Visit https://ebookmass.

com to download the full version and

browse more ebooks or textbooks

Introduction to Responsible AI: Implement Ethical

AI Using Python 1st Edition Manure

_____ Press the link below to begin your download _____

https://ebookmass.com/product/introduction-to-responsible-
ai-implement-ethical-ai-using-python-1st-edition-manure/

Access ebookmass.com now to download high-quality

ebooks or textbooks
 We believe these products will be a great fit for you. Click
the link to download now, or visit ebookmass.com
to discover even more!

Introduction to Datafication : Implement Datafication

Using AI and ML Algorithms Shivakumar R. Goniwada

https://ebookmass.com/product/introduction-to-datafication-implement-
datafication-using-ai-and-ml-algorithms-shivakumar-r-goniwada-2/

Introduction to Datafication: Implement Datafication Using

AI and ML Algorithms Shivakumar R. Goniwada

https://ebookmass.com/product/introduction-to-datafication-implement-
datafication-using-ai-and-ml-algorithms-shivakumar-r-goniwada/

Explainable AI Recipes: Implement Solutions to Model

Explainability and Interpretability with Python 1st
Edition Pradeepta Mishra
https://ebookmass.com/product/explainable-ai-recipes-implement-
solutions-to-model-explainability-and-interpretability-with-
python-1st-edition-pradeepta-mishra/

Explainable AI Recipes: Implement Solutions to Model

Explainability and Interpretability with Python 1st
Edition Pradeepta Mishra
https://ebookmass.com/product/explainable-ai-recipes-implement-
solutions-to-model-explainability-and-interpretability-with-
python-1st-edition-pradeepta-mishra-2/
Productionizing AI: How to Deliver AI B2B Solutions with
Cloud and Python 1st Edition Barry Walsh

https://ebookmass.com/product/productionizing-ai-how-to-deliver-
ai-b2b-solutions-with-cloud-and-python-1st-edition-barry-walsh/

Productionizing AI: How to Deliver AI B2B Solutions with

Cloud and Python 1st Edition Barry Walsh

https://ebookmass.com/product/productionizing-ai-how-to-deliver-
ai-b2b-solutions-with-cloud-and-python-1st-edition-barry-walsh-2/

Introduction to Prescriptive AI: A Primer for Decision

Intelligence Solutioning with Python Akshay Kulkarni

https://ebookmass.com/product/introduction-to-prescriptive-ai-a-
primer-for-decision-intelligence-solutioning-with-python-akshay-
kulkarni/

Introduction To Computing And Problem Solving Using Python

1st Edition E. Balaguruswamy

https://ebookmass.com/product/introduction-to-computing-and-problem-
solving-using-python-1st-edition-e-balaguruswamy/

The AI Book Chishti

https://ebookmass.com/product/the-ai-book-chishti/
Avinash Manure, Shaleen Bengani and Saravanan S

Introduction to Responsible AI
Implement Ethical AI Using Python
Avinash Manure
Bangalore, Karnataka, India

Shaleen Bengani
Kolkata, West Bengal, India

Saravanan S
Chennai, Tamil Nadu, India

ISBN 978-1-4842-9981-4 e-ISBN 978-1-4842-9982-1

https://doi.org/10.1007/978-1-4842-9982-1

© Avinash Manure, Shaleen Bengani, Saravanan S 2023

Apress Standard

The use of general descriptive names, registered names, trademarks,

service marks, etc. in this publication does not imply, even in the
absence of a specific statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general
use.

The publisher, the authors, and the editors are safe to assume that the
advice and information in this book are believed to be true and accurate
at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the
material contained herein or for any errors or omissions that may have
been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Apress imprint is published by the registered company APress

Media, LLC, part of Springer Nature.
The registered company address is: 1 New York Plaza, New York, NY
10004, U.S.A.
Any source code or other supplementary material referenced by the
author in this book is available to readers on GitHub
(github.com/apress). For more detailed information, please visit
https://www.apress.com/gp/services/source-code.
Table of Contents
Chapter 1:​Introduction
Brief Overview of AI and Its Potential
Foundations of AI:​From Concept to Reality
AI in Action:​A Multifaceted Landscape
The Promise of AI:​Unlocking Boundless Potential
Navigating the AI Frontier
Importance of Responsible AI
Ethics in the Age of AI:​The Call for Responsibility
Mitigating Bias and Discrimination:​Pioneering Fairness
and Equity
Privacy in the Age of Surveillance:​Balancing Innovation and
Security
Human-Centric Design:​Fostering Collaboration Between
Man and Machine
Ethics in AI Governance:​Navigating a Complex Landscape
Conclusion:​The Ongoing Dialogue of Responsibility
Core Ethical Principles
1.​Bias and Fairness:​Cornerstones of Responsible AI
2.​Transparency and Explainability
3.​Privacy and Security
4.​Robustness and Reliability
Conclusion
Chapter 2:​Bias and Fairness
Understanding Bias in Data and Models
Importance of Understanding Bias
How Bias Can Impact Decision-Making Processes
 Types of Bias
Examples of Real-world Cases Where Models Exhibited
Biased Behavior
Techniques to Detect and Mitigate Bias
Techniques to Detect Bias
Techniques to Mitigate Bias
Implementing Bias Detection and Fairness
Stage 1:​Data Bias
Dataset Details
Stage 2:​Model Bias
Conclusion
Chapter 3:​Transparency and Explainability
Transparency
Explainability
Importance of Transparency and Explainability in AI Models
Real-world Examples of the Impact of Transparent AI
Methods for Achieving Explainable AI
Explanation Methods for Interpretable Models:​Decision
Trees and Rule-Based Systems
Generating Feature Importance Scores and Local
Explanations
Tools, Frameworks, and Implementation of Transparency and
Explainability
Overview of Tools and Libraries for AI Model Transparency
Implementation of Explainable AI
Stage 1:​Model Building
Stage 2:​SHAP
 Stage 3:​LIME
Stage 4:​ELI5
Challenges and Solutions in Achieving Transparency and
Explainability
Addressing the “Black Box” Nature of AI Models
Balancing Model Performance and Explainability
Trade-offs between Model Complexity, Performance, and
Explainability
Conclusion
Chapter 4:​Privacy and Security
Privacy Concerns in AI
Potential Threats to Privacy
Privacy Attacks in AI Models
Mitigating Privacy Risks in AI
Security Concerns in AI
Potential Threats to Security
Mitigating Security Risks in AI
Conclusion
Chapter 5:​Robustness and Reliability
Concepts of Robustness and Reliability
Importance in AI Systems
Metrics for Measuring Robustness and Reliability
Challenges in Achieving Robustness
Sensitivity to Input Variations
Model Overfitting
Outliers and Noise
Transferability of Adversarial Examples
 Challenges in Ensuring Reliability
Data Quality
Model Drift
Uncertainty in AI Models
Conclusion
Chapter 6:​Conclusion
Summary of Key Findings
Role of Responsible AI in Business Adoption
Call to Action for Developers, Businesses, and Policymakers
Developers
Businesses
Policymakers
Final Thoughts
Future Outlook
Index
About the Authors
Avinash Manure
is a seasoned machine learning
professional with more than ten years of
experience in building, deploying, and
maintaining state-of-the-art machine
learning solutions across different
industries. He has more than six years of
experience in leading and mentoring
high-performance teams in developing
ML systems that cater to different
business requirements. He is proficient
in deploying complex machine learning
and statistical modeling algorithms and
techniques for identifying patterns and
extracting valuable insights for key stakeholders and organizational
leadership.
He is the author of Learn Tensorflow 2.0 and Introduction to
Prescriptive AI, both with Apress.
Avinash holds a bachelor’s degree in electronics engineering from
Mumbai University and earned his master’s in business administration
(marketing) from the University of Pune. He resides in Bangalore with
his wife and child. He enjoys traveling to new places and reading
motivational books.

Shaleen Bengani
is a machine learning engineer with more than four years of experience
in building, deploying, and managing cutting-edge machine learning
solutions across varied industries. He has developed several
frameworks and platforms that have significantly streamlined
processes and improved efficiency of machine learning teams.
Bengani has authored the book Operationalizing Machine Learning
Pipelines as well as multiple research papers in the deep learning space.
 He holds a bachelor’s degree in
computer science and engineering from
BITS Pilani, Dubai Campus, where he
was awarded the Director’s Medal for
outstanding all-around performance. In
his leisure time, he likes playing table
tennis and reading.

Saravanan S
is an AI engineer with more than six
years of experience in data science and
data engineering. He has developed
robust data pipelines for developing and
deploying advanced machine learning
models, generating insightful reports,
and ensuring cutting-edge solutions for
diverse industries.
Saravanan earned a master’s degree
in statistics from Loyola College from
Chennai. In his spare time, he likes
traveling, reading books, and playing
games.
About the Technical Reviewer
Akshay Kulkarni
is an AI and machine learning evangelist
and thought leader. He has consulted
with several Fortune 500 and global
enterprises to drive AI- and data
science–led strategic transformations.
He is a Google developer expert, author,
and regular speaker at major AI and data
science conferences (including Strata,
O’Reilly AI Conf, and GIDS). He is a
visiting faculty member for some of the
top graduate institutes in India. In 2019,
he was also featured as one of the top 40
under 40 data scientists in India. In his
spare time, he enjoys reading, writing,
coding, and building next-gen AI products.
© The Author(s), under exclusive license to APress Media, LLC, part of Springer
Nature 2023
A. Manure et al., Introduction to Responsible AI
https://doi.org/10.1007/978-1-4842-9982-1_1

1. Introduction
Avinash Manure1 , Shaleen Bengani2 and Saravanan S3
(1) Bangalore, Karnataka, India
(2) Kolkata, West Bengal, India
(3) Chennai, Tamil Nadu, India

In a world permeated by digital innovation, the rise of artificial

intelligence (AI) stands as one of the most remarkable advancements of
our era. AI, the simulated intelligence of machines capable of mimicking
human cognitive processes, has ignited a transformative wave that
spans across industries, from health care and finance to education and
entertainment. As the boundaries of AI continue to expand, so too does
its potential to reshape the very fabric of our society.
In this chapter, we shall embark on a journey to explore a concise
overview of AI and the vast potential it holds. Subsequently, we will
delve into the compelling reasons behind the significance of
responsible AI. In the end, we will cast our gaze upon the foundational
ethical principles that underpin the realm of responsible AI.

Brief Overview of AI and Its Potential

Artificial intelligence, once a realm of science fiction, has evolved into a
transformative force shaping our contemporary world. This
technological marvel, rooted in the emulation of human intelligence,
has unveiled an era of unprecedented possibilities. In this section, we
will delve into a succinct exploration of AI’s foundational concepts, its
diverse manifestations, and the remarkable potential it holds across
various domains.
Foundations of AI: From Concept to Reality
At its core, AI is an interdisciplinary domain that seeks to develop
machines capable of executing tasks that typically require human
intelligence. It encompasses a spectrum of technologies and techniques,
each contributing to the advancement of AI’s capabilities.
AI’s journey traces back to the mid-twentieth century, with pioneers
like Alan Turing laying the groundwork for the field’s theoretical
underpinnings. The development of early AI systems, often based on
symbolic reasoning, marked a significant step forward. These systems
aimed to replicate human thought processes through the manipulation
of symbols and rules.
However, it was the advent of machine learning that revolutionized
AI’s trajectory. Machine learning empowers computers to acquire
knowledge from data, allowing them to adjust and enhance their
performance over time. Neural networks, inspired by how human
brains work, enabled the emergence of revolutionary deep learning
technology, responsible for groundbreaking achievements in vision
(image recognition), speech (natural language processing), and more.

AI in Action: A Multifaceted Landscape

AI’s potential is vast and extends across a spectrum of applications,
each amplifying our ability to address complex challenges. One of AI’s
prominent manifestations is in the realm of data analysis. The ability of
AI algorithms to sift through vast datasets and extract meaningful
insights has revolutionized industries like finance, health care, and
marketing. For instance, financial institutions employ AI-powered
algorithms to detect fraudulent activities and predict market trends,
enhancing decision making and risk management.
AI’s prowess shines in its capacity for automation. Robotic process
automation (RPA) streamlines routine tasks, freeing human resources
for more strategic endeavors. Manufacturing, logistics, and customer
service have all witnessed the efficiency and precision AI-driven
automation can bring.
Another notable domain is natural language processing (NLP),
which empowers machines to comprehend and generate human
language. This technology finds applications in chatbots, language
translation, and sentiment analysis, transforming the way businesses
engage with customers and analyze textual data.
Health care, a sector perpetually seeking innovation, is experiencing
a revolution through AI. Diagnostic tools fueled by AI aid in the early
detection of diseases, while predictive analytics assist in identifying
outbreaks and planning resource allocation. The amalgamation of AI
with medical imaging is enhancing diagnostic accuracy, expediting
treatment decisions, and potentially saving lives.

The Promise of AI: Unlocking Boundless Potential

The potential of AI extends beyond incremental advancements; it
possesses the capacity to reshape industries, enhance our quality of life,
and address societal challenges. One such promise lies in autonomous
vehicles. AI-powered self-driving cars have the potential to reduce
accidents, optimize traffic flow, and redefine urban mobility.
In the realm of environmental conservation, AI plays a pivotal role.
Predictive models analyze complex climate data to anticipate natural
disasters, aiding in disaster preparedness and response. Additionally,
AI-driven precision agriculture optimizes crop yields, reduces resource
wastage, and contributes to sustainable food production.
Education, too, stands to benefit immensely from AI. Personalized
learning platforms leverage AI to adapt content to individual learning
styles, ensuring effective knowledge absorption. Moreover, AI-powered
tutoring systems provide students with immediate feedback, fostering a
deeper understanding of subjects.

Navigating the AI Frontier

As we stand on the precipice of the AI revolution, the horizon brims
with potential. From streamlining industries to revolutionizing health
care and empowering education, AI’s transformative influence is
undeniable. Yet, with its soaring capabilities comes the responsibility of
harnessing its potential ethically and responsibly, ensuring that
progress is accompanied by compassion, inclusivity, and accountability.
In the chapters that follow, we will delve deeper into the ethical
considerations and guiding principles that underpin the responsible
integration of AI into our lives.
Importance of Responsible AI
In the ever-evolving landscape of technology, AI emerges as a beacon of
innovation, promising to revolutionize industries, elevate human
capabilities, and redefine problem-solving paradigms. Yet, as AI takes
center stage, the imperative of responsibility looms larger than ever
before. In this exploration, we delve into the profound importance of
responsible AI, unraveling its ethical dimensions, societal implications,
and the critical role it plays in shaping a sustainable future.

Ethics in the Age of AI: The Call for Responsibility

As AI’s capabilities flourish, its potential to influence human lives,
societies, and economies becomes increasingly apparent. However, with
this potential comes an inherent ethical dilemma—the power to create
and wield machines capable of decision making, learning, and even
autonomy. Responsible AI emerges as the lodestar guiding the
development, deployment, and governance of AI technologies.
At its core, responsible AI calls for a deliberate alignment of
technological innovation with societal values. It beckons developers,
policymakers, and stakeholders to uphold ethical principles,
accountability, and transparency throughout the AI lifecycle. Its
significance transcends mere technology; it signifies a commitment to
safeguarding human well being and ensuring equitable benefits for all.

Mitigating Bias and Discrimination: Pioneering

Fairness and Equity
A glaring concern in the AI landscape is the potential for bias and
discrimination to be embedded in algorithms. AI systems trained on
biased data can perpetuate societal prejudices and exacerbate existing
inequalities. Responsible AI takes the mantle of addressing this issue
head-on, demanding rigorous data preprocessing, algorithmic
transparency, and the pursuit of fairness.
Through principled design and ethical considerations, responsible
AI strives to create systems that reflect the diverse fabric of human
society. It urges a concerted effort to bridge digital divides, ensuring
that AI’s impact is not marred by discriminatory practices. By
championing fairness and equity, responsible AI paves the way for a
future where technology is a tool of empowerment, rather than an
agent of division.

Privacy in the Age of Surveillance: Balancing

Innovation and Security
The era of digital advancement has resulted in an unparalleled rise in
data creation, raising worries about individual privacy and the security
of data. The insatiable appetite of AI for data necessitates a careful
equilibrium between creativity and the protection of individual rights
in its learning algorithms. Responsible AI highlights the significance of
safeguarding data by promoting strong encryption, secure storage, and
rigorous access management.
By championing responsible data-handling practices, responsible AI
cultivates a sense of trust between technology and individuals. It
empowers individuals to retain agency over their personal information
while enabling organizations to harness data insights for positive
transformations. Thus, it fortifies the pillars of privacy, ensuring that
technological advancement does not come at the cost of individual
autonomy.

Human-Centric Design: Fostering Collaboration

Between Man and Machine
Amidst the AI revolution, the concern that machines will supplant
human roles resonates strongly. Responsible AI dispels this notion by
embracing a human-centric approach to technology. It envisions AI as
an enabler, amplifying human capabilities, enhancing decision making,
and fostering innovative synergies between man and machine.
The importance of maintaining human oversight in AI systems
cannot be overstated. Responsible AI encourages the development of
“explainable AI,” wherein the decision-making processes of algorithms
are comprehensible and traceable. This not only engenders trust but
also empowers individuals to make informed choices, thereby ensuring
that AI operates as a benevolent ally rather than an enigmatic force.
Ethics in AI Governance: Navigating a Complex
Landscape
Responsible AI extends its purview beyond technology development
and encapsulates the intricate realm of governance and regulation. In
an era where AI systems traverse legal, social, and cultural boundaries,
ensuring ethical oversight becomes paramount. Responsible AI calls for
the establishment of robust frameworks, codes of conduct, and
regulatory mechanisms that govern the deployment of AI technologies.
The importance of responsible AI governance lies in its ability to
avert potential harms, address accountability, and align AI’s trajectory
with societal aspirations. It prevents a chaotic proliferation of
unchecked technology and ensures that AI is wielded for the collective
good, ushering in an era of collaborative progress.

Conclusion: The Ongoing Dialogue of Responsibility

As AI embarks on its transformative journey, the importance of
responsible AI remains steadfast. It reverberates through technological
corridors and resonates in ethical debates, reminding us of the
profound influence technology exerts on our lives. The responsibility of
shaping AI’s trajectory lies in our hands—developers, policymakers,
citizens alike—and requires a collective commitment to the tenets of
ethical innovation, societal benefit, and accountable stewardship.
In the sections that follow, we navigate deeper into the
multidimensional landscape of responsible AI. We unravel its core
principles, illuminate real-world applications, and scrutinize its
implications on diverse sectors. As we embark on this exploration, we
hold the torch of responsibility high, illuminating a path that aligns AI’s
capabilities with humanity’s shared vision for a just, equitable, and
ethically enriched future.

Core Ethical Principles

Responsible AI encapsulates a set of guiding principles that govern the
ethical development, deployment, and impact of AI technologies. These
principles (see Figure 1-1) serve as a compass by which to navigate the
intricate intersection of innovation and societal well being. In this
summary, we distill the essence of these core principles.

Figure 1-1 Evolution of artificial intelligence

1. Bias and Fairness: Cornerstones of Responsible

AI
In the realm of AI, the evolution from creativity to ethical obligation has
given rise to the notion of responsible AI. Among its guiding principles,
Bias and Fairness holds a greater urgency to be tackled than do the
others. With the growing integration of AI technologies into our daily
lives, assuring the absence of bias and the adherence to fairness
principles has risen as a crucial focal point. In this summary, we delve
into the intricacies of bias and fairness as foundational elements of
responsible AI, exploring their implications and challenges, and the
imperative of addressing them in the AI landscape.

Unveiling Bias: The Hidden Challenge

Bias, a deeply ingrained human tendency, can inadvertently seep into AI
systems through the data used to train them. AI algorithms learn
patterns and associations from vast datasets, which may inadvertently
contain biases present in human decisions and societal structures. This
can result in discriminatory outcomes, perpetuating stereotypes and
exacerbating social disparities.
Responsible AI acknowledges that eliminating all biases may be
unfeasible, but mitigating their impact is crucial. The focus shifts to
addressing glaring biases that lead to unjust or harmful consequences,
while also striving to ensure that AI systems promote equitable
treatment for all individuals.

Fairness as a North Star: Ethical Imperative

Fairness in AI underscores the creation of systems that treat all
individuals equitably, regardless of their background, demographics, or
characteristics. It transcends statistical definitions, delving into ethical
considerations to guarantee just and unbiased outcomes. Responsible
AI champions fairness as a moral and societal imperative, emphasizing
the need to redress historical and systemic inequities.
A critical facet of fairness is algorithmic fairness, which strives to
ensure that AI systems’ decisions are not influenced by sensitive
attributes such as race, gender, or socioeconomic status. Various
fairness metrics, algorithms, and techniques have emerged to assess
and rectify bias, promoting equitable outcomes and bolstering societal
trust in AI technologies.

The Challenge of Quantifying Fairness

The pursuit of fairness encounters challenges in quantification and
implementation. Defining a universally acceptable notion of fairness
remains elusive, as different contexts demand distinct definitions.
Striking a balance between competing notions of fairness poses a
significant challenge, with some approaches favoring equal treatment
while others prioritize addressing historical disparities.
Quantifying fairness introduces complexities, requiring the
calibration of algorithms to meet predefined fairness thresholds. The
trade-offs between different types of fairness can be intricate,
necessitating careful consideration of their implications for
marginalized groups and overall societal well being.

Mitigation Strategies and the Path Forward

Responsible AI advocates for proactive strategies to mitigate bias and
ensure fairness in AI systems, as follows:
Awareness and education play a pivotal role, fostering a deep
understanding of biases’ manifestation and their potential
consequences.
Data preprocessing techniques, such as re-sampling, re-weighting,
and augmentation, offer avenues to alleviate bias in training data.
Moreover, algorithmic interventions like adversarial training and
fairness-aware learning guide AI systems to produce fairer outcomes.
The incorporation of diversity in data collection, model
development, and evaluation reduces the risk of perpetuating biases.
We will dig deeper into different mitigation strategies in the coming
chapters.

Ethical Considerations and Societal Impact

Addressing bias and fostering fairness in AI transcends technical
algorithms; it delves into ethical considerations and societal impact.
Responsible AI obligates developers, stakeholders, and policymakers to
engage in an ongoing dialogue about the ethical dimensions of bias and
fairness. It prompts organizations to adopt comprehensive AI ethics
frameworks, infusing ethical considerations into the AI development
lifecycle.
 Societal implications underscore the urgency of addressing bias and
promoting fairness. Biased AI systems not only perpetuate existing
inequalities but can also erode trust in technology and exacerbate
social divisions. By championing fairness, responsible AI cultivates a
technological landscape that mirrors society’s aspiration for a just and
equitable future.

Conclusion: Toward Equitable Technological

Frontiers
In the pursuit of responsible AI, addressing bias and ensuring fairness
is not a mere checkbox; it is a transformative endeavor that demands
collaboration, ingenuity, and ethical conviction. As AI technologies
continue to reshape industries and touch countless lives, upholding the
principles of bias mitigation and fairness is an ethical imperative. The
path forward involves a multidisciplinary approach, where
technological innovation converges with ethical considerations, paving
the way for a future where AI fosters inclusivity, equity, and the
betterment of humanity as a whole.

2. Transparency and Explainability

In the realm of AI, where algorithms make decisions that impact our
lives, the principles of transparency and explainability emerge as
critical safeguards. These principles are integral components of
responsible AI, a framework designed to ensure ethical, fair, and
accountable AI development and deployment. In this summary, we
explore the significance of transparency and explainability as
cornerstones of responsible AI, delving into their implications,
challenges, and the transformative potential they offer.

Transparency: Illuminating the Black Box

Transparency in AI refers to the openness and comprehensibility of an
AI system’s decision-making process. It addresses the “black box”
nature of complex AI algorithms, where inputs and processes result in
outputs, without clear visibility into the reasoning behind those
outcomes. Responsible AI demands that developers and stakeholders
make AI systems transparent, enabling individuals to understand how
decisions are reached.
Transparency serves multiple purposes. It fosters accountability,
allowing developers to identify and rectify biases, errors, or unintended
consequences. It also empowers individuals affected by AI decisions to
challenge outcomes that seem unfair or discriminatory. Moreover,
transparency cultivates trust between AI systems and users, a crucial
element for widespread adoption.
However, achieving transparency is no trivial task. AI models often
consist of intricate layers and nonlinear transformations, making it
challenging to extract human-interpretable insights. Balancing the need
for transparency with the complexity of AI algorithms remains a
delicate endeavor.

Explainability: Bridging the Gap

Explainability complements transparency by providing insights into the
rationale behind AI decisions in a human-understandable manner.
While transparency reveals the overall decision-making process,
explainability delves into the specifics, unraveling the factors that
contributed to a particular outcome.
Explainability addresses the cognitive gap between the inherently
complex nature of AI processes and human cognition. It strives to
answer questions like “Why was this decision made?” and “How did the
algorithm arrive at this conclusion?” By translating AI outputs into
explanations that resonate with human reasoning, explainability
empowers users to trust and engage with AI systems more confidently.
However, achieving meaningful explainability is not without its
challenges. Striking a balance between simplicity and accuracy,
especially in complex AI models like deep neural networks, requires
innovative techniques that synthesize complex interactions into
interpretable insights.

Implications and Applications

The implications of transparency and explainability extend across a
spectrum of AI applications. In sectors like finance and health care,
where AI-driven decisions can have profound consequences,
transparency and explainability help stakeholders understand risk
assessments, diagnoses, and treatment recommendations. In the
criminal justice system, these principles can ensure that AI-driven
predictive models do not perpetuate racial or socioeconomic biases.
Furthermore, transparency and explainability are essential for
regulatory compliance. As governments and institutions craft AI
governance frameworks, having the ability to audit and verify AI
decisions becomes pivotal. Transparent and explainable AI systems
enable regulators to assess fairness, accuracy, and compliance with
legal and ethical standards.

Challenges and Future Directions

While the importance of transparency and explainability is widely
recognized, challenges persist, such as the following:
The trade-off between model complexity and interpretability
remains a fundamental conundrum. Developing techniques that
maintain accuracy while providing clear explanations is an ongoing
research frontier.
The dynamic nature of AI models also poses challenges.
Explainability should extend beyond initial model deployment to
cover model updates, adaptations, and fine-tuning. Ensuring
explanations remain accurate and meaningful throughout an AI
system’s lifecycle is a complex task.
Moreover, balancing transparency with proprietary considerations is
a delicate tightrope walk. Companies may be reluctant to reveal
proprietary algorithms or sensitive data, but striking a balance
between intellectual property protection and the public’s right to
transparency is imperative.

Conclusion
Transparency and explainability are not mere technical prerequisites
but rather essential pillars of responsible AI. They foster trust,
accountability, and informed decision making in an AI-driven world. By
shedding light on AI’s decision-making processes and bridging the gap
between algorithms and human understanding, transparency and
explainability lay the foundation for an ethical, fair, and inclusive AI
landscape. As AI continues to evolve, embracing these principles
ensures that the journey into the future is guided by clarity, integrity,
and empowerment.

3. Privacy and Security

In the age of rapid technological advancement, the integration of AI into
various facets of our lives brings with it a plethora of benefits and
challenges. As AI systems become more widely used and more complex,
the preservation of individual privacy and the assurance of data
security emerge as crucial facets within the scope of responsible AI.
This summary delves into the intricate interplay between privacy and
security, outlining their significance, implications, and the pivotal role
they play as core principles in the responsible development and
deployment of AI technologies.

Privacy in the Digital Age: A Precious Commodity

Privacy, a cornerstone of personal freedom, takes on new dimensions in
the digital era. As AI systems accumulate vast amounts of data for
analysis and decision making, the preservation of individuals’ rights to
privacy becomes paramount. Responsible AI recognizes the necessity of
preserving privacy as an inherent human entitlement, guaranteeing
that personal information is managed with the highest level of
consideration and reverence.
One of the key tenets of responsible AI is informed consent.
Individuals have the right to know how their data will be used and
shared, granting them the agency to make informed decisions.
Transparent communication between AI developers and users fosters a
sense of trust and empowers individuals to maintain control over their
personal information.
Furthermore, data minimization is a fundamental principle
underpinning responsible AI. It advocates for the collection, processing,
and retention of only that data essential for a specific AI task. This
approach minimizes the risk of unintended exposure and helps mitigate
potential breaches of privacy.

Data Security: Fortifying the Digital Fortress

The inseparable companion of privacy is data security. Responsible AI
recognizes that the data collected and utilized by AI systems is a
valuable asset, and safeguarding it against unauthorized access,
manipulation, or theft is imperative. Robust data security measures,
including encryption, access controls, and secure storage, form the
backbone of a trustworthy AI ecosystem.
Responsible AI developers must prioritize data protection
throughout the AI life cycle. From data collection and storage to data
sharing and disposal, security protocols must be rigorously
implemented. By fortifying the digital fortress, responsible AI
endeavors to shield sensitive information from malicious intent,
preserving the integrity of individuals’ identities and experiences.

Challenges and Opportunities

While privacy and security stand as cornerstones of responsible AI,
they also present intricate challenges that demand innovative solutions,
as follows:
The vast quantities of data collected by AI systems necessitate
sophisticated anonymization techniques to strip away personal
identifiers, ensuring that individuals’ privacy is upheld even in
aggregated datasets.
Additionally, the global nature of dataflows necessitates a
harmonized approach to privacy and security regulations.
Responsible AI advocates for the establishment of international
standards that guide data-handling practices, transcending
geographical boundaries and ensuring consistent protection for
individuals worldwide.
In the face of these challenges, responsible AI opens doors to
transformative opportunities. Privacy-preserving AI techniques, such as
federated learning and homomorphic encryption, empower AI systems
to learn and generate insights from decentralized data sources without
compromising individual privacy. These innovative approaches align
with the ethos of responsible AI, fostering both technological progress
and ethical integrity.
Trust and Beyond: The Nexus of Privacy, Security,
and Responsible AI
The interweaving of privacy and security within the fabric of
responsible AI extends far beyond technical considerations. It is an
embodiment of the ethical responsibility that AI developers and
stakeholders bear toward individuals whose data fuels the AI
ecosystem. By prioritizing privacy and security, responsible AI
cultivates trust between technology and humanity, reinforcing the
societal acceptance and adoption of AI technologies.
Responsible AI acknowledges that the preservation of privacy and
security is not a matter of mere regulatory compliance, but rather one
of ethical duty. It encompasses the commitment to treat data as a
stewardship, handling it with integrity and ensuring that it is leveraged
for the collective good rather than misused for unwarranted
surveillance or exploitation.
In conclusion, privacy and security emerge as inseparable twins
within the constellation of responsible AI principles. Their significance
extends beyond the technological realm, embodying the ethical
foundation upon which AI technologies stand. By embracing and
upholding these principles, responsible AI charts a path toward a future
where technological advancement and individual rights coexist
harmoniously, empowering society with the transformative potential of
AI while safeguarding the sanctity of privacy and data security.

4. Robustness and Reliability

In the ever evolving landscape of AI, ensuring the robustness and
reliability of AI systems stands as a paramount principle of responsible
AI. Robustness entails the ability of AI models to maintain performance
across diverse and challenging scenarios, while reliability demands the
consistent delivery of accurate outcomes. This summary delves into the
significance of these intertwined principles, their implications, and the
measures essential for their realization within AI systems.

Robustness: Weathering the Storms of Complexity

Robustness in AI embodies the capacity of algorithms to remain
effective and accurate amidst complexity, uncertainty, and adversarial
conditions. AI systems that lack robustness may falter when confronted
with novel situations, data variations, or deliberate attempts to deceive
them. A robust AI model can adeptly generalize from its training data to
novel, real-world scenarios, minimizing the risk of errors and biases
that could undermine its utility and trustworthiness.
The importance of robustness resonates across numerous domains.
In self-driving cars, a robust AI system should reliably navigate various
weather conditions, road layouts, and unexpected obstacles. In medical
diagnostics, robust AI models ensure consistent accuracy across diverse
patient profiles and medical settings. Addressing the challenges of
robustness is crucial to building AI systems that excel in real-world
complexity.

Reliability: A Pillar of Trust

Reliability complements robustness by emphasizing the consistent
delivery of accurate outcomes over time. A reliable AI system maintains
its performance not only under challenging conditions but also through
continuous operation. Users, stakeholders, and society as a whole rely
on AI systems for critical decisions, making reliability a foundational
element of trust.
Unreliable AI systems can lead to dire consequences. In sectors such
as finance, where AI aids in risk assessment and investment strategies,
an unreliable model could lead to substantial financial losses. In health
care, an unreliable diagnostic AI could compromise patient well being.
The pursuit of reliability ensures that AI consistently upholds its
performance standards, enabling users to confidently integrate AI-
driven insights into their decision-making processes.

Challenges and Mitigation Strategies

Achieving robustness and reliability in AI systems is no small feat, as
these principles intersect with multiple dimensions of AI development
and deployment, as follows:
Data Quality and Diversity: Robust AI requires diverse and
representative training data that encompass a wide array of
scenarios. Biased or incomplete data can undermine robustness.
Responsible AI emphasizes data quality assurance, unbiased
 sampling, and continuous monitoring to ensure that AI models learn
from a comprehensive dataset.
Adversarial Attacks: AI systems vulnerable to adversarial attacks can
make erroneous decisions when exposed to subtly altered input data.
Defending against such attacks involves robust training strategies,
adversarial training, and constant model evaluation to fortify AI
systems against potential vulnerabilities.
Transfer Learning and Generalization: The ability to generalize
knowledge from one domain to another is crucial for robustness. AI
developers employ transfer learning techniques to ensure that
models can adapt and perform well in new contexts without
extensive retraining.
Model Monitoring and Feedback Loops: To ensure reliability,
continuous monitoring of AI models in real-world scenarios is
imperative. Feedback loops allow models to adapt and improve based
on their performance, enhancing reliability over time.
Interpretable AI: Building AI systems that provide transparent
insights into their decision-making processes enhances both
robustness and reliability. Interpretable AI empowers users to
understand and trust AI-generated outcomes, fostering reliability in
complex decision domains.
Collaborative Ecosystems: The collaborative efforts of researchers,
developers, policymakers, and domain experts are vital for advancing
robust and reliable AI. Open dialogues, knowledge sharing, and
interdisciplinary cooperation facilitate the identification of
challenges and the development of effective mitigation strategies.

Conclusion: Building Bridges to Trustworthy AI

Robustness and reliability stand as pillars of responsible AI, nurturing
the growth of AI systems that can navigate complexity, deliver accurate
results, and engender trust. In the pursuit of these principles, AI
practitioners tread a path of continuous improvement, where
technological advancement intertwines with ethical considerations. As
AI takes on an increasingly pivotal role in our lives, the pursuit of
robustness and reliability ensures that it remains a tool of
empowerment, enhancing human endeavors across sectors while
safeguarding the foundations of trust and accountability.
Conclusion
In this chapter, we have provided a comprehensive introduction to the
world of AI, highlighting its immense potential to transform industries
and societies. We delved into the imperative need for responsible AI,
acknowledging that as AI’s influence grows, so too does the significance
of ensuring its ethical and moral dimensions. By examining the core
principles of responsible AI, such as fairness, transparency, security,
and reliability, we’ve underscored the essential framework for guiding
AI development and deployment.
It is clear that responsible AI is not merely an option, but also an
ethical obligation that ensures technology serves humanity in a just and
equitable manner. As we move forward, embracing responsible AI will
be pivotal in shaping a future where innovation and ethical
considerations harmoniously coexist. In the next chapters, we will
deep-dive into each of the core principles and showcase how they can
be achieved through sample examples with code walkthroughs.
© The Author(s), under exclusive license to APress Media, LLC, part of Springer Nature 2023
A. Manure et al., Introduction to Responsible AI
https://doi.org/10.1007/978-1-4842-9982-1_2

2. Bias and Fairness

Avinash Manure1 , Shaleen Bengani2 and Saravanan S3
(1) Bangalore, Karnataka, India
(2) Kolkata, West Bengal, India
(3) Chennai, Tamil Nadu, India

In the artificial intelligence (AI) landscape, bias’s impact on decisions is paramount. From individual
choices to complex models, bias distorts outcomes and fairness. Grasping bias’s nuances is essential for
equitable systems. It’s a complex interplay of data and beliefs with profound implications. Detecting and
mitigating bias is empowered by technology, nurturing transparent and responsible AI. This ongoing quest
aligns with ethics, sculpting AI that champions diversity and societal progress.
In this chapter, we delve into the intricate relationship between bias, fairness, and artificial
intelligence. We explore how bias can impact decision making across various domains, from individual
judgments to automated systems. Understanding the types and sources of bias helps us identify its
presence in data and models. We also delve into the importance of recognizing bias for creating fair and
equitable systems and how explainable AI aids in this process. Additionally, we touch on techniques to
detect, assess, and mitigate bias, as well as the trade-offs between model complexity and interpretability.
This comprehensive exploration equips us to navigate the complexities of bias and fairness in the AI
landscape, fostering ethical and inclusive AI systems.

Understanding Bias in Data and Models

Bias in data and models refers to the presence of systematic deviations that lead to inaccuracies or
unfairness in decision-making processes. It emerges when data-collection or model-construction
processes inadvertently favor certain groups, attributes, or perspectives over others. This bias can stem
from various sources, such as historical inequalities, flawed data-collection methods, or biased algorithms.
Addressing bias requires a deep understanding of its manifestations in both data and model outcomes,
along with the implementation of strategies that ensure equitable and unbiased decision making in
artificial intelligence systems.

Importance of Understanding Bias

Understanding bias is an indispensable cornerstone when striving to establish systems that are fair and
equitable, especially within the domain of artificial intelligence (AI) and machine learning (ML). Bias holds
the potential to instigate systemic inequalities, perpetuate discrimination, and reinforce social disparities.
Recognizing and comprehending its significance is paramount for fostering inclusivity, upholding ethical
practices, and ensuring that AI technologies make a positive contribution to society. Delving deeper, let’s
explore the profound importance of understanding bias for the creation of fair and equitable systems,
beginning with the following:
Avoiding Discrimination and Inequity: Bias, whether embedded within data or woven into models,
can be the catalyst for generating discriminatory outcomes. In instances where AI systems are crafted
without careful consideration of bias, they run the risk of disproportionately disadvantaging specific
groups, thereby perpetuating pre-existing inequalities. A profound understanding of the origins of bias
and its far-reaching implications empowers developers to embark on the journey of crafting systems
that treat all individuals impartially, irrespective of factors like background, gender, race, or any other
defining characteristic.
 Ensuring Ethical AI Deployment: Ethics and responsibility form the bedrock of AI development
endeavours. The comprehension of bias equips developers with the capability to align their work with
ethical principles and legal mandates. The essence of ethical AI lies in its determination to steer clear of
accentuating or prolonging societal biases. Instead, ethical AI strives ardently for fairness, transparency,
and accountability, serving as a beacon of responsible technological advancement.
Building Trust in AI Systems: The acceptability and trustworthiness of AI hinge upon its perceived
fairness and impartiality. An AI system that consistently generates biased outcomes erodes public trust
and undermines confidence in its efficacy. By proactively addressing bias in its myriad forms,
developers embark on a journey to construct systems that radiate trustworthiness and credibility,
reinforcing the belief that AI technologies are designed to function equitably.
Enhancing Decision-Making Processes: AI systems are progressively integrated into decision-making
processes that wield a tangible impact on individuals’ lives—be it hiring, lending, or criminal justice.
Bias within these systems can give rise to outcomes that are unjust and inequitable. Herein lies the
critical role of understanding bias: it lays the foundation for AI-driven decisions that are well informed,
transparent, and free from any semblance of discriminatory influence.
Promoting Innovation: Bias possesses the potential to shackle AI systems, limiting their efficacy and
applicability. A system tainted by bias may fail to accurately represent the diverse spectrum of human
experiences and perspectives. Addressing bias serves as a catalyst for innovation, creating an
environment conducive to the development of AI systems that are adaptive, versatile, and potent across
various contexts.
Reducing Reproduction of Historical Injustices: The shadows of historical biases and injustices can
unwittingly find their way into AI systems that learn from biased data. In this context, understanding
these latent biases proves instrumental. It empowers developers to take proactive measures, preventing
AI from inadvertently perpetuating negative historical patterns and detrimental stereotypes.
Encouraging Diversity and Inclusion: Understanding bias emerges as a driving force behind fostering
diversity and inclusion within the realm of AI development. By acknowledging biases and their potential
impact, developers take on the responsibility of ensuring that their teams are a microcosm of diversity,
ushering in an array of perspectives that contribute to more-comprehensive system design and
judicious decision making.
Contributing to Social Progress: AI possesses an immense potential to be a conduit of positive
transformation, capable of precipitating societal progress. Through the lens of addressing bias and
architecting fair systems, AI emerges as a tool that can bridge disparities, champion equal opportunities,
and propel social aspirations forward.
Long-Term Viability of AI: With AI poised to permeate diverse sectors, ranging from health care to
education to finance, the need for long-term viability and sustainable adoption becomes evident. This
enduring viability is anchored in the creation of AI technologies that are inherently equitable, acting as
catalysts for positive change and responsible technological advancement.
Understanding bias extends beyond theoretical recognition; it serves as a guiding beacon that informs
ethical practices, shapes technological landscapes, and steers the trajectory of AI’s contribution to society.

How Bias Can Impact Decision-Making Processes

Bias can have a profound impact on decision-making processes across various domains, from individual
judgments to complex automated systems. It can distort perceptions, influence choices, and lead to unjust
or discriminatory outcomes. Understanding how bias affects decision making is crucial for developing fair
and equitable systems. Here’s an in-depth explanation of how bias can impact decision-making processes:
Distorted Perceptions: Bias can alter how information is perceived and interpreted. When bias is
present, individuals may focus more on certain aspects of a situation while overlooking others. This can
lead to incomplete or skewed understandings, ultimately influencing the decisions made.
Unconscious Biases: Human decision making is influenced by unconscious biases, often referred to as
implicit biases. These biases stem from cultural, societal, and personal factors and can unconsciously
shape perceptions, attitudes, and judgments. Even well-intentioned individuals can be impacted by
these biases without realizing it.
Confirmation Bias: Confirmation bias occurs when individuals seek out or favor information that
confirms their existing beliefs or biases. This can result in decisions that are not well informed or
 balanced, as contradictory information may be ignored or dismissed.
Stereotyping: Bias can lead to stereotyping, where individuals make assumptions about a person or
group based on preconceived notions. Stereotyping can result in decisions that are unfair, as they are
based on generalizations rather than individual merits.
Unequal Treatment: Bias can lead to unequal treatment of different individuals or groups. This can
manifest in various ways, such as offering different opportunities, resources, or punishments based on
factors like race, gender, or socioeconomic status.
Discriminatory Outcomes: When bias influences decisions, it can lead to discriminatory outcomes.
Discrimination can occur at both individual and systemic levels, affecting people’s access to education,
employment, health care, and more.
Impact on Automated Systems: In automated decision-making systems, bias present in training data
can lead to biased predictions and recommendations. These systems may perpetuate existing biases and
further entrench inequality if not properly addressed.
Feedback Loops: Biased decisions can create feedback loops that perpetuate and amplify bias over
time. For example, if biased decisions lead to limited opportunities for a particular group, it can
reinforce negative stereotypes and further marginalize that group.
Erosion of Trust: When individuals perceive that decision-making processes are influenced by bias, it
erodes trust in those processes and the institutions responsible for them. This can lead to social unrest
and a breakdown of societal cohesion.
Reinforcing Inequalities: Bias in decision making can reinforce existing social inequalities. If certain
groups consistently face biased decisions, their opportunities and access to resources are limited,
perpetuating a cycle of disadvantage.

Types of Bias
Bias in machine learning refers to the presence of systematic and unfair errors in data or models that can
lead to inaccurate or unjust predictions, decisions, or outcomes. There are several types of bias that can
manifest in different stages of the machine learning pipeline (see Figure 2-1).

Figure 2-1 Types of bias

1. Data Bias: Data bias encompasses biases present in the data used to train and test machine learning
models. This bias can arise due to various reasons, such as the following:
Sampling Bias: When the collected data is not representative of the entire population, leading to
over- or under-representation of certain groups or attributes. For instance, in a medical diagnosis
dataset, if only one demographic group is represented, the model might perform poorly for
underrepresented groups.
Measurement Bias: Errors or inconsistencies introduced during data-collection or measurement
processes can introduce bias. For example, if a survey is conducted in a language not understood by
a specific community, their perspectives will be omitted, leading to biased conclusions.
Coverage Bias: Occurs when certain groups or perspectives are missing from the dataset. This can
result from biased data-collection methods, incomplete sampling, or systemic exclusion.

2. Model Bias: Model bias emerges from the learning algorithms’ reliance on biased data during
training, which can perpetuate and sometimes amplify biases, as follows:
 Representation Bias: This occurs when the features or attributes used for training
disproportionately favor certain groups. Models tend to learn from the biases present in the training
data, potentially leading to biased predictions.
Algorithmic Bias: Some machine learning algorithms inherently perpetuate biases. For example, if
a decision-tree algorithm learns to split data based on biased features, it will reflect those biases in
its predictions.
Feedback Loop Bias: When models’ predictions influence real-world decisions that subsequently
affect the data used for future training, a feedback loop is created. Biased predictions can perpetuate
over time, reinforcing existing biases.
3. Social Bias: Social bias pertains to the biases present in society that get reflected in data and models,
as follows:
Cultural Bias: Cultural norms, beliefs, and values can shape how data is collected and interpreted,
leading to biased outcomes.
Gender Bias: Historical and societal gender roles can result in unequal representation in datasets,
affecting model performance.
Racial Bias: Biased historical practices can lead to underrepresentation or misrepresentation of
racial groups in data, impacting model accuracy.
Economic Bias: Socioeconomic disparities can lead to differences in data availability and quality,
influencing model outcomes.

Understanding these types of bias is essential for developing strategies to detect, mitigate, and prevent
bias. Addressing bias involves a combination of careful data collection, preprocessing, algorithm selection,
and post-processing interventions. Techniques such as reweighting, resampling, and using fairness-aware
algorithms can help mitigate bias at various stages of model development.
However, ethical considerations play a crucial role in addressing bias. Being aware of the potential
impact of bias on decision-making processes and actively working to mitigate it aligns AI development
with principles of fairness, transparency, and accountability. By understanding the different types of bias,
stakeholders can work toward creating AI systems that promote equitable outcomes across diverse
contexts and populations.

Examples of Real-world Cases Where Models Exhibited Biased Behavior

Several real-world examples illustrate how machine learning models have exhibited biased behavior,
leading to unfair and discriminatory outcomes. These cases highlight the importance of addressing bias in
AI systems to avoid perpetuating inequality and to ensure ethical and equitable deployments. Here are
some detailed examples:
1. Amazon’s Gender-Biased Hiring Tool: In 2018, it was revealed that Amazon had developed an AI-
driven recruiting tool to help identify top job candidates. However, the system displayed a bias against
female applicants. This bias resulted from the training data, which predominantly consisted of
resumes submitted over a ten-year period, mostly from male candidates. As a result, the model
learned to favor male applicants and downgrade resumes that included terms associated with women.

2. Racial Bias in Criminal Risk Assessment: Several criminal risk assessment tools used in the
criminal justice system have been criticized for exhibiting racial bias. These tools predict the
likelihood of reoffending based on historical arrest and conviction data. However, the historical bias in
the data can lead to overestimating the risk for minority groups, leading to discriminatory sentencing
and parole decisions.

3. Google Photos’ Racist Labeling: In 2015, Google Photos’ auto-tagging feature was found to label
images of Black people as “gorillas.” This was a result of the model’s biased training data, which did
not include enough diverse examples of Black individuals. The incident highlighted the potential harm
of biased training data and the need for inclusive datasets.
4. Biased Loan Approval Models: Machine learning models used for loan approval have shown bias in
favor of certain demographic groups. Some models have unfairly denied loans to minority applicants
or offered them higher interest rates, reflecting historical biases in lending data.

5. Facial Recognition and Racial Bias: Facial recognition systems have been criticized for their racial
bias, where they are more likely to misidentify people with darker skin tones, particularly women.
This bias can result in inaccurate surveillance, racial profiling, and infringement of civil rights.

These real-world examples underscore the urgency of addressing bias in AI systems. To prevent such
biased behavior, it’s crucial to carefully curate diverse and representative training data, use fairness-aware
algorithms, implement bias detection and mitigation techniques, and continuously monitor and evaluate
model outputs for fairness. By proactively addressing bias, developers can ensure that AI systems
contribute positively to society and uphold ethical standards.

Techniques to Detect and Mitigate Bias

Detecting and mitigating bias in machine learning models and data is essential to create fair and equitable
AI systems. Let’s look at some techniques to identify and address bias (Figure 2-2).

Figure 2-2 Techniques to detect and mitigate bias

Techniques to Detect Bias

Bias-detection techniques are essential tools for identifying and quantifying biases present in data,
models, and their outputs. These techniques help ensure that AI systems are fair, equitable, and free from
discriminatory tendencies. Here’s a detailed explanation of various bias-detection techniques:
Exploratory Data Analysis (EDA): EDA involves analyzing the distribution and characteristics of data
to identify potential sources of bias. By visualizing data distributions and exploring patterns across
different groups or attributes, data scientists can spot disparities that might indicate bias.
Fairness Metrics: Fairness metrics quantify and measure bias in machine learning models’ predictions.
Common fairness metrics include disparate impact, equal opportunity difference, and statistical parity
difference. These metrics compare outcomes between different groups to determine if there’s an unfair
advantage or disadvantage.
Benchmark Datasets: Benchmark datasets are designed to expose bias in machine learning models.
They contain examples where fairness issues are intentionally present, making them useful for
evaluating how well models handle bias.
Group Disparity Analysis: Group disparity analysis compares outcomes for different groups across
various attributes. By calculating differences in outcomes, such as acceptance rates, loan approvals, or
hiring decisions, developers can identify disparities that indicate bias.
Sensitivity Analysis: Sensitivity analysis involves testing how small changes in data or model inputs
impact outcomes. This can reveal how sensitive predictions are to variations in the input, helping
identify which features contribute most to biased outcomes.
Adversarial Testing: Adversarial testing involves deliberately introducing biased data or biased inputs
to observe how models respond. By observing how models react to these adversarial inputs, developers
can gauge their susceptibility to bias.
Real-world Performance Analysis: Deployed models can be monitored in real-world settings to assess
whether they generate biased outcomes in practice. Continuous monitoring allows developers to detect
emerging bias patterns over time.
Proxy Variable Analysis: Proxy variables are attributes correlated with protected characteristics (e.g.,
gender, race). Analyzing how strongly proxy variables affect model outcomes can indicate the presence
of hidden bias.
Interpretability Techniques: Interpretability techniques, like feature importance analysis, can help
understand which features contribute most to model predictions. Biased features that contribute
disproportionately might indicate bias.
Human Evaluation and Feedback: Involving human evaluators from diverse backgrounds to review
model outputs and provide feedback can help identify bias that might not be apparent through
automated techniques.
Fairness Audits: Fairness audits involve a comprehensive review of data collection, preprocessing, and
model development processes to identify potential sources of bias.
Synthetic Testing Scenarios: Creating controlled scenarios with synthetic data can help simulate
potential bias sources to observe their impact on model predictions.

Techniques to Mitigate Bias

Mitigating bias in machine learning models is a critical step to ensure fairness and equitable outcomes.
There are various strategies and techniques that can be employed to reduce bias and promote fairness in
AI systems. Here’s a detailed explanation of mitigation bias strategies:
1. Resampling: Balancing class representation by either oversampling underrepresented groups or
undersampling overrepresented ones can help reduce bias present in the data.

2. Reweighting: Assigning different weights to different classes or samples can adjust the model’s
learning process to address imbalances.

3. Fairness-Aware Algorithms:
Adversarial Debiasing: Incorporates an additional adversarial network to reduce bias while
training the main model, forcing it to disregard features correlated with bias.
Equalized Odds: Adjusts model thresholds to ensure equal opportunity for positive outcomes
across different groups.
 Reject Option Classification: Allows the model to decline to make a prediction when uncertainty
about its fairness exists.

4. Regularization Techniques:
Fairness Constraints: Adding fairness constraints to the model’s optimization process to ensure
predictions are within acceptable fairness bounds.
Lagrangian Relaxation: Balancing fairness and accuracy trade-offs by introducing Lagrange
multipliers during optimization.

5. Post-processing Interventions:
Calibration: Adjusting model predictions to align with desired fairness criteria while maintaining
overall accuracy.
Reranking: Reordering model predictions to promote fairness without significantly compromising
accuracy.

6. Preprocessing Interventions:
Data Augmentation: Adding synthesized data points to underrepresented groups to improve
model performance and reduce bias.
De-biasing Data Preprocessing: Using techniques like reweighting or resampling during data
preprocessing to mitigate bias before training.

7. Fair Feature Engineering: Creating or selecting features that are less correlated with bias, which
can help the model focus on relevant and fair attributes.

8. Ensemble Methods: Combining multiple models that are trained with different strategies can help
mitigate bias, as biases in individual models are less likely to coincide.

9. Regular Monitoring and Updates: Continuously monitoring model performance for bias in real-
world scenarios and updating the model as new data becomes available to ensure ongoing fairness.

10. Ethical and Inclusive Design: Prioritizing diverse representation and ethical considerations in data
collection, preprocessing, and model development to prevent bias from entering the system.

11. Collaborative Development: Involving stakeholders from diverse backgrounds, including ethicists
and affected communities, to collaboratively address bias and ensure that mitigation strategies align
with ethical values.

12. Transparency and Communication: Being transparent about the steps taken to mitigate bias and
communicating these efforts to users and stakeholders to build trust in the system.

13. Legal and Regulatory Compliance: Ensuring that the AI system adheres to relevant laws and
regulations concerning discrimination and bias, and actively working to comply with them.

Implementing Bias Detection and Fairness

The purpose of this exercise was to start exploring bias, including potential methods to reduce bias as well
as how bias may very easily become exacerbated in ML models. In this exercise, for the sake of brevity,
Let’s check for bias in relation to race, though bias should be checked against the other protected classes
as well.

Stage 1: Data Bias

In this task, we begin by training a model on the original dataset. However, upon analyzing and visualizing
the data, we detect the presence of racial bias. To address this issue, we implement a resampling technique
to promote a fair and unbiased representation.
Resampling is a technique used in machine learning to create new training data by altering the original
data. It can include both oversampling and undersampling and aims to create a more balanced and
representative training dataset, which helps models learn more effectively.
We then retrain the model on the balanced data and evaluate the accuracy. This process helps to
mitigate race bias and ensure fair prediction.

Dataset Details
We have used an individual’s annual income results from various factors. Intuitively, it is influenced by the
individual’s education level, age, gender, occupation, etc.
Source: https://archive.ics.uci.edu/dataset/2/adult
The dataset contains the following 16 columns:
Age: Continuous
Workclass: Private, Self-emp-not-inc, Self-emp-inc, Federal-gov, Local-gov, State-gov, Without-pay,
Never-worked
Fnlwgt: Continuous
Education: Bachelor’s, Some-college, 11th, HS-grad, Prof-school, Assoc-acdm, Assoc-voc, 9th, 7th-8th,
12th, Master’s, 1st-4th, 10th, Doctorate, 5th-6th, Preschool
Marital Status: Married-civ-spouse, Divorced, Never-married, Separated, Widowed, Married-spouse-
absent, Married-AF-spouse
Occupation: Tech-support, Craft-repair, Other-service, Sales, Exec-managerial, Prof-specialty, Handlers-
cleaners, Machine-op-inspect, Adm-clerical, Farming-fishing, Transport-moving, Priv-house-serv,
Protective-serv, Armed-Forces
Relationship: Wife, Own-child, Husband, Not-in-family, Other-relative, Unmarried
Race: White, Asian-Pac-Islander, Amer-Indian-Eskimo, Other, Black
Sex: Female, Male
Capital Gain: Continuous
Capital Loss: Continuous
Hours-per-week: Continuous
Native Country: United-States, Cambodia, England, Puerto-Rico, Canada, Germany, Outlying-US, India,
Japan, Greece, South, China, Cuba, Iran, Honduras, Philippines, Italy, Poland, Jamaica, Vietnam, Mexico,
Portugal, Ireland, France, Dominican-Republic, Laos, Ecuador, Taiwan, Haiti, Columbia, Hungary,
Guatemala, Nicaragua, Scotland, Thailand, Yugoslavia, El-Salvador, Trinadad&Tobago, Peru, Hong,
Holand-Netherlands
Income (>50k or <=50k): Target variable

Getting Started
The following is the process for implementing data bias detection and mitigation process in Python.

Step 1: Importing Packages

The following shows how to import all the necessary packages:

[In]:
# Import necessary libraries
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score, classification_report,
confusion_matrix
from sklearn.utils import resample
from sklearn.preprocessing import LabelEncoder, StandardScaler
from sklearn.metrics import classification_report
Step 2: Loading the Data
[In]:
# Read the dataset into a pandas DataFrame
df = pd.read_csv(" Income.csv")

Step 3: Checking the Data Characteristics

Check if there are any discrepancies in the data, like missing values, wrong data types, etc.:

[In]:
# Display basic information about the dataset
df.info()

[Out]:
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 48842 entries, 0 to 48841
Data columns (total 15 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 age 48842 non-null int64
1 workclass 48842 non-null int32
2 fnlwgt 48842 non-null int64
3 education 48842 non-null int32
4 education-num 48842 non-null int64
5 marital-status 48842 non-null int32
6 occupation 48842 non-null int32
7 relationship 48842 non-null int32
8 race 48842 non-null int32
9 sex 48842 non-null int32
10 capital-gain 48842 non-null int64
11 capital-loss 48842 non-null int64
12 hours-per-week 48842 non-null int64
13 native-country 48842 non-null int32
14 income 48842 non-null int32
dtypes: int32(9), int64(6)
memory usage: 3.9 MB

There are no null values present in the data, so we can proceed with the data preprocessing steps.

Step 4: Data Preprocessing

Create a list of columns with categorical columns to be encoded:

[In]:
# Define a list of categorical columns to be encoded and perform label
encoding for categorical columns
categorical_columns = ['sex', 'race', 'education', 'marital-status',
'occupation', 'relationship', 'native-country', 'workclass', 'income']
label_encoders = {}
for column in categorical_columns:
label_encoders[column] = LabelEncoder()
df[column] = label_encoders[column].fit_transform(df[column])

Categorical columns contain multiple categorical values. To use these categorical values ​for model
building, apply dummy variable-creation techniques to columns having more than two unique values.
[In]:
# Perform one-hot encoding for columns with more than 2 categories
get_dummies = []
label_encoding = []
for i in categorical_columns:
print('Column Name:', i, ', Unique Value Counts:', len(df[i].unique()),
', Values:', df[i].unique())
if len(df[i].unique()) > 2:
get_dummies.append(i)
else:
label_encoding.append(i)
df = pd.get_dummies(df, prefix=get_dummies, columns=get_dummies)

[Out]:
Column Name: sex, Unique Value Counts: 2, Values: [1 0]
Column Name: race, Unique Value Counts: 2, Values: [1 0]
Column Name: education, Unique Value Counts: 16, Values: [ 9 11 1 12 6
15 7 8 5 10 14 4 0 3 13 2]
Column Name: marital-status, Unique Value Counts: 7, Values: [4 2 0 3 5 1
6]
Column Name: occupation, Unique Value Counts: 15, Values: [ 1 4 6 10 8
12 3 14 5 7 13 0 11 2 9]
Column Name: relationship, Unique Value Counts: 6, Values: [1 0 5 3 4 2]
Column Name: native-country, Unique Value Counts: 42, Values: [39 5 23
19 0 26 35 33 16 9 2 11 20 30 22 31 4 1 37 7 25 36 14 32
6 8 10 13 3 24 41 29 28 34 38 12 27 40 17 21 18 15]
Column Name: workclass, Unique Value Counts: 9, Values: [7 6 4 1 2 0 5 8 3]
Column Name: income, Unique Value Counts: 2, Values: [0 1]

[In]:
# Gender distribution graph
df['sex'].value_counts().plot(kind='bar')

[Out]:

Figure 2-3 Gender distribution, male vs. female

As shown in Figure 2-3, with 67% of the population identified as male and 33% as female, which is
considered as imbalanced dataset in the context of machine learning. After comparing both gender and
demographic features in the dataset, its more critical to prioritize and address the demographic imbalance
because it’s more severe in this context.

[In]:
# Race distribution graph
df['race'].value_counts().plot(kind='bar')

[Out]:

Figure 2-4 Demographic comparison (black vs. white population)

As shown in Figure 2-4, a class distribution of 83% to 17% is considered imbalanced in the context of
machine learning. When the distribution of classes is highly imbalanced like this, a model can be biased
toward the majority class, which can lead to poor performance on minority class samples.

Step 5: Model Building

Model building is a crucial step in machine learning and data analysis to make predictions or gain insights.
It involves choosing the right algorithm, training the model, fine-tuning, and evaluating its performance.
Let’s build a model with biased data and see how the model performs.
Split the dataset into train and test data and build a randomforest model on the training data as follows:

[In]:
# Split the data into features (x) and target (y)
x = df.drop(['income'],axis=1)
y = df['income']
# Split the data into training and testing sets
x_train,x_test, y_train, y_test =
train_test_split(x,y,test_size=0.2,random_state=42)

# Create a Random Forest Classifier model and train it

model = RandomForestClassifier(random_state=42)
model.fit(x_train,y_train)

[Out]:
RandomForestClassifier(random_state=42)

Step 6: Predicting for Test Data

Prediction on the test data using the model with bias dataset looks like the following:

[In]:
# Make predictions on the test set and calculate original accuracy
y_pred = model.predict(x_test)
original_accuracy = accuracy_score(y_test,y_pred)
# Display the original accuracy and classification report
print('Original accuracy : ', original_accuracy)
print(classification_report(y_pred, y_test))

[Out]:
Original accuracy: 0.8525949431876344
precision recall f1-score support
0 0.93 0.88 0.91 7814
1 0.61 0.73 0.67 1955

accuracy 0.85 9769

macro avg 0.77 0.81 0.79 9769
weighted avg 0.87 0.85 0.86 9769
The model performed well, with an accuracy of 85 percent. The F1 score provides a balanced measure
of model performance, with higher values indicating a better balance between precision and recall. But as
you can see the f1 score for 1 is 67 percent compared to 91 percent for 0 due to the bias in the data.
Let’s mitigate the bias by resampling. For this, we will consider the race variable since it has 83 percent
white and 17 percent black.

Step 7: Mitigating Bias

To mitigate gender bias, let’s perform resampling. It specifically upsamples the minority class, which is the
Black race in this example. To balance the data, we have used the resample from sklearn.

[In]:
# Handle class imbalance for the race column
black_candidates = df[df['race'] == 0]
white_candidates = df[df['race'] == 1]
upscaled_black =
resample(black_candidates,replace=True,n_samples=len(white_candidates),
random_state=42)
balanced_data= pd.concat([upscaled_black,white_candidates])

Step 8: Modeling with the Balanced Data and Predicting using Test Data
Let’s build a model with balanced data and see how the model performs:

[In]:
# Split the balanced data into features (x) and target (y)
x_balanced = balanced_data.drop(['income'],axis=1)
y_balanced = balanced_data['income']
# Split the balanced data into training and testing sets
x_train_balanced, x_test_balanced, y_train_balanced, y_test_balanced =
train_test_split(x_balanced,y_balanced,test_size=0.2,random_state=42)
# Create a Random Forest Classifier model for balanced data and train it
model_balanced = RandomForestClassifier(random_state=42)
model_balanced.fit(x_train_balanced, y_train_balanced)
# Make predictions on the balanced test set and calculate balanced accuracy
y_pred_balanced = model.predict(x_test_balanced)
balanced_accuracy = accuracy_score(y_test_balanced,y_pred_balanced)
# Display the balanced accuracy and classification report
print('Original accuracy : ', balanced_accuracy)
print(classification_report(y_pred_balanced, y_test_balanced))
[Out]:
Original accuracy: 0.9736605806644717

precision recall f1-score support

0 0.99 0.98 0.98 13410
1 0.92 0.95 0.93 3295

accuracy 0.97 16705

macro avg 0.95 0.96 0.96 16705
weighted avg 0.97 0.97 0.97 16705
The classification report provides insight into the performance metrics of the machine learning model
for both the original and a resampled dataset.
For the original dataset, the model’s accuracy is 85 percent, indicating that it correctly classifies
instances 85 percent of the time. For class 0, the precision and recall are 93 percent and 88 percent,
respectively, indicating that the model is correctly identifying negative instances with a high precision rate
and recall rate. Conversely, for class 1, the precision and recall are 61 percent and 73 percent, respectively,
suggesting that the model is less accurate at identifying positive instances. The F1 score, which is a
weighted average of precision and recall, is 91 percent and 67 percent for classes 0 and 1, respectively.
For the resampled dataset, the model’s accuracy is 97 percent, significantly higher than that of the
original dataset. The precision and recall values for class 0 are high, with both being above 98 percent,
indicating that the model is excellent at identifying negative instances. The precision and recall for class 1
are around 92 percent and 95 percent, respectively, demonstrating that the model has improved in its
performance for identifying and classifying positive instances. The F1 score for classes 0 and 1 are also
improved, with scores of 98 percent and 93 percent, respectively.
Overall, the resampling process has significantly improved the performance of the model, leading to
improvements in accuracy, precision, recall, and F1 score.

Stage 2: Model Bias

In this activity, let’s see the process of detecting bias in the model and its mitigation process. For detecting
and mitigating we will be using a debiasing algorithm by AIF360 (The AI Fairness 360 is an open source
python library containing techniques to help detect and mitigate bias). The goal of this algorithm is to
mitigate bias in datasets and models, specifically for binary classification problems where the protected
attribute is binary (e.g., gender, race, age).
The Adversarial Debiasing algorithm, part of the AI Fairness 360 toolkit, is a machine learning
algorithm that aims to mitigate bias in datasets and models by training two neural networks in an
adversarial manner. The algorithm has shown promise in removing bias related to protected attributes,
but it is just one solution, and more work needs to be done to tackle bias in machine learning.

Dataset Details
The dataset used for data bias is also utilized here, but instead of using the raw data, we will import
preprocessed data from the AIF360 package.
Source: https://github.com/Trusted-AI/AIF360

Step 1: Importing Packages

[In]:
# Import necessary libraries and modules
from aif360.datasets import BinaryLabelDataset
from aif360.datasets import AdultDataset, GermanDataset, CompasDataset
from aif360.metrics import BinaryLabelDatasetMetric
from aif360.metrics import ClassificationMetric
from aif360.metrics.utils import compute_boolean_conditioning_vector
from
aif360.algorithms.preprocessing.optim_preproc_helpers.data_preproc_functions
import load_preproc_data_adult, load_preproc_data_compas,
load_preproc_data_german
from aif360.algorithms.inprocessing.adversarial_debiasing import
AdversarialDebiasing
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler, MaxAbsScaler
from sklearn.metrics import accuracy_score
from IPython.display import Markdown, display
import matplotlib.pyplot as plt
import tensorflow.compat.v1 as tf
tf.disable_eager_execution()

Step 2: Importing the Preprocessed Dataset

For this step, instead of importing raw data and then implementing data cleaning, we use the
preprocessed data from AIF360 since we covered all the preprocessing steps in data bias. See the
following:

[In]:
# Get the dataset and split it into training and testing sets
df = load_preproc_data_adult()
privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]
df_train, df_test = df.split([0.7], shuffle=True)

Step 3: Model Building with Biased Dataset

The AdversarialDebiasing function from the AI Fairness 360 toolkit is used to define the model’s
scope name, privileged groups, unprivileged groups, and session. The debias parameter is set to false,
indicating that the model will not be debiased:

[In]:
# Create a TensorFlow session
session = tf.Session()

# Create a plain model without debiasing

model = AdversarialDebiasing(privileged_groups=privileged,
unprivileged_groups=unprivileged,
scope_name='plain_classifier', debias=False,
sess=session)
model.fit(df_train)

[Out]:
epoch 0; iter: 0; batch classifier loss: 0.680384
epoch 0; iter: 200; batch classifier loss: 0.415313
epoch 1; iter: 0; batch classifier loss: 0.459850
epoch 1; iter: 200; batch classifier loss: 0.389641
epoch 2; iter: 0; batch classifier loss: 0.412084
epoch 2; iter: 200; batch classifier loss: 0.570294
epoch 3; iter: 0; batch classifier loss: 0.424368
epoch 3; iter: 200; batch classifier loss: 0.409188
epoch 4; iter: 0; batch classifier loss: 0.405540
epoch 4; iter: 200; batch classifier loss: 0.347155
epoch 5; iter: 0; batch classifier loss: 0.462146
epoch 5; iter: 200; batch classifier loss: 0.350587
.
.
.
epoch 48; iter: 0; batch classifier loss: 0.394173
epoch 48; iter: 200; batch classifier loss: 0.460599
epoch 49; iter: 0; batch classifier loss: 0.464685
epoch 49; iter: 200; batch classifier loss: 0.447618

[In]:
# Predict using the plain model
# dataset_nodebiasing_train = plain_model.predict(df_train)
y_test = model.predict(df_test)

# Close the session

session.close()
tf.reset_default_graph()

Step 4: Model Building with Debiased Dataset

This code initializes an adversarial debiasing model to remove bias from the model output. The
“privileged” and “unprivileged” groups are specified to ensure that decisions are not biased against the
unprivileged group:

[In]:
# Create a new TensorFlow session
session = tf.Session()

# Create a model with debiasing

debiased_model = AdversarialDebiasing(privileged_groups=privileged,
unprivileged_groups=unprivileged, scope_name='debiased_classifier',
debias=True, sess=session)
debiased_model.fit(df_train)

# Predict using the model with debiasing

# dataset_debiasing_train = debiased_model.predict(dataset_orig_train)
y_debiasing_test = debiased_model.predict(df_test)

[Out]:
epoch 0; iter: 0; batch classifier loss: 0.636262; batch adversarial loss:
0.612745
epoch 0; iter: 200; batch classifier loss: 0.486372; batch adversarial
loss: 0.658000
epoch 1; iter: 0; batch classifier loss: 0.549870; batch adversarial loss:
0.658346
epoch 1; iter: 200; batch classifier loss: 0.472538; batch adversarial
loss: 0.644015
epoch 2; iter: 0; batch classifier loss: 0.453551; batch adversarial loss:
0.646959
epoch 2; iter: 200; batch classifier loss: 0.445221; batch adversarial
loss: 0.619725
epoch 3; iter: 0; batch classifier loss: 0.454317; batch adversarial loss:
0.663603
epoch 3; iter: 200; batch classifier loss: 0.386634; batch adversarial
loss: 0.583133
epoch 4; iter: 0; batch classifier loss: 0.444041; batch adversarial loss:
0.633082
Other documents randomly have
different content
 CHAPTER 4
Samuel's First Specimen

"What do you make of that?" puffed Samuel Salt as Roger dropped

breathlessly down on his shoulder. "Well, 'Hung wung wah HEEE!' it
is. Let's give them a cheer for luck." Lifting his great voice, the Royal
Discoverer for Oz, helped out by his two shipmates, sent the weird
call booming back across the water.
An answering call came from the Island, and then, with a hiss and
thud, a small glowing object fell on the deck. Fortunately the fire
tongs were still handy and picking up the offending object before it
could do any damage, Ato marched sternly off to the galley.
Stopping long enough for another wave to the Island, which was
growing smaller and smaller as the Crescent Moon sped away,
Samuel hastened after his cook, jotting down hurried notes in his
journal as to latitude and longitude as he ran along.
"There's something written on this piece of lava," announced Ato,
who had dropped the smoking souvenir from Lavaland on the stove.
Peering over his shoulder, Samuel could see queer raised symbols
and signs on the sulphurous surface of the rock.
"There's something crawling on it, too," volunteered Roger, who was
perched on the towel rack above the stove, and had a better view,
"a golden frog or a lizard."
"Merciful mustard! What next?" groaned Ato.
"Why, this—this—" Samuel's voice quivered with excitement and
disbelief, "this, Mates, is as fine a specimen of a Preoztoric Monster
as a scientist could hope for; a real live salamander, a fire lizard,
straight from the burning depths of yonder crater. Stars! Tar and
Tarrybarrels! This is even better than the baby and will prove my
point just as well."

"Does it have to live on my stove?" asked Ato ominously, as the

Salamander slid merrily backward and forward over the red hot
plates of the range. "Home on the range!" snickered Roger, winking
at the Pirate.
"Just till I can fix up a hot box for it," apologized Samuel, "but don't
fret, old Toff, it doesn't bite and if it falls on the floor, all you have to
do is scoop it up and put it back before it goes out."
"Not only cook, mate and swab, but now I'm nursemaid to a fire
lizard." Ato shuddered, and reaching for his tall cook's cap, jammed
it down hard on his shiny bald head.
"You can keep it in an iron pot while you cook," suggested Roger
practically, "and after all, King dear, it's the only Salamander in
captivity. Here, Sally, here Sal—this way, my little crater critter."
Tilting the pot on the back of the stove, Roger was delighted to find
the Salamander quite willing to answer to her new name. As she slid
adventurously into the small cooking vessel, the Read Bird quickly
righted the pot and clapped on the cover. "There," he exclaimed with
a satisfied nod at his Master, "how's that?"
"Well, I suppose I'll have to put up with it," sighed Ato resignedly.
"But in some ways pirating was easier than discovering, Sammy. At
least, we never kept the captives on the stove. And NOW—" Ato
waved his arms determinedly. "Clear out, both of you. It's three bells
and time to stir up the food. And just take that pesky rock along
with you. I've meat to broil!"
"When this cools, maybe I'll be able to figure out the language,"
exulted Samuel, removing the offending piece of lava with a cake
turner. "All in all, a most interesting and profitable day, eh, Roger?
An island, a visit from a fire baby, and a real live Preoztoric monster."
"Not bad," agreed the Read Bird, transferring himself to the
Captain's shoulder. Depositing the piece of lava on an iron hatchway
to cool, Samuel strode happily along the deck, stopping to light the
red lamps on the port and the green lights on the starboard. Roger
himself had just hung a white light in the rigging when a lusty call
from the galley sent him flying off to help Ato serve the dinner.
"What could be cozier than a life at sea?" he reflected, winging
jauntily into the main cabin with a dish of roast potatoes. Ato puffed
cheerfully behind, bearing a huge tray. On the tray a steaming
tureen of soup, a pot of coffee, seven dishes of vegetables and two
of smoking meats sent up tantalizing whiffs and fragrances. Later
when the Read Bird brought in the pudding, he and Sammy soberly
agreed it was the tastiest feast Ato had served on the voyage.
The main cabin of the Crescent Moon, with its red leather couches
under the ports, its easy chairs and tables clamped to the floor to
keep them from shifting, with its ship's clock and ship's lanterns, was
a cheery place to be when the day's work was ended. There was a
huge fireplace for foggy evenings and every visible space on the wall
was covered with pictures of pirate ships, ancient sailing vessels and
rough maps and charts of strange and curious islands. While Samuel
and Ato sat at their ease to finish off the pudding, Roger took his
upon the wing, darting in and out between bites to assure himself
that all was well on deck. There was a tiny crescent moon sliding
down the sky, and the slap of waves against the side of the ship and
the wind creaking in the cordage made as pleasant a tune as the
heart of a seaman could wish for.
"Now what could be better than this?" said Samuel Salt exhaling a
cloud of smoke from his pipe and stretching his legs luxuriously
under the long table. "A tidy ship, a good wind and the whole wide
sea to sail on."
"Suits me!" grinned Ato scraping up the last of the hard sauce and
settling back with a grunt of sheer content. "Did you mark up our
volcano on the chart Sammy, and what are we calling it Mates? An
island must have a name you know."
"I know." Samuel blew another cloud of smoke upward and cleared
his throat. "If it's agreeable to all hands and Roger, I'd like to call it
Salamander Island after Sally."
"Why not? There's a Sally in our galley and a real nice gal is Sally,"
warbled Roger, settling on the back of Samuel's chair and wagging
his head in time to the music.
"Sing like a bird, don't ye?" muttered Samuel striding over to the
map of Oz and surrounding countries and oceans that covered the
west wall.
"I AM a bird," screamed Roger fluttering up to his shoulder. "'Bout
here she would lie, Master Salt, sixty leagues from Octagon Island."
As Roger talked on, making numerous suggestions, the Captain of
the Crescent Moon drew with red chalk a small but effective picture
of Salamander Island showing the volcano in action and the
Lavaland Islanders grouped around the crater's top.
"Taken this day without a shot or the loss of a single man," printed
Samuel in neat letters under his sketch.
"Don't forget, you shot the baby," twittered Roger raising a claw
argumentatively.
"Oh, we can't put in small details like that," sniffed the Captain
stepping back to admire his drawing.
"Seems odd for us to be discovering and taking possession of islands
for a country we know so little about," mused Ato, looking
thoughtfully at the map on the west wall. "Why, we've only been to
Oz once ourselves."
"Yes, but everybody knows about Oz," Samuel said putting the red
chalk back in the table drawer. "Our business is with wild new
countries that have never been seen or heard of. Besides, anyone
can see that Oz is overpopulated and needs new territories and sea
ports. And since Ozma is so clever at governing, and her subjects all
so happy and prosperous, the more people who come under her rule
the better!"
"Aye! Aye!" agreed Roger, peering with deep interest at the map.
Small wonder the Read Bird was interested, for Oz is one of the
most exciting and enchanting countries ever discovered. There are
four large Kingdoms in Ozma's realm, the Northern Land of the
Gillikens, the Eastern Empire of the Winkies, the Southern Country of
the Quadlings and the Western domain of the Munchkins. Each
forms a triangle in the oblong of Oz. The Emerald City which is the
capital, is in the exact center where all these triangles meet. Each of
these Kingdoms has its own ruler, but all four are under the
sovereign rule and control of Ozma, the small but powerful fairy who
lives in the Emerald City. On all sides, Oz is surrounded by a deadly
desert and beyond the desert lie the independent Kingdoms of No-
Land, Low Land, Ix, Play, Ev, the Dominions of the Gnome King, and
many other strange and important Principalities. These countries
form a narrow rim around the desert, and beyond this rim lies the
Nonestic Ocean itself, stretching in all directions and to no one
knows what far and undiscovered shores. Each of the four Kingdoms
in Oz shown on Samuel's map was so dotted with smaller Kingdoms,
cities, towns, villages and the holdings of ancient Knights and
Barons, there was scarcely room for another castle. With young
Princes growing up on every hand, Roger could well sympathize with
the need of Ozma for more territory.
"Won't the Ozians have too long a way to come before they reach
these new islands and countries we discover?" inquired the Read
Bird, after staring at the map for some moments in silence.
"Not a bit of it!" Samuel dismissed Roger's objection with a snap of
his fingers. "I hear the Wizard of Oz is working on a new fleet of
airships, that will make crossing the desert and Nonestic a real lark
and enable new settlers to reach these outlying islands in a day or
less. So all we have to do is to proceed with our discovering. Ozma
will attend to the rest. This volcanic island may not be as useful as
some of the others, but one can never tell. How about picking up a
few islands for you, Ato, as we ride along?" The former pirate
dropped his arm affectionately round the shoulders of his Royal
Cook.
"No, thanks," grunted Ato, rolling cheerfully to his feet. "One's
enough. What would I want with any more islands? Why I'd never
get off on a voyage. But pick yourself a couple, Sammy, why don't
you?"
"Who, ME?" Samuel Salt shook his head emphatically. "A ship's all I
can handle and I wouldn't trade you two buckets of sea water for all
the islands in the Nonestic. One ship and one crew's enough for me,
and since you're my crew, you'd better turn in—we've had a hard
day and another one coming. I'll take first watch, Cooky, here, shall
have middle, and you Roger can be the early bird on morning
watch."
"Ho hum! I'm right sleepy at that," admitted Ato, starting to heap up
plates. "Give me a lift with the dishes, Roger, will you?"
"Oh, throw 'em overboard," directed Samuel Salt recklessly. "There's
plenty more in the hold and I'm agin all extry labor."
"Hurray!" screamed Roger seizing the coffee pot and winging merrily
through an open port.

"Avast! Avast there! Not my coffee pot!" pleaded Ato, making after
the Read Bird with surprising speed considering his tonnage. "Stop
you great Gossoon! How many times must I tell you I'm boss of the
galley?" Catching Roger by the leg just as he reached the rail, Ato
snatched back his precious coffee pot and hugged it protectively to
his bosom. "Why I've just got this contraption broken in proper," he
panted indignantly. "A coffee pot's like a pipe, it's got to be
sweetened and seasoned. Heave over the plates and cups if you
like," he went on, relenting a bit as he noted the keen
disappointment on Roger's face, "but save the soup tureen. I'll
wager there's not another that size on the ship and the Captain must
have his soup. What a splendid pot of soup THIS would make,"
murmured Ato looking dreamily down at the sea, "a bit salty,
perhaps, but full of snapper and porgy and tender young sea shoots.
Why that foam's as near to whipping cream as anything I've ever
gazed on."
Tearing himself reluctantly from the appetizing sight, the Royal Cook
padded off to put the galley in order for the night, while Roger with
loud squalls of glee dropped the plates and saucers one by one over
the side. In this way the dishes were soon done, the cabin tidy and
shipshape, and by eight bells the King and the Read Bird were
sleeping soundly and Samuel Salt had the ship to himself.
First, he made a complete round of all decks, glanced at the
barometer and compass, and furled the fore and mizzen topsails.
Then he took the cooled piece of lava down to the hold. The strange
signs and symbols had hardened, and labeling it carefully with the
date and name of Salamander Island, Samuel placed it on his
shelves for further study. Then returning to the main deck he set a
portable ship's lantern on a coil of rope and settled down to fix a hot
box for the Salamander. Selecting from the material he had brought
from the hold an iron box with a glass lid, he covered the bottom
with sand and pebbles. Knowing salamanders require hot water as
well as hot air, he placed a tiny flat pan of water in the corner of the
box to serve as a swimming pool. A burning glass in the day time
and an alcohol lamp under the box at night would supply the
necessary heat, and setting the whole contrivance on an iron tray in
the cabin, Samuel went joyfully off to fetch the fire lizard.
The Salamander was still in the pot on the back of the stove, and
giving her an experimental poke with his finger, Samuel was
astonished to find her quite cool to the touch. This was surprising
considering she could only live in the most intense heat. But without
stopping to figure it out, the Captain picked her up between thumb
and forefinger, carried her to the cabin and popped her into the iron
box. He had already lighted the lamp under the box so that
everything was red hot and cozy for her. The small captive seemed
to appreciate her new quarters, wriggling over the hot pebbles and
sand, then splashing gaily in her swimming pool.
"Quite a girl!" sighed the pirate, resting his elbows on the table and
gazing happily down at the first prize of the voyage. "You're going to
be great company for me, Sally." As if she really understood, the
lizard gave a squeak and tapped loudly on the glass lid with her tail.
The pipe almost dropped from Samuel's mouth at Sally's strange
behavior, and lifting the lid he peered inquisitively down at her.
Before he had a chance to clap it shut, the Salamander hurled
herself upward, landing smartly on the bridge of the Pirate's nose,
from where she slid cleverly into the pipe itself.

"Well I'll be scuppered!" gasped the Royal Explorer looking slightly

cross-eyed down the bridge of his nose as Sally coiled up
comfortably in the bowl of the pipe. "The little rascal wants to keep
me company, and so she shall, bless my boots, so she shall! Why
this is plumb cute and cozy and something to write in my journal."
Puffing away delightedly Samuel stepped out of the cabin and all
during his watch, the little Salamander rested contentedly in his
pipe. Sometimes she peered up inquisitively over the edge, but
mostly she lay quietly on the smoking tobacco, looking with calm
interest at the sky and the rippling sails over her head. Not only did
she keep his pipe from going out, but never had it drawn so well.
So, filled with a vast wonder and content, Samuel strode up and
down the deck. Not till midnight when he roused Ato could he bear
to put Sally back in her box and only then, after he had promised
her another ride in the morning. But when morning came, Samuel
had no time to keep his promise, for while Ato was cooking
breakfast and the Captain himself catching forty winks in the cabin,
the raucous voice of the Read Bird came whistling down from the
foremast.
"Land Ho! Land! More Land. Island tuluward, Captain!"
 CHAPTER 5
Patrippany Island

"All hands on deck! Come on! Come on!" yelled Samuel Salt running
past Ato's galley dragging on his clothes as he ran. "There's an
island tuluward, you lubber."
"Well, 'tain't a flying island is it?" Ato stuck a very red face out the
door. "I guess it'll stay there till I turn the bacon, won't it? No cause
to burn the biscuits just 'cause an island's sighted is there?" But in
spite of his pretended indifference, the ship's cook shoved all his
pans on the back of the stove and hurried out on deck. "Rich and
jungly, this one," he observed, resting his arms comfortably on the
rail, "and from what I can see a good place to grow bananas and
whiskers. Look, Sammy, even the trees have beards."
"Moss," muttered Samuel Salt striding over to the wheel. "Fly ashore
Roger and see whether there's a good place to put in."
Twittering with importance and curiosity, the Read Bird flung himself
into the air. In ten minutes he was back to report a wide river cutting
through the center of the island from end to end. The foliage was so
dense, Roger had not been able to discover any signs of habitation,
but after viewing the mouth of the river through his glasses, the
Captain decided to take a chance, and sail through.
"Now, Sammy, let's not do anything hasty," begged the ship's cook
lifting his floury hands in warning, "nor try to conquer a country on
an empty stomach. This may be an important island, so after we
eat, let us put on our proper clothes and plant the Oz flags with
dignity and decorum."
"Spoken like a King and a seaman," approved Samuel Salt, "and if
my eye does not deceive me, I'll have the ship in the river as soon
as you have the coffee in the pot. Then we'll ride in with the tide,
put on our discovering togs and proceed with the business of the
day."
So while Ato returned to his galley and the Read Bird to his post in
the foremast, Samuel swung the Crescent Moon in toward the
island. Each felt a slight twinge of uneasiness as the ship left the
open sea and began to slip rapidly up the broad new and
unnavigated jungle stream. Vine covered trees pressed close to the
banks, and birds and monkeys in the branches kept up an incessant
screech and chattering. A flock of greedy pelicans flopped comically
after the ship and as they penetrated deeper and deeper into the
jungle it almost seemed as if they were entering some dim green
land of goblins.
"A fine target we make for anyone who cares to shoot at us,"
moaned Ato, as he waddled backward and forward between the
cabin and galley with cups and covered dishes. "Ugh!"
"Yes, I wouldn't be surprised to feel an arrow in my back any minute
now," assented Samuel Salt brightly, "though I must say I'd much
prefer a fried mackerel in my stomach."
"Come on then," shuddered Ato, in no wise cheered by Samuel's
remarks, "breakfast's ready and we may as well eat before we die."
"Now never say die!" roared the Royal Explorer of Oz, touching the
buttons to furl sail and yelling to Roger to let go the anchor. "Never
say die—say dee—dee-scovery is our aim and purpose, Mates. Dee-
scovery with a hi de di dide di dough!" sang Samuel vociferously to
keep up his own spirits. Finally with the ship motionless amidstream
the three shipmates sat down to breakfast. Their nerves were tense
and their ears cocked for signs of approaching natives, but except
for the noise of the birds and monkeys and the occasional splash of
some river creature, there was no sound to indicate the ship had
been sighted by the islanders.
"Nobody's home," concluded Samuel, finishing off his third cup of
coffee at one toss and hurrying off to his cabin. Roger, having only
Oz flags and no shore togs to bother him, generously offered to
clear away the dishes and amused himself by throwing scraps and
the rest of the biscuits to the pelicans. He had just tossed over the
last biscuit when Ato appeared in a grand satin coat and breeches,
long cape and three-cornered hat. The elegance of his apparel was
somewhat marred by the bread board he had belted round his
middle and the bread knife and blunderbuss he had stuck through
his sash.
"Ha, hah!" roared Samuel Salt, giving the bread board a resounding
whack. "Something to stay your stomach, EH?" Samuel himself was
as stylishly attired as the King, his three-cornered hat at a dashing
angle. Under his arm he had two pairs of tremendously long stilts.
"No need for us to get all grubby lowering the boat. We'll wade
ashore this time," explained Samuel as Ato's eyes grew round and
questioning. "Easy as walking on crutches; just watch me, Mate."
Now Samuel, it must be confessed, had been practicing stilting on
Elbow Island, so naturally it came easy to him. First he put his stilts
over the side, then vaulting the rail, he seized the tops and settled
his feet in the cross pieces at one jump and started walking calmly
up and down gleefully calling for Ato to follow. It all looked so
simple, Ato handed the basket of lunch he had packed to Roger, and
seizing his stilts began anxiously feeling around for the river bottom.
Satisfied that it was solid, he climbed boldly up on the rail.
"That's it! That's it!" applauded Samuel. "Now grab the tops, Mate,
and start coming."
"Chee tree—tee—hee—!" screeched the monkeys derisively as Ato
clung precariously to the rail with one hand and maneuvered his
stilts with the other. By some miracle of balance the fat King actually
managed to mount and hold on to his perilous walking sticks. Then
with a long quivering breath he heaved one forward. He was about
to take another step when a desperate scream from Roger almost
caused him to topple over backwards.
"'Gators!" croaked the Read Bird, beating his wings together
violently. "Watch out for those 'gators."
"Why bother him with gaiters at a time like this? They look perfectly
all right to me." Samuel Salt frowned up at Roger.
"Not his gaiters, river 'gators, alligators, CROCODILES!" wailed
Roger, beginning to fly in agonized circles. "Crocodiles and WORSE."
Samuel, eyeing what he had supposed to be a pile of rotten logs on
the river bank, saw dozens of the slimy saurians slide into the water
and come savagely toward them.
"Back to the ship! Back to the ship!" babbled the Read Bird,
clutching Ato's collar with a frantic claw. But the King was too
frightened to move. The sight of the bleary-eyed river monsters
made him tremble so violently his stilts twittered and swayed like
trees in a hurricane. He could not for the life of him take a step in
either direction. With a loud cry Samuel started to help him, but a
crocodile reached Ato first. Its jaws closed with a vicious snap on the
King's left stilt and with a heart-rending shriek Ato plunged into the
slimy river.
"There, there! Now you've done it!" sobbed Roger. "Fed the kindest
soul who ever served a ship's company to a parcel of crocodiles!"
Dropping the Oz flags and lunch basket, he made an unsuccessful
grab for his Master's arm. But even if he had caught it, Ato's great
weight would have pulled them both under, and now only a circle of
bubbles showed where the luckless explorer had disappeared. Firing
his blunderbuss to frighten off the rest of the crocodiles, Samuel,
striking left and right with his stilts, propelled himself forward, while
Roger pecked futilely at the monster that had felled his Master. But
just as Samuel, after boldly driving off the dragon-like creature,
prepared to dive in and save Ato or perish with him, a dripping head
appeared above the water.
"Thank you. Thank you very much!" murmured a mild voice. "I
haven't had as nice a present as this since I was an itty bitty baby.
Now what can I do for YOU?" Neither Samuel nor Roger could speak
a word, for where the King had gone down, a tremendous
hippopotamus was coming up, the lunch basket hanging carelessly
out of a corner of its mouth. For a wild moment Samuel thought his
enormous friend and shipmate had been transformed by some
witchcraft into this ponderous beast. He even imagined he caught an
expression of Ato's in the monster's moist eye. But this gloomy idea
was soon dispelled, for, as the creature rose higher out of the water,
they could see a desperate and bedraggled figure sprawled across its
slippery back.
"Ahoy, Mate!" choked Samuel, his heart thumping like a trip hammer.
"Is it really you? Are you safe, then?"
"Safe!" quavered the half-drowned and mud-covered King of the
Octagon Isle. "SAFE?" He peered dizzily at the churning crocodiles
just a boat's length away, and his voice cracked and broke. "I never
felt safer in my life. What am I riding, a whale or an elephant?"
"A river horse," explained the hippopotamus, looking kindly over her
shoulder. Then, as the crocodiles began to hiss and roar and come
rolling toward them, she gave a ferocious bellow and snort. "Away
with you! Be off, you river scum!" she squealed viciously. "These
travelers are MINE. Shoot your fire stick, Master Long Legs. That will
fix them." For a moment the crocodiles held their post, then, as
Samuel fired his gun repeatedly, they began to slide sullenly across
the river to the opposite bank. "Hold fast, Master Short Legs, and I'll
soon have you ashore," wheezed the hippopotamus, speaking out of
the corner of her mouth so as not to drop the picnic basket.
"Yes, yes, but what then?" shuddered Ato, trying to get a finger hold
on the monster's slippery neck.
"Why, then, we'll both tell our stories, and after that I'll eat," snorted
the river horse, paddling joyously toward the bank.
"You'll EAT!" groaned Ato, ready to roll back into the river. "Oh, my
father and mother and maiden aunts!"
"Did you hear that?" Dropping to Samuel's shoulder, Roger
whispered fiercely. "Quick now, a shot behind the ear, before it gets
any further. Are you going to do nothing while this ravenous monster
carries off my poor Master?"
"Sh-hh!" warned Samuel, holding up his finger. "These creatures do
not eat meat or men. They're herbivorous, my lad, and this one
seems uncommonly kind and friendly. But what puzzles me—" the
Royal Explorer looked intently into the face of the Read Bird. "What
puzzles me is to find this one talking our language. To my
knowledge, only animals in Oz, a few in Ev and you on the Octagon
Isle have the gift of speech. And I tell you, Mate, this is a valuable
discovery, and a simply splendid specimen of a pachydermatous
talking aquatic." Whether the last few words in this sentence or a
stone in the river bottom tripped up the Captain, Roger never knew,
but without any warning Samuel turned a sudden back somersault
into the river, going under as completely as Ato had done.
"Ugh—gr—ugh!" he gurgled, coming up full of mud and disgust.
"How did that happen?"
"Stilts!" sniffed Roger, whose wings had saved him from going down
with Samuel. "A splendid way to get ashore, Master Salt, so neat
and tidy. And a fine Discoverer you look now."
Sighing deeply, Samuel watched his stilts floating out of reach, then
shaking his head violently to get the water out of his eyes, he swam
thoughtfully after the hippopotamus. As he dragged himself up on
the bank, a monkey swinging by its tail from the lower branches of a
tree snatched his three-cornered hat and scittered all the way to the
tree top, at which all the other monkeys let out shrill hoots of
mocking merriment.
"Ah! The welcome committee!" sniffled Ato, rolling off the
hippopotamus. "Well, Sammy, wherever it is, here we are and a nice
mess you've made of the landing. Clothes ruined, weapons gone,"
(Ato felt his middle dejectedly for his bread knife and blunderbuss),
then hitching up the bread board at his waist looked long and
accusingly at the Leader of the Expedition.
"Now you mustn't mind a little mud," said the hippopotamus, setting
down the picnic basket and gazing from one to the other with frank
interest and curiosity. "Mud is beautiful and SO healthy."
"Not for me," frowned Samuel Salt, endeavoring to remove the thick
green slime from his hair and ears with his damp silk handkerchief.
"But I suppose we'll dry off in time and—"
"Proceed with the business of the day," finished Ato sarcastically, as
he squeezed the water out of his silk pantaloons and coat tails. "But
I hope you don't mind my saying that a seaman should stick to his
boats, Samuel. If I had not fallen in with this kind and obliging
hippopotamus, I'd have been a crocodile's lunch by this time."
"Oh, I'd have got you out somehow," muttered Samuel, smoothing
back his hair sulkily. "And those stilts really saved your life. Suppose
that animal had bitten your leg instead of your stilt? By the way,
what's the name of this island, Mate?" Anxious to change the
subject, Samuel turned to Ato's tremendous rescueress.
"Mate?" repeated the hippopotamus, wiggling her ears inquiringly,
"What may that mean?"
"It is what a seaman calls his crew and his friends," explained
Samuel, grinning in spite of himself.
"Seaman? Mate?" mused the hippopotamus in a rapt voice. "How
cozy, how beautiful!" Overcome with emotion, the mighty monster
leaned forward and lapped up the picnic basket, Oz flags, lunch and
everything. "I shall remember this as long as I live," she assured
them with a gulp as one of the flags went sideways down her throat.
"Nikobo, Little Daughter of the Biggenlittle River People, bids you
welcome to Patrippany Island."
"Little daughter!" exclaimed Ato in a smothered voice. "Ha, ha!
Patrippany Island. Ho, ho! This is interesting. I knew there was a trip
in it somewhere, a wet trip for us, eh, Samuel?"
"But what I don't understand," said the Royal Explorer of Oz, briskly
massaging his beard with his handkerchief, "is how you happen to
speak our language. Do all the creatures on this Island talk? I don't
mean that monkey chatter above."
"No, none of the other creatures here speak the language of man,"
answered Nikobo solemnly. "I never knew I could speak it myself till
five moons ago last Herb Day."
"Herb Day? Dear, dear and dear! How confusing it all grows," sighed
Ato, emptying the water out of his hat which had somehow survived
his river ducking. "Do you suppose she means Thursday? Roger!
ROGER! Keep away from those monkeys. Do you wish to lose all
your tail feathers?"
"Oh, it's all very simple," Nikobo rolled her eyes from side to side.
"One day I eat herbs and that is Herb Day. One day I eat twigs and
that is Twig Day, and one day I eat grass and that is Grass Day, and
—"
"And one day you eat lunch baskets and Oz flags, and I suppose
that makes it Flag Day," chuckled Roger, coming down from a little
excursion in the tree tops. "She's swallowed the Oz flags, Skipper,
and if that doesn't make her a citizen of Oz, I'll eat my feathers."
"Go ahead, if it will keep you any quieter," said Samuel Salt, who did
not want this interesting conversation interrupted by Roger's
nonsense. "So you only began to speak our language five moons
ago last Herb Day? What made you do that?"
"A boy," confided Nikobo with a ponderous wag of her head.
"Ah, now we're getting somewhere." Feeling in his pocket, Samuel
pulled out a small note book and pencil, still damp but usable. "Was
it a native boy?" he asked eagerly.
"No, no, certainly NOT." The hippopotamus panted a little at the very
idea of such a thing. "The Leopard Men speak a strange roaring
language I have never been able to make head or tail of. Besides, to
speak to them would not be safe nor desirable. The Leopard Men
have long tusks and spears and—"
"Leopard Men!" yelled Ato, flinging both arms round the trunk of a
tree. "Oh! Oh! and OH! I wish we were safely back at pirating,
Sammy. Here we are marooned on this miserable monkey island,
inhabited by Leopard Men, surrounded by crocodiles and no way of
getting back to the ship."
"You forget me," murmured the hippopotamus. Lumbering over to
Ato, she gave him a gentle nudge with her moist pink snout.
"Nikobo, Little Daughter of the Biggenlittle River People, will carry
you anywhere you wish to go."
 CHAPTER 6
A Little Wild Man

"Not yet, not yet," protested Samuel Salt as Ato made a clumsy
attempt to mount the hippopotamus. "Why, we've only just come,
Mate. We can't go without seeing these Leopard Men and this
strange boy who speaks our language."
"Oh, CAN'T we?" Drawing in his breath, Ato made a flying leap at
Nikobo, and this time managing an ear hold, pulled himself
determinedly up on her moist, slippery back. "Goodbye, Samuel,"
said the King with a firm wave of his hand. "If you bring any Leopard
Men back to the Crescent Moon, you can discover yourself another
cook. No Leopard Men. Mind, now!"
"Oh, you needn't worry about that." The hippopotamus closed one
eye and smiled knowingly to herself. Thoroughly annoyed by the
desertion of Ato and the superior grin of the river horse, Samuel
snatched a long rapier from his belt and glowered belligerently
around him.
"Shiver my timbers! You think I'm not strong enough nor smart
enough to fight these savages? HUWHERE are these Leopard Men?"
roared the former Pirate in such a reverberating voice the monkeys
fled silently to the tree tops, and even Roger put his head under his
wing.
"Gone, all gone!" explained Nikobo as she started calmly down
toward the river bank.
"You mean there are no Leopard Men on this Island now?" Looking
with horror and aversion at the crocodile-infested river, Ato began
tugging at Nikobo's ear. "Not so fast, my good creature! Wait a
moment, my buxom lass! Perhaps I'll stay with Sammy after all."
"Well, just as you say." With scarcely a pause in her stride, the
hippopotamus turned round and waddled amiably back to the strip
of sand where Samuel Salt stood staring sternly into the jungle
beyond.
"This is a great disappointment to me, Mates," sighed the Captain of
the Crescent Moon mournfully wringing out the lace ruffles of his
cuffs. "To have taken a Leopard Man back to the Court of Oz would
have been an achievement worth the whole voyage."
"Now there's where we're different," murmured Ato, settling into a
more comfortable position on the back of the river horse. "I myself
would rather be disappointed than speared by a savage, and I don't
care how many Leopard Men I miss seeing. Rather be spared than
speared, ha, ha! Tee, HEE, HEE!" Ato chuckled from sheer relief.
"Shall I fly back to the ship for some more Oz flags?" Roger flapped
his wings inquiringly. "If the Leopard Men are really gone, then
Patrippany Island is ours without a spear thrown."
"That's so," mused Samuel Salt, thrusting his rapier back into its
sheath and beginning to show a little interest in the island itself. "Fly
ahead, my Hearty."
"And bring back some ship's biscuit," called Ato. "All this diving and
mud turtling has left me weak as a fish. And while we're waiting for
Roger, perhaps Nikobo will tell us a little about these Islanders. Were
they little or big, black or brown?"
"Yellow," answered the hippopotamus gravely. "Big and yellow with
brown spots all over their hides. They had brown hair, mane and
eyes, and rough snarling voices. They used neither huts nor shelter,
but roamed like the animals through the jungle, hunting, fishing and
fighting. They had hollowed out logs for use in the water and last
Twig Day every Leopard man, woman and child climbed into the
long boats and paddled out to sea. Shortly afterward—" Nikobo's
eyes grew round and shiny at the mere memory, "shortly afterward
a great hurricane arose and my family and I, watching from the
mouth of the Biggenlittle River, saw the boats and men swept under
the waves. Some of the logs floated back to the islands, but the
Leopard Men and women we never saw again."
"Not even ONE?" exclaimed Samuel peevishly.
"Not even one," Nikobo assured him solemnly. "And to tell the truth,"
the hippopotamus flashed a sudden and expansive sigh, "it is much
better and safer without them. The one problem is the boy, and I've
been feeding him myself."
"Oh, yes, the boy who speaks our language," mused Samuel, still
lost in bitter reflections of the Leopard Men he should never see face
to face.
"What've you been feeding him?" asked Ato, suspiciously. "How
would a hippopotamus know what to feed a boy?"
"I do the best I can," said Nikobo in a hurt voice. "Every day I collect
fresh roots, herbs and grasses for him."
"Roots, herbs—grasses! Merciful Mustard! A boy's being fed on
roots, herbs and grasses, Sammy. Did you ever hear of anything
more ridiculous in your life?"
"No worse than spinach," mumbled Samuel Salt. "But SAY, look here
—" The Royal Explorer of Oz raised his arm imperiously. "What is a
small boy doing on this island? How'd he get here in the first place,
and where is he now?"
"Follow me," directed Nikobo in a dignified voice. "Follow me and
you shall know all." As Roger appeared at that moment with the Oz
flags and biscuits, the little procession immediately got under way,
Ato calmly riding behind.
On her many visits to the strange boy, Nikobo had worn a path
through the tangled growth of vines and bush. Tenuous trees
dropped their branches over this path and stretched out their
gnarled roots to trip the unwary traveler. Several times Roger let out
hoarse squeals as a huge snake coiled along the limb of a tree,
thrust out its ugly head. Gaudy flowers from the vines that closely
entwined every tree, filled the air with a damp sleepy fragrance, and
Samuel Salt, darting his eyes left and right, held his blunderbuss
ready for any savage beast that might spring upon them. But the
jungle creatures, thinking the Leopard Men had returned, slunk
further and further into the green shadows and without any mishaps
or encounters, Nikobo brought the explorers to a small clearing in
the whispering tangle of green.
Here they were suddenly confronted by a stoutly built cage, its bars
constructed of saplings set scarcely an inch apart. On a heap of
grass in a corner of the cage crouched the lonely figure of a little
boy clothed in a single leopard skin.
"Well, goosewing my topsails!" panted Samuel Salt, deceived at first
by the leopard skin. "A little wild man, a Leopard boy, as I'm a salt
sea sailor!"
"It's nothing of the kind," Nikobo contradicted him sharply. "Can't
you see he is white and has teeth as straight as your own instead of
tusks? He's not like the Leopard Men at all."
"But who put him in this cage? What's he done, and what's he doing
here?" Slipping off Nikobo's back, Ato pressed his face close to the
bars of the strange prison.

"I am waiting for my people to come and rescue me," stated the
boy, rising with great dignity from his bed of grass. Folding his arms,
he looked haughtily out at the explorers. "Who are these men,
Nikobo?" he inquired sternly. "Why have you brought them here?"
"Because they seemed friendly and speak your language," puffed
the hippopotamus, beaming lovingly at her small charge. "Because I
thought they might break these bars and set you free. They have a
hollow log seventy times as large as the hollowed logs of the
Leopard Men. In this they could easily carry you over the waters and
back to your own people. I've tried to break this miserable hutch
dozens of times," explained Nikobo, turning to Samuel Salt. "But the
saplings are sunk so deep, I've been afraid I'd crush Tandy as well
as the cage if I pushed too hard."
"Quite likely," said Samuel Salt, rapping the bars with his knuckles.
"We'll have to fetch an ax from the ship. But who shut you up here,
little Lubber, and how long have you been prisoner on this island?"
"Five months and a half," answered the boy after consulting one of
the bars in the corner of his cage. "I've made a nick in this bar with
my teeth for every day I have been here."
"Well, that's all over now, you poor child, you!" Ato's voice shook
with indignation as he looked in at the little boy whose every rib
showed plainly under the skin. In fact, a heap of grass and dried
roots in the cage made the kind-hearted monarch shudder with
distaste and sympathy. "You shall come with us and eat like a King,"
he promised, nodding his head cheerfully, "and learn to be an able-
bodied seaman to boot." Instead of looking grateful or pleased, the
boy whom the hippopotamus had called "Tandy" merely stood
looking between the bars of his cage.
"Why should I go with you?" he said finally and wearily. "You look
wild and dangerous to me, and far worse than the Leopard Men.
Here, at least I have Kobo to take care of me, and who knows what
further perils and hardships I should suffer at sea?"
"Ho! HO! And how do you like that, my lads?" Roger rocked
backward and forward on Samuel Salt's shoulder. "The young one
speaks truly. If you could but see yourselves, my Hearties." Now
both Ato and Samuel had forgotten their plunge in the river, but with
their hair and clothing still covered with mud and slime they looked
the veriest rogues and rascals. And while Ato regarded himself with
embarrassment and discomfiture, Samuel took a quick step forward.
"SO!" roared the great seaman angrily. "So, you don't trust us, eh?
Well, stay here if you wish and grow up like a monkey. You look like
a little wild man already."
"STOP!" Nikobo quivered all over with resentment. "You must not
call Tandy a wild man."

"Don't mind." The boy drew the leopard skin around him with quiet
dignity. "I can bear it. I have borne far worse. I can bear anything. I
am a KING and the son of a King's son! Tell them to go away, Kobo."
"Now, Now, NOW! This is nothing but nonsense." Ato clapped his
hands sharply. "However we look, my young squab, you are in good
and royal company. My mate here, Captain Salt, is Captain of the
Crescent Moon, Royal Explorer of Oz, and a Knight, besides. I,
though at present a ship's cook, am King of the Octagon Isle, and
Roger, here, is as Royal a Read Bird as ever wagged a bill and wing.
If you say you are a King, we will have to believe you, though 'tis
hardly credible." Ato stared with round eyes at the matted hair and
dirty body of the little prisoner. "If you say you are a King we must
believe you, but in return you must believe us, and stop all this hoity
toity talk and clishmaclatter."
"He speaks the plain truth." Nikobo pressed her huge snout close to
the bars. "Even I can detect the signs of royalty in this fat and
goodly person whom I just this morning helped out of the river. You
must go with them, Tandy, and they will carry you back to your own
Kingdom."
"But I tell you, I'd rather stay here with YOU," wailed the little boy,
relaxing a moment from his kingly and overbearing attitude.
"Roger, fetch the AX." Samuel Salt spoke so loud and sternly Nikobo
lapsed into a shocked silence and Tandy hastily drew back into a far
corner of his cage.
"Never argue with a sea-going man," whispered Ato, winking
solemnly as Roger flew off to obey Samuel's order. Having settled
the matter in his own mind, Samuel turned his back on Tandy and
began to examine with deep interest the fungus growth on one of
the gnarled old trees. "So you really are a King?" Leaning against the
huge body of Nikobo, Ato folded his hands comfortably on his
stomach and regarded the boy in the leopard skin earnestly. "Now
what country do you hail from and what do they call you at home?"
"I am Tazander Tazah of Ozamaland," announced the boy proudly,
"the land of the creeping bird and flying reptile. Ozamaland on the
long continent of Tarara is my home."
"OZAMALAND!" shouted Samuel Salt, swinging round like a
teetotum. "So there really IS such a place. I have always said so,
Ato, but no one would believe me. Lies to the east of here, doesn't
it, sonny, and is twice as large as any known land bordering on the
Nonestic?" Somewhat impressed to find that Samuel Salt knew
something of his homeland, the little boy nodded. "And do you
suppose we could snare one of those creeping birds and flying
reptiles if we managed to reach Ozamaland?" Grasping the bars of
the cage, Samuel peered anxiously into the young King's face.
"Do you suppose you ever could reach Ozamaland?" sighed
Tazander, returning Samuel's eager look with gloomy aloofness. "Do
you know that a ship has never touched our shores?"
"Then the Crescent Moon shall be the first!" cried Samuel Salt,
snapping his fingers joyfully. "Why, this will be tremendous and the
most momentous discovery in a thousand years! But how do you
happen to be so far from Ozamaland yourself?" asked Samuel Salt
immediately afterward. "Did you come by air or sea?"
"That I cannot tell." Tazander seated himself soberly on a log before
he continued. "One night I was sleeping soundly in my tower in the
White City, next thing I remember I was here in this jungle. The
Leopard Men, wild and savage as they were, fed me when they
remembered on raw fish and chunks of hard, bitter bread they made
from the roots of the Brima Tree. But I could not understand their
talk, nor they mine, and till Kobo found me a month after my
imprisonment I had no one to talk to at all. But she has come every
day to keep me company and try to set me free, and since the
Leopard Men were drowned she has fed me, too. See, through this
little door." Tazander opened a small door in the bars and stuck both
hands through.

"But how did you learn the language?" asked Ato, turning round to
gaze up into Nikobo's huge face.
"I don't know," said Nikobo with an excited gulp. "I just started to
say 'Hello!' and instead of saying it in hippopotamy, there I was
talking a strange language which I could understand as well as my
own. And in this language Tandy answered me, much to my delight
and pleasure."
"Strange, very strange." Ato shook his head in a puzzled manner.
"Well, all I say is, it was lucky for this small fellow that you
happened along, and once we have him aboard he'll soon forget all
these hardships and unpleasant experiences."
"I'll never forget Kobo," said the young King, backing stiffly away
from the outstretched arms of Ato.
"And Kobo'll never forget YOU," sniffed the hippopotamus. "The talk
of the river people seems dull and stupid since I've talked to Tandy.
None of the herd really need me and I don't know what I'm going to
do—whoo—Hoo HOO WHOOO!" Rocking from side to side, Nikobo
began to sob as if her heart would break, so violently in fact, Samuel
Salt covered both ears and Ato, alarmed at the enormous grief of
the gigantic beast, tried to put his arms around her.
"Here, here!" begged the ship's cook, thumping her hard upon the
back. Opening the bag of biscuits Roger had brought from the ship,
Ato handed two to Tandy and began shoving the rest as fast as he
could down the vast throat of the grief-stricken hippopotamus. After
each biscuit, Nikobo choked and sobbed to herself, but on the whole,
they seemed to comfort her, and when the Read Bird finally returned
with the ax she watched almost cheerfully as Samuel Salt, with well-
aimed blows, demolished Tandy's jungle cage. As the last side
crashed down and without giving Tandy time to argue any further,
Samuel Salt seized the boy firmly in both arms and set him down on
the back of the hippopotamus. Then, giving Ato a hand up behind
him, the Captain of the Crescent Moon sternly led the way to the
edge of the island. Roger, waving an Oz flag, flew ahead screaming
defiantly to the monkeys and parrots that infested the island, "WAY,
WAY! Way for the Royal Discoverer of Oz! Way for the King of the
Octagon Isle! Way for Nikobo, Little Daughter of the Biggenlittle
River People. Way for Tazander Tazah, King and son of a King's son!
WAY—ay—ayyyy!"
 CHAPTER 7
Strange Specimens for Samuel Salt

With no one to challenge their going but the birds and monkeys, the
little band made its way back to the sandy beach. Tandy, perhaps
because he had been so long pent up in the silent jungle and
because he was by nature a naturally sober and solemn little boy,
said nothing. Not even the Crescent Moon, riding so proudly at her
anchor, seemed to arouse any interest or enthusiasm in this strange
young Ozamalander.
"Well, here we are!" exclaimed Ato, heartily thankful to be in sight of
the ship again. "And I hope you'll not mind ferrying us out to the
boat, Nikobo; those crocodiles still look hungry and I've no notion of
being crocked for the rest of my life."
"Any time you say," grunted the hippopotamus, squeaking a listless
greeting to a company of her own relatives who were rolling lazily
about in the muddy river water.
"Avast and belay and what's the hurry?" Leaning his ax against a
tree, Samuel moistened a finger and held it up. "The wind's against
us, Mate, so we'll have to wait for the tide. Not only that, but Roger
and I must survey the island and dig up some more interesting
specimens to take back to the ship." After a long and rather quizzical
look at Tandy, Samuel turned and swung along the beach, the Read
Bird flapping joyously behind him.
"Run up and down a bit," advised Ato, sliding down from Nikobo's
back. "Your legs must need stretching. Wonder if there's anything to
eat around here or hereabouts? Aha, those look like oranges, a wild
orange grove, as I'm a cook and a seaman. Come along, young one,
and help me gather a few."
"A King and son of a King's son does not come and go at another's
bidding," announced Tandy, stiffly, alighting from the hippopotamus.
"Merciful mothers! What's this?" gasped Ato, blinking his eyes
rapidly. "As complete a case of ingrowing Royalitis as I've ever had
the misfortune to encounter. Well, since it's every King for himself,
then I'll be leaving you, sonny and son of a King's sonny. Watch out
for him, Kobo, he's probably real important to himself."
"You should not speak like that," reproved the hippopotamus as Ato
disappeared into the orange grove, "after all, the big and fat one is
himself a King."
"Pooh, King of some potty little island," sniffed Tandy, leaning
wearily against a palm. "Break me a cocoanut, Kobo, I'm thirsty."
With a discouraged sigh Nikobo trod on one of the cocoanuts,
cracking it from end to end and then, because she was a generous
and kindly creature, she cracked several more for Ato when he
should return. Sitting back on her haunches, she anxiously watched
while Tandy downed the cocoanut milk, then, stretching out in the
sand, fell unconcernedly asleep. Thus Ato found them when he
emerged from the orange grove an hour later. His elegant explorer's
cape was knotted to form a sack and bursting full of the small sweet
fruit of the wild orange trees.
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookmasss.com