UNIT 4

-PART-

AT

Generative Al

AN

LEARNING OBJECTIVES

Supervised Leaming and Discriminative Modelling

Unsupervised Learning and Generative Modelling

Timeline of Generative Al

Characteristics of Generative Al

importance of Generative Al

Conventional Al and Generative Al

Types of Generative Al

GANPaint

Generative Al Tools

Benefits of Generative Al

Limitations of Generative Al

Ethical Considerations of Ising Generative Al

The Potential Negative mpact on Society

esponsible Use of nerative Al

ture Trends in nerative Al

Generative Al is a branch of artificial intelligence that uses intelligent models to create

various types of content, such as text, images, audio, video, and even code. Unlike
traditional artificial intelligence, which is concerned with using training data to process,
analyse, and then give similar outputs based on that data, generative Al focuses on creating
new and original content.

These models use algorithms and deep learning techniques to find patterns and structures in
existing data and generate new content similar to the input data.

Generative Al continuously learns and improves by analysing historical data, user

interactions, and feedback. This process helps it identify patterns and trends, enhancing its
understanding and problem-solving abilities. A popular example of Generative Al is
ChatGPT. ChatGPT generates responses to user queries and prompts based on the data it
has been trained on and the information provided by the user. Generative AI, like ChatGPT
and models such as DALL-E, is employed when requesting tasks such as suggesting names
for a flower shop or designing a logo for a company.

Fig. 4.1 Generative Al

Al Fact File!

DALL-E is a neural network model developed by OpenAl that generates images from textual
descriptions. It belongs to the class of generative Al models capable of creating visual
content based on the input provided.

Supervised Learning and Discriminative Modelling

In supervised learning, the model is trained using labelled data, where each data point has
an input (features) and a corresponding output (label). The algorithm learns to map these
inputs to their desired outputs based on the labelled examples it has been trained on.
Discriminative modelling specificall focuses on learning how to distinguish between different
classes or categories based on the features and characteristics present in the input data. It
aims to build a model that can accurately predica training examples. seen before, using what
it has learned from

Unsupervised Learning and Generative Modelling

In unsupervised learning, the model is trained using unlabelled data, it discovers patterns
and hidden Hructures. The model learns to represent the distribution of data by itself as it
has no specific oupur struct Generative Modelling focuses on generating new data points
that resemble the training data Using which they generate new images or text based on
learned patterns, allowing for the creation of unique and realistic outputs.

Fig. 4.2 Supervised learning

Timeline of Generative Al

Fig. 4.3 Unsupervised learning

Adobe Fyelly
2025 LLAMA 2023 Google Band

2011 IBMY Watson

Generative Adversarial Networks

2016-2019 OpenAl Five

2020 GPT

AVART

2023 Bing Chat 2003 GPT-4

Vartationsal Autoencoders VAEs

Tensorflow

2017 2018 Transformers GPT

2019 GPT-2

2022 Midjourney

2022 magen 2022

2016 Google DeepMind AlphaGo

Stable Diffusion

2022 CharGPT

Fig. 4.4 Generative Al timeline

In 2011, IBM's Watson won jeopardy, the first time a computer beat humans in a game.

In 2013, Variational Autoencoder algorithms were introduced in a paper titled

"Auto-Encoding Variational Bayes" by Diederik P. Kingma and Max Welling.

In 2014, Generative Adversarial Networks were developed by lan Goodfellow and his
colleagues June, 2014.

⚫ In 2015, TensorFlow was released by Google. TensorFlow is a free and open-source

software library which became a crucial tool for developers to create neural networks.

From 2016-2019, OpenAl Five was launched. It is a team of artificial intelligence agents
developed that plays the multiplayer online battle arena kurning from this was used to
develop ChatGPT
2016, Google's DeepMind AlphaGo beat the highest level players of Go

In 2017. Transformers were introduced in the paper "Attention is All You Need" by Vaswani.

2018. Chat hay most aderative Pre-trained Transformer) arbitere was Inroduced, which may
have been the most advanced Al model at that time.

jabsequently, newer models of ChatGPT were released in 2019, 2020, 2022, and 2023 with
updated architectures. 2022, there was a rise of designs Generative Al technologies, such
as DALL-E, CLIP, Imagen, and

Diffusion. These models made significant contributions to the field of artificial image
generation. in 2023, Bing Chat, Google Bardid LLAMA, and Adobe Firefly were launched,
each making

significant contributions to their field as advanced Al language models and creative tools. •
Bing Chat is an Al chatbot by Microsoft for information retrieval and task assistance.

• Google Bard is a Google-developed Al language model for generating coherent text and
storytelling. • LLAMA is a model for META (previously known as Facebook) specialising in
meta-learning across various tasks with minimal data.

Adobe Firefly is an Al tool for image and video creation and enhancement.

Al Review

Match the Advancements in Generative Al in column A to their respective descriptions in

column B.

Column A

Advancement in Al

1. Watson

2. Variational Autoencoder

3. Generative Adversarial Networks

4. TensorFlow

OpenAl Five

6. AlphaGo

7. Transformers
8. ChatGPT

9. DALL-E, CLIP, Diffusion

10. Bing Chat, Google Bard, LLAMA, Adobe Firefly

Column B

Description

a. DeepMind's Al program that defeated top Go players in 2016

b. Introduced in 2017, a neural network architecture for sequence modelling

c. Developed by lan Goodfellow in 2014, used for generating synthetic data

d. Released by Google in 2015, a crucial tool for neural network development

e. IBM'S Al system that won Jeopardy in 2011

f. Al agents developed by OpenAl for Dota 2, influencing ChatGPT development

g. Introduced in 2013, it learns a probabilistic latent space of input data

h. Based on GPT-3.5 architecture, introduced in 2018, a milestone in Al models

1. 2022 technologies contributing to artificial image generation

J. 2023 Al models known for language. and understanding creativity

Characteristics of Generative Al

Generative Al has various characteristics:

1. Content Creation: It generates new and unique content using learned patterns in text,
images, audio, video, and code.

2. Pattern Recognition: It identifies and learns from patterns found in large datasets, through
machine learning techniques

3. Continuous Learning: It continuously learns and improves 4. Data Analysis: It analyses

the historical data, user interactions and feedback.

5. Creativity: It produces creative and original content through advanced pattern recognition
and adaptation.

6. Vastliny: It is applicable across various domains and industries, from creative arts to
scientif research, adapting its methods to different contexts.
7. Scalability: It is capable of handling large datasets and complex tasks.

8. Adaptability: It adapts to new trends and industry information, adjusting its models and
outputs to reflect current standards and user preferences.

Al Fact File!

Generative Al has gained traction in recent times. Companies like Linkedin are using Al to
enhance job hunting, while Amazon is investing hundreds of millions of dollars in generative
Al startups.

Importance of Generative Al

Generative Al represents a significant advancement in the field of artificial intelligence, with

the potential to improve and impact various fields such as marketing, product development,
education, problem-solving automation, and user-machine interaction.

1. Content Creation: Generative Al's primary importance lies in content creation. We can use
it to generate innovative, original, and creative content in various mediums, such as text,
images, music, and videos.

2. Personalisation: Generative Al enhances the user experience by using their past data,
preferences, and history to create customised content tailored to their preferences and
needs. Streaming platforms like Netflix use it to provide recommendations based on the
user's viewing and search history.

Fig. 4.5 Streaming platforms

3. Productivity: It enhances productivity by automating the routine and repetitive tasks,

document processing, and streamlining workflows across various industries. It helps improve
one's ability to make decisions and work effectively.

Al Fact File!

Generative Al is being used by 73% of marketing departments across multiple businesses.

Generatie Al

20

Healthcare: Generative Al helps doctors analyse medical images, diagnose diseases, create
and improve the effectiveness and efficiency of treatments. uses
Baslacions to boost marketing efforts creates targeted content and personalised customer
teractions to beinventoring efforts. It also helps improve decision-making by analysing
demand, managing inventory, and making logistics more efficiend

Education: Generative le can be used to adapt lessons to fit student's needs, helping them
Hnderstand concepts more comprehensively. It also empowers teachers to create interactive
earning materials and automate administrative tasks, thereby enhancing classreflect and
improving student

Al Brainstorming!

List a few places, finnliga tiith where you have encountered Generative Al in your daily
routine, and compare your findings with those of your classmates to see if there are common
or different experiences with Generative Al.

• Conventional Al and Generative Al

Conventional AΙ

Conventional Al, also known as Narrow Al, is used for performing specific tasks by analysing
data it has been trained on previously. This type of Al uses techniques like machine learning,
where algorithms learn from large sets of labelled data to recognise patterns and make
predictions. For example, in image recognition, a conventional Al system could learn to
distinguish between different objects such as cats and dogs by analysing thousands of
labelled images. Another example is, in speech recognition, it can understand the user's
speech and convert it into text by recognising the patterns in audio data.

Conventional Al's main strength lies in its ability to operate within well-defined parameters
and rules, this makes it reliable for performing tasks that require precise analysis and
decision-making based on existing data patterns.

The applications of conventional Al are vast; ranging from recommendation systems that
suggest products based on user preferences to medical diagnostic tools that analyse
symptoms and predict potential illnesses. By leveraging its ability to identify and utilise
patterns, it enhances the efficiency and accuracy in various fields.

Generative Alises on creating new content, unlike conventional Al gwhich focuses on

analysing Existing data. This type of Al uses models trained on vast dates While convent
outputs in vario existing data. This type of ages, music, videos, and even codeutputs that are
creative Al relies on predefined rules, generative Al harnesses its learning to produce
outputs that are creative and original

similar to how a human creates things. One of the notable techniques used in generative Al
is Generative Adversarial Networks (GANs) One of the notable tephritions in various fields,
as discussed carliarticles. and des can use it Generative Al finds applove their writing style,
or even draft entire articles. Artists and designers can Employ generative Al to explore new
visual styles, generate artwork, or assist in architectural design Conventional Al vs
Generative Al
Conventional Al and Generative Al both have merits and demerits; there is no definitive way
to labe one approach as better than the other since each has its own use cases.
Conventional Al excels in problem-solving and accuracy based on data analysis, making it
the preferred choice for tasks such as fraud detection, recommendation systems, and
data-driven decision-making. In contrast, Generative Al is more versatile and creative,
suitable for applications that require the generation of new and unique content, such as
writing, art, and music. By utilising large datasets and advanced neural networks, Generative
Al can produce innovative outputs beyond the capabilities of Conventional Al Al models often
combine both approaches to create a well-rounded solution, capable of addressing diverse
user needs, integrating the analytical strengths of Conventional Al with the creative potential
of Generative Al.

Table 4.1 Conventional Al vs Generative Al

Focus and Output

Conventional Al

Analyses data to recognise patterns and make predictions within a specific domain

Generative Al

Creates new content based on learned patterns and user input, enhancing creativity and
producing unique outputs

Learning Approach

Learns from labelled data to recognise patterns

Learns from large datasets to generate new content

Scope

Narrow and restricted scope

Versatile and wide scope

Output Quality

Quality is generally consistent and controllable

Quality can vary and is difficult to control

Flexibility

Limited to predefined tasks and rules

Adaptable across various domains and creative tasks

User Involvement

Minimal user involvement after deployment

Requires significant user input to guide content creation

Data Dependency

Relies heavily on labelled data

Can use large amounts of uniabelled data and unstructured data

Applications

Generative A

Used in analytical applications such as recommendation systems, fraud detection, and

speech recognition

Aspect

Used in content creation applications such as writing, art generation, and music composition

Al Review

Which image in each pair is the real image, and which one is generated by Al?

1.

12

2.

Types of Generative Al

Generative Al can be classified in various ways, with the most common method based on
architecture and mechanisms. However, from a user's perspective, it is more practical to
classify Generative Al models by the type of content they generate text, images, music,
videos, or speech-as users are more concerned with the output rather than the underlying
technology.

Types of Generative AlX

On the Basis of Output

On the Basis of Architecture and Mechanism

Text

Image

Music

Video

Speech

Code

VAE'S

Transformers

Autoregressive Models

Fig. 4.6 Classification of generative Al

RNNG

Based on Output Type

Classification based on output type, focuses on the data generated by the Generative Al
models, such as text, images, music, videos, or speech. This classification is focused on
what is being generated.

1. Text Generation: Text Generation models are designed to understand and generate
human-like text. Models like GPT-3, GPT-4, and BERT excel at tasks, such as answering
queries, engaging in conversations, and creating text-based questions.

2. Image Generation: Image Generation Models are used to generate realistic or artistic
images, providing visual content. Models like DALL-E, MidJourney, and StyleGAN are
utilised to create artwork, graphic content, and realistic images. Tools such as GANPaint
enable users to edit and modify images using GAN technology.

3. Music Generation: Music generation models are used to compose music; they create
original compositions or improve upon existing ones. Models like OpenAl's MuseNet and
AIVA specialise in generating music of various styles and genres.

Fig. 4.7 Text generation with generative Al

4. Video Generation: Video generation models can generate or modify video content,
creating dynamic visual media. Generative Al models like Synthesia are used to create
Al-driven videos with avatars that can speak multiple languages based on text inputs.
Another example is DeepDream, which modifies existing videos using neural network-based
artistic enhancements.

5. Speech Generation: Speech generation models are used to produce human-like speech,
which help to enhance text-to-speech applications. Models like DeepMind's WaveNet
generate natural-sounding and highly intelligible speech from text. This has helped improve
human-machine interactions.

6. Code Generation: Code generation models are used to generate code snippets or entire
programs based on natural language prompts or code context. Models like GitHub Copilot,
OpenAl's Codex, and CodeBERT help developers by generating code suggestions,
auto-completing code, and debugging the program.

GitHub Copilot

Fig. 4.8 GitHub Copilot model

Based on Architecture and Mechanisms

Classification of Generative Al based on their underlying neural network architectures and

mechanisms focuses on how the data is generated by these models. This approach
emphasises the operational methods and generative processes of the models, rather than
the specific type of data they produce.

Before we learn more about the types of Generative Al on the basis of their architecture, we
must first understand what a Neural Network is.

A Neural Network is a type of computer program designed to recognise patterns and make
decisions, similar to how the human brain works. It is essentially a network made up of
neurons, these neurons are similar to brain cells and are used to process information. Each
neuron takes input, processes it, and passes the output to the next neuron.

Al Fact File!

In 2017, the album "I AM AI" was released, which was the first ever music album created by
Al This album was created using the Al software Amper Music.

network consists of various layers: the input layer, which takes the initial input in the form of
datak the hidden layers, which process the information and recognise patterns, and the
output ree which provides the final output. The data is processed through each of these
layers by passing ormation from one neuron to the next until it reaches the output layer.

Input Layer

Hidden Layer

Output Layer
TAN

Fig. 4.9 Architecture of an artificial neural network

Now that you have a fundamental understanding of what a neural network and neurons are,
let us discuss the types of generative Al on the basis of the neural network architectures and
mechanisms.

1. Generative Adversarial Networks (GANs): Generative Adversarial Network (or GAN) is a

type of machine learning model used in Generative AI, first introduced by Ian Goodfellow
and his colleagues in June 2014. This model comprises two neural networks: a generator
and a discriminator. These networks are pitted against each other to improve the overall
system. The primary objective of GANs is to generate synthetic data that resembles real
data, achieving this through a competitive process between the generator and discriminator.

• Generator: The task of the generator is to create synthetic data or content that could be
used in place of the real data, It learns to generate realistic data through feedback received
from the discriminator.

• Discriminator: The task of the discriminator is to

distinguish between synthetic data (produced by the generator) and real data (from the
training set). It is supposed to discriminate between the two to determine which outputs are
artificially manufactured. You performed a task similar to that of the discriminator in Activity
2, where you had to compare two images-one real and one fake.

GANs perform this process iteratively, with the generator producing fake data and improving
it in each iteration based on feedback from the discriminator, while the discriminator gets
better at distinguishing between real and fake data.

2. Variational Autoencoders (VAEs): Variational Autoencoders (VAEs) use an

encoder-decoder architecture to generate new data points by learning a probabilistic
mapping from input data to a latent space

Fig. 4.10 GAN

Generative Al
To illustrate this concept, consider the encoder as a translator that comes a fange data into a
simpler language known as the latent space. This encoding process provides trange of
interpretation rather than a single translation, allowing for multiple representations of the
data.

The latent space serves as a compressed representation where data is expressed in a

simplified format. By selecting different points in this space (or selecting different
translations), new versions for the data can be generated. This process is similar to altering
ingredients in a recipe to achieve varying tastes.
Conversely, the decoder functions as a reverse translator, converting representations from
the latent space back into the original format. This decoding process enables the VAE to
generate new data points that closely resemble the original input. VAEs are used to generate
photorealistic data, such as images, videos, and audio content.

3. Transformers: Transformer-based models use self-attention mechanisms to generate

sequences of data. This mechanism allows the model to evaluate the importance of each
word or token in relation to other words or tokens within the sentence, capturing
relationships and dependencies across the entire sequence. The primary objective of this
type of Generative Al is to generate contextual data that comprehends the relationships
between words or tokens in a sentence. This technology is widely applied in text generation,
content completion, translation, and has significantly advanced the field of Natural Language
Processing (NLP).

NLP

Fig. 4.11 Natural language processing

4. Recurrent Neural Networks (RNNs): RNNs utilise recurrent connections to generate

sequences of data. They take the output from one time step and feed it back into the
network as input for the next time step. This recursive feedback mechanism makes RNNs
well-suited for modelling time-series and sequential data, such as text and speech. Variants
like Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs) address issues
like long-term dependencies, enhancing their effectiveness in generating meaningful
sequences.

5. Autoregressive Models: Autoregressive models generate data using a step-by-step

approach, where each new data point is conditioned on the preceding ones, establishing a
sequential dependency. Examples include PixelRNN, which generates images pixel by pixel
and WaveNet, which produces high-fidelity audio by generating sound waveforms sample by
sample. These models excel in tasks where maintaining the sequential order and context of
data points is critical.

• GANPaint

GANPaint enables users to manipulate images by directly drawing on them. These drawings
trigger the GAN to generate realistic modifications by manipulating neurons with each brush
stroke. Users can add or remove objects such as trees, doors, and grass, with each object
corresponding to specific neural activations. Additionally, it also enables users to modify
attributes, such as adding windows to buildings.

GANPaint operates by selectively activating and deactivating neurons in a deep network that
correspond to objects and attributes within the image. This modification process generates a
new output image based on the user's interactions.

Generative Al

Generative Al Tools
Generative Al tools are software applications and platforms that utilise generative Al
techniques to create new content, solutions, or data. Some of the most popular Generative
Al tools at the moment include ChatGPT, DALL-E, and Google Gemini. These tools excel in
generating text, images, and engaging in human-like conversations. They demonstrate the
remarkable capabilities of Al in producing creative and contextually relevant content across
various media. Given here are short descriptions of some of the most popular Generative Al
tools.

1. ChatGPT: Developed by OpenAI, ChatGPT is a conversation Al model capable of

generating human-like responses across a wide range of applications. It excels in customer
support, answering queries, content creation, interactive conversations, and engaging in
interactive conversations.

2. DALL-E: Also developed by OpenAL, DALL-E is an Al model designed to generate unique

images from textual descriptions. It demonstrates the capability to create diverse and
intricate visual content based solely on written prompts.

3. Google Gemini (Formerly Known as Bard): Developed by Google, it is a conversational Al

that generates coherent and relevant text, enabling engaging and informative human-like
conversations.

Gemini

Fig. 4.12 Generative Al tools

Fig. 4.13 Google Gemini

4. GANPaint: Developed by the MIT-IBM Watson AI Lab in 2019, this tool utilises GAN
(Generative Adversarial Network) technology for image editing. It enables users to
manipulate and create scenes using advanced Al capabilities.

5. Midjourney: Developed by Midjourney Al, this tool generates high-quality images from
textual descriptions. Midjourney is renowned for its ability to produce visually stunning and
artistically inspired content based on descriptive input.

6. Github Copilot: Developed by Github in collaboration with OpenAI, it is a code completion

tool powered by advanced machine learning models. It assists developers by suggesting
code snippets and providing programming assistance to enhance efficiency and accuracy in
coding tasks.

Artbreeder: Developed by Joel Simon is a popular Al art application, used by artists to create
new images by leveraging various GAN models. Users can select and combine different
models to generate new and unique visual compositions.

8. RunwayML: RunwayML. is a platform designed to make machine learning models

accessible and usable for creatives, artists, and developers. It offers a user-friendly interface
that simplifies the development of generative models without requiring deep expertise in data
science or programming

9. Notion Al: Developed by Notion, this text-based generative tool combines various features
to enhance productivity in workspace, such as note-taking, databases management, and
project organisation.

10. Galileo.ai: Galileo.ai is a text-to-UI generative model designed to create user interface
designs based on textual prompts. This tool aims to boost the creativity and efficiency of
designers by automating the UI design process.

Ice-Breaker Activity 1

Visit the link: https://gemini.google.com/app. Use Gemini to learn how to write a prompt to
generate your desired output. Write the prompt you used and your findings. Design your
dream home. Write down the innovations that you would like to see in your future home.

Benefits of Generative Al

Generative Al has numerous benefits across different industries, transforming how we create
content and personalise user experiences. It automates tasks, boosts creativity, and
discovers innovative solutions, making it indispensable in today's competitive world. These
advantages have diverse applications, revolutionising businesses operations and customer
interactions, resulting productivity and efficiency. Some specific benefits efits include:
include: in increased

Content Generation

Generative Al is an invaluable tool for creating various forms of content, including text,
images, music, videos, and even code. It empowers users to generate creative, unique, and
novel content by suggesting ideas and producing outputs, thereby enhancing creativity and
efficiency. Additionally, it automates content creation processes, enabling users to
concentrate on strategic tasks while maintaining a consistent brand voice and style. For
example, marketers and sales specialists use generative Al for basic text creation and
copywriting.

Creativity Enhancement

Generative Al enhances creativity by generating innovative and unique content,

fundamentally changing how individuals express themselves. It facilitates the generation of
new ideas, provides suggestions to overcome creative blocks, and offers starting points for
further development. For example, a designer could use it to create initial website layouts or
receive suggestions for a brand logo, thereby making the creative process more efficient and
inspiring.

Personalisation
Generative Al enables the creation of content and recommendations tailored to the individual
needs and preferences of users and customers. This enhances user experience and
increase customer satisfaction through targeted content delivery. For example, booking
platforms can recommend events, shows, and movies based on user's past activities and
purchases. Additionally, voice assistants like Siri and Alexa personalise interactions by
learning users' likes and dislikes.

Frite

Cost Reduction and Time Savings

Generative Al significantly reduces operational costs and saves time by automating

repetitive tasks. this allows organisations to reallocate resources to more critical areas,
enhancing efficiency and maintaining a competitive edge. For example, a company could
automate tasks, such as generating customer support responses, enabling employees to
concentrate on strategic and high-priority initiatives.

Enhanced Problem-Solving Capabilities

Generative Al can enhance problem-solving capabilities by analysing complex scenarios,

simulating outcomes, and proposing innovative solutions that humans may not have
considered. This enhances the creativity and strategic thinking of users, enabling them to
explore new possibilities and solutions. For example, in marketing, it can brainstorm creative
campaign concepts and content strategies tailored to specific audiences, providing a
competitive edge in the market through innovative approaches,

ent ve se ed

Fig. 4,14 Benefits of Generative Al

Adaptive Learning

Generative Al continuously learns and adapts to new information, trends, and scenarios to
improve the quality of its outputs. It personalises the educational experience by adjusting
content and learning pathways based on real-time feedback and user interactions. This
customisation makes learning more accessible and comprehensible for students, catering to
their individual pace and understanding. For example, one student might benefit from more
illustrative examples, while another may prefer straightforward textbook definitions.

Limitations of Generative Al

Generative Al, despite its numerous advantages, possesses several limitations. Let us
understand each of these limitations in detail:

Data Dependency
Generative Al models heavily rely on the quantity and quality of data they are trained on. If
the training data is inaccurate, biased, incomplete or of low quality the outputs of the Al will
also be low quality. This dependency can lead to inaccuracies, biases, or inappropriate
content generation. For example, a generative Al trained on biased data may produce
outputs that exclude certain

ps.

DATA

DATA DATA

ATA

DATA

Fig.4.15 Dependency on data

Generative Af

Lack of Understanding

Generative Al models do not possess genuine understanding of the content they generate,
despite their advanced capabilities. They simply analyse patterns in data and produce
outputs, based on these patterns, but they lack the ability to understand the context and
meaning of what they generate. This can result in outputs that are inappropriate or
nonsensical. For example, a generative Al might create a grammatically correct sentence
that is contextually irrelevant to the topic.

Ethical Concerns

Generative Al raises several ethical concerns including intellectual property issues, and the
potential for misuse. Al-generated content can be used to infringe on copyright laws, create
deepfakes, and spread rumours and misinformation. These ethical concerns must be
addressed and carefully considered to prevent harm. For example, deepfake technology has
been used to create false information about popular public figures, damaging their
reputations.

High Computational Costs

Generative Al incurs significant computational costs, requiring substantial resources for both
training and operation. This can have environmental implications, as energy consumption
associated with these processes contributes to carbon emissions. Additionally, these costs
can pose a barrier to entry for smaller organisations. For example, the extensive
computational power required to train models like GPT-4 is costly and often not feasible for
all businesses.

Limited Creativity and Originality

Generative Al is known for its ability to generate creative outputs, but this creativity is
constrained by scope of its training data. These models generate content by mimicking
patterns in existing data to produce new outputs, but they may struggle to generate truly
novel or groundbreaking ideas. For example, an Al-generated artwork may resemble existing
styles but lack the unique creative vision and innovation of a human artist.

Dependence on Human Oversight

Generative Al is a valuable tool for content creation; however, it requires constant human
input and oversight to refine content to meet desired standards and avoid errors. Continuous
review, editing, and adjustment of the content are necessary. This dependency on human
intervention can diminish the benefits of automation and content generation. For example,
when generating an image of a tree, detailed input is essential, followed by ongoing
refinements to meet expectations. In some cases, the generated content may need to be
discarded if it fails to meet these standards.

Ethical Considerations of Using Generative Al

With the increasing popularity and advancement of generative Al, it is crucial to address and
resolve ethical concerns promptly.

The primary ethical issues of generative Al include authenticity, ownership, and misuse. One
significant implication is the potential for Al-generated content to deceive or manipulate the
public by blurring the lines between what is real and what is fake.

Fig. 4.16 Al ethics

‫هللا‬

this poses risks in areas such as misinformation, where Al can be used to create convincing
fake ws or forged documents. Additionally, there are concerns about intellectual property
rights and the ownership of Al-generated creations, raising questions about who holds
produced by machines. responsibility for content

eve these across various steel involves ensuring transparency, consent, and is deploytrisks
and to mis sectors, from art and entertainment to healthc. and accountability

gate these and to maximise benefits responsibly. Let us look at some of these ethical
implications and negative impacts of Generative Al in detail.

Bias and Discrimination: Generative Al models trained on biased datasets can amplify
existing societal biases, potentially leading to unfair outcomes or discrimination. For
example, biased language models may generate offensive or discriminatory content. It is
crucial to ensure that Al systems treat all users fairly, regardless of race, gender, or other
characteristics. Fairness considerations involve designing algorithms that mitigate biases
and promote equitable outcomes.
Privacy: Another major concern regarding Generative Al is its potential to use personal data
to create hyper-personalised content raising questions about data security. Protecting
individuals' identities and personal information becomes increasingly challenging as Al
techniques evolve.

Copyright and Legal Exposure: Generative Al tools trained on large datasets from various
sources, including the internet, may inadvertently generate content that infringes upon
copyright. Users may unknowingly face legal exposure if the generated content resembles
copyrighted material.

Lack of Accountability: The accountability associated with artificially generated content is

often unclear. In cases of copyright infringement or privacy violations, the lack of clear
ownership and responsibility can lead to legal disputes and uncertainty.

Misuse and Manipulation: Generative Al, while offering societal benefits, can also be
misused. The propagation of misinformation and fake content, facilitated by generative Al,
poses serious risks. This includes creating fake news, impersonating individuals through
deepfakes, and other forms of manipulation that can deceive the public and cause harm.

Al Fact File!

The term Deepfake has been derived from deep learning' and 'fake'. It refers to the use of
generative Al techniques to manipulate audiovisual content, typically to depict someone
saying or doing something they have not actually said or done.

→The Potential Negative Impact on Society

It is crucial to understand and learn more about the potential negative impacts of Generative
Al on society. These risks and impacts are causing major displacement in the society in the
present time and are likely to expand in the future. Let us examine some of these impacts.

• Facilitation of Plagiarism: Generative Al can generate content that mimics original works,
raising

concerns about intellectual property rights and authenticity.

Creation and Spread of Fake News and Misinformation: Generative AI can be exploited to
create and disseminate false information, influencing public opinion and potentially
destabilising societies.

Generative Al
Dependency on Al for Content Generation: Relying on Al models for content creation may
hinder individual creativity and innovation.

Job Displacement: The automation capabilities of Generative Al could lead to job

displacement as companies seek to reduce costs by using Al models to perform tasks
formerly done by hunans.
Privacy Violations: Generative Al's ability to process vast amounts of personal data raises
concerns about privacy rights and the potential for surveillance and exploitation of
individuals.

Responsible Use of Generative Al

There must be proper regulation and accountability to ensure a balance between innovation
and deployment. It is essential to enforce robust regulatory frameworks and mechanisms to
maintain accountability. Achieving this requires collective efforts from policymakers,
technology developers, and society at large. Here are a few ways in which generative Al can
be used responsibly.

Ethical guidelines and standards must be established to govern the development and
deployment of Generative Al technologies, along with the adoption of transparent practices.
This will help minimise the negative effects and potential harms of generative Al, while
maxirnising its advantages and fostering innovation in Al technologies, safeguarding
individuals' rights, privacy, and societal well-being

• It is crucial to use diverse and representative training data that is fair and unbiased.
Outputs generated by Al models must be scrutinised and periodically checked for bias and
misinformation. It is also important to ensure that the model does not generate hallucinated
or nonsensical outputs.

Al Fact File!

In generative Al, a hallucination occurs is when the model produces content that is not based
on facts or present in its training data. These hallucinations may happen in any form, such
as images, text or other creative outputs. For example, an Al model asked to generate an
image of a tiger may provide an image of a tiger with a trunk like an elephant or three ears.

It is crucial to maintain the privacy and consent of users and ensure that the rights of
creators are not violated when generating content. Proper attribution and clear guidelines on
ownership should be provided

Future Trends in Generative Al

Multimodal Generative Al

Generative Al has made significant strides in creating diverse types of content such as text,
video, images, music, and code. However, current models typically specialise in generating
one type of content at a time-text-based models generate text, image-based models create
visual content. Engineers are now advancing towards multimodal generative Al, which aims
to generate multiple types of content simultaneously. For example, the third version of the
text-to-image tool DALI-E can embed high-quality textual descriptions within its generated
images. This capability provides it with a competitive edge over other image-generating
models that lack the ability to incorporate textual context directly into their visual outputs.

DALL-E 3
Fig.4.17 DALL-E 3

Enhanced Creativity

One of the limitations of Generative Al is its restricted scope of creativity often constrained
by the patterns in its training data. However, as Generative Al advances, it holds the promise
of generating more creative and versatile content beyond the scope of its training data,
approaching the creativity of the human mind. This evolution could potentially approach the
levels of creativity seen in human minds, allowing Al to produce complex and innovative
content across various industries. For example, in the future, Al may have the capability to
create new genres of music by blending elements from different styles in novel ways that no
human has thought of.

Interactive Al

Mustafa Suleyman, co-founder of DeepMind and CEO of Microsoft Al, has proposed the
concept of interactive Ation or task con beyond generative Al. In this phas proposed the
concept simple conversation or partial task completion; they are capable of
autonomousleiend tasks either independently or by coordinating with other software tonis
pusly performing entire can ask bots to create cold emails, send them to the larges
audience, respond to any queries, follow up, and even schedule meetings for users. Another
emerging application is exemplified by tools like ChatGPT, which can currently assist in
writing and debugging code snippets. In the future, sach Al could potentially develop entire
applications with minimal user guidance, handling tasks from initial creation through testing
and ongoing maintenance.

Improved Natural Language Understanding The ability of generative Al models to

understand natural

language is continually advancing, enabling them to grasp the human language more
effectively. As Al continues to improve in this area, it will lead to enhancement in its ability to
comprehend and generate human-like text, facilitating more natural and intuitive
communication between machines and humans.

Ethically Responsible Al and the Open-Source Wave

There is a growing emphasis on developing ethical guidelines and frameworks to ensure the
responsible use of generative AL. This effort aims to address current challenges such as
bias, misinformation, and copyright issues. The goal is to establish standards that promote
fairness, transparency, and accountability in the development and deployment of Al
technologies. Furthermore, the open-source movement enhances transparency by enabling
broader participation in discussing and implementing ethical considerations in Al
development.

Generative Al in Healthcare and Biotechnology

Generative Al is increasingly being utilised in the healthcare and biotechnology industries,
enabling advancements in diagnostics, discovery of new drugs, creation of personalised
medicine, and patient care. Al-powered tools analyse medical images, predict patient
outcomes, optimise treatment plans, and accelerate the discovery of new drugs. These
innovations have the potential to save lives and reduce healthcare costs worldwide.

DeepMind

Fig. 4.18 Google DeepMind

Fig. 4.19 Natural language processing