artwork and
videos, etc. However, due to lack of regulations, there are many
ways it can be misused as well. Like other forms of AI, generative
AI can cause a number of ethical issues and risks surrounding
data privacy, security, policies and workforces.

Let’s look into some of these concerns.

Copyright and Data Theft Issues

Generative AI can potentially cause copyright and data theft

issues in several ways:

Creation of Copyrighted Content

Generative AI can generate content, such as images, music, or

text, that closely resembles copyrighted material. If this
generated content is distributed or used without permission, it
could infringe on the original creator’s copyright.

Plagiarism

Content generated by AI could be used to plagiarize existing

works, such as academic papers, articles, or creative works. If AI-
generated content is passed off as original work without proper
attribution, it can lead to copyright infringement and academic
dishonesty.

Data Reuse and Replication

Generative AI models trained on datasets containing proprietary

or sensitive information may inadvertently generate content that
exposes confidential data. For example, text generators trained
on private chat logs or medical records could produce sensitive
information, leading to data breaches and privacy violations.

Forgery and Fraud

Generative AI can create realistic-looking images, videos, or
documents that mimic official or authenticated materials. This
could be exploited for forgery and fraud, such as creating fake
identification documents, counterfeit products, or deceptive
marketing materials.

Reverse Engineering

Generative AI models trained on copyrighted or proprietary data

may inadvertently reveal insights or patterns that could be
reverse engineered by competitors. This could lead to intellectual
property theft and unfair competition.

Harmful Content Distribution

Generative AI can contribute to the distribution of harmful

content in several ways:

Creation of Fake Content

Generative AI algorithms can produce highly realistic fake

images, videos, audio, and text. These creations can be used to
spread misinformation, fabricate evidence, or deceive individuals
and organizations.

Deepfakes

Deepfake technology, a specific application of generative AI,

allows for the manipulation of audiovisual content to make it
seem like someone said or did something they didn’t. This can be
used maliciously to create fake videos of public figures,
celebrities, or ordinary people engaging in inappropriate or
harmful behavior.

Automated Content Generation

Generative AI can automate the creation of large volumes of

content, such as spam emails, fake reviews, or malicious
messages. This can overwhelm online platforms and
communities with low-quality or harmful content, making it
difficult to distinguish between genuine and fake information.

Privacy Violations

Generative AI can generate synthetic images or videos that

resemble real individuals, potentially leading to privacy violations
if these creations are used without consent or for malicious
purposes, such as impersonation or defamation.

Generative AI can also be used to create surveillance footage or

tracking data that mimics real-life scenarios, enabling invasive
monitoring of individuals’ activities without their knowledge or
consent. This raises concerns about mass surveillance, stalking,
and other forms of privacy intrusion.
Summary
Generative AI has huge potential to completely change several
sectors, from healthcare to education, from gaming to
manufacturing, by creating new content and enhancing
productivity.

However, it also brings with it significant ethical concerns,

including the distribution of harmful content, copyright
infringements, data privacy violations, and many more. As we
continue to harness the power of Generative AI, it is very
important to ensure ethical best practices.

Hopefully generative AI regulations will soon be established by

governments. In the meantime, many companies are taking the
lead and developing their own ethical generative AI policies to
protect themselves and their customers. For example, SAP is
focusing on AI ethics to protect it’s customers and their data.
Step 8 - Challenges and Limitations in Generative AI
So far, we have learnt how generative AI and associated
technologies like LLM, NLP, etc. are changing the entire world. On
a high-level Generative AI seems to be an absolutely game
changer which can generate any text, image, video, audio, music
etc. like we human do and perhaps replace many types of jobs
that we human pursue today. However, in reality, there are still
some major challenges and limitations are there which needs to
be solved before generative AI becomes part of our daily life.
Let’s have a close look into some of these challenges and
limitations.

Lack of real creativity — Not able to think out of the box

We human are amazing at creativity. Its not just about creating a
new music or write a poetry or a new painting, its also about
altogether start a new genre of an art. For example, Pablo Picasso
amazed the world by his new style of painting by breaking down
objects into geometric shapes and presenting multiple
perspectives simultaneously. Similarly, Kool Herc, known as
“fathers of hip hop”, innovated a new genre of music by mixing
different techniques.
Generative AI is still has a long way to go before it reaches the
true level of creativity as we human have. Although generative AI
has capability to generate new ideas, music, poetries, stories, it
still goes by the defined rules and genres. It cannot altogether
start a new type of story telling techniques or start a new genre of
music.
Let’s simplify this with an example. If we train a generative model
to solve Rubik’s cube, it can generate different ways to solve the
cube. However, it may not propose the idea to unassembled the
Rubik’s cube into individual pieces and reassemble it again.
Intensive resource requirements for training generative
AI models
One of the significant limitations of generative AI is intensive
resource requirements for training the generative AI models.
Generative AI models, especially large-scale ones, for example,
OpenAI’s GPT, require extensive hardware and energy
consumption. Let’s break this down:

Training Time
Generative AI models are trained on huge datasets, often
consisting of millions or even billions of examples. The training
process can take weeks or even months, depending on the
complexity of the model and the available computational
resources.

Hardware Requirements
Training and running generative AI models require substantial
computational resources. It includes high-performance CPUs or
GPUs, memories, and other hardware components. These
resources are often expensive to acquire and maintain. It’s one of
the reasons that only big organizations are able to build the
generative AI models who can afford huge cost and other
resources.

Energy Consumption
Training generative AI models consumes a significant amount of
energy. This also leads to environmental concerns, particularly
the carbon emissions.
Ethical Considerations and Challenges
This is probably the most critical challenge we are facing with
generative AI. Due to lack of regulations, there are many ways
generative AI can be misused. Copyright and data theft Issues,
plagiarism, distribution of harmful content, deepfakes, identity
theft are some of the major ethical concern we have.

Challenges with Training Data

The training data plays a crucial role in generative AI. Any
generative model depends highly on quality and quantity of
training data. There are several challenges associated with
training data, such as:
Quality of Data
The quality of training data directly impacts the performance of
Generative AI models. If the training data is noisy, incomplete,
biased, or of poor quality, it can lead to the generation of
inaccurate or undesirable outputs by the AI model.
Quantity of Data
Generative AI models require large amounts of diverse training
data to learn effectively and generalize well to new scenarios.
However, obtaining and curating large datasets can be
challenging and resource-intensive, particularly for niche or
specialized domains.
Data Bias
Training data may contain biases present in the real-world data it
was collected from. If not properly addressed, these biases can
be perpetuated and amplified by Generative AI models, leading to
unfair or discriminatory outcomes in generated content.
Data Privacy and Security
Accessing and utilizing large datasets for training Generative AI
models can raise concerns about data privacy and security.
Ensuring compliance with data protection regulations and
safeguarding sensitive information is crucial to prevent
unauthorized access or misuse of data.
Technical Challenges
The architectures of generative AI models are often very complex,
comprising millions or even billions of parameters. Just to give
you some examples, OpenAI’s GPT-4 has more than 1 trillion
parameters. Even GPT-3 has around 175 billion parameters.
Google’s BERT has 110 million parameters. Handling this
technical complexity requires high expertise in machine learning
algorithms and programming skills.
Quality control and Human Intervention
The content generated by Generative AI may contain inaccurate
results. In industries where the impact of these generated
content is affecting direct human life, for example, healthcare,
legal services, it’s important to have a proper quality control and
human review in place.
For example, if a generative AI tool in healthcare generate a wrong
report based on the medical data of patient, the recommended
treatment should not be given to the patient. We still need
qualified doctors to review the generated report and provide their
confirmation.

Summary
With this final step, we have completed the . Hope you have a
crystal clear understanding of generative AI and all its related
concepts.