100% found this document useful (2 votes)
18 views

Building Responsible AI Algorithms: A Framework for Transparency, Fairness, Safety, Privacy, and Robustness 1st Edition Toju Duke 2024 scribd download

AI

Uploaded by

sandriagdan
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
100% found this document useful (2 votes)
18 views

Building Responsible AI Algorithms: A Framework for Transparency, Fairness, Safety, Privacy, and Robustness 1st Edition Toju Duke 2024 scribd download

AI

Uploaded by

sandriagdan
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 55

Download Full Version ebook - Visit ebookmeta.

com

Building Responsible AI Algorithms: A Framework


for Transparency, Fairness, Safety, Privacy, and
Robustness 1st Edition Toju Duke

https://ebookmeta.com/product/building-responsible-ai-
algorithms-a-framework-for-transparency-fairness-safety-
privacy-and-robustness-1st-edition-toju-duke/

OR CLICK HERE

DOWLOAD NOW

Discover More Ebook - Explore Now at ebookmeta.com


Instant digital products (PDF, ePub, MOBI) ready for you
Download now and discover formats that fit your needs...

Start reading on any device today!

Cambridge IGCSE and O Level History Workbook 2C - Depth


Study: the United States, 1919-41 2nd Edition Benjamin
Harrison
https://ebookmeta.com/product/cambridge-igcse-and-o-level-history-
workbook-2c-depth-study-the-united-states-1919-41-2nd-edition-
benjamin-harrison/
ebookmeta.com

Responsible AI Implementing Ethical and Unbiased


Algorithms 1st Edition Sray Agarwal

https://ebookmeta.com/product/responsible-ai-implementing-ethical-and-
unbiased-algorithms-1st-edition-sray-agarwal/

ebookmeta.com

Responsible AI 1st Edition Patrick Hall

https://ebookmeta.com/product/responsible-ai-1st-edition-patrick-hall/

ebookmeta.com

Aptitude Learning and Instruction Volume 3 Conative and


Affective Process Analyses Richard E. Snow (Editor)

https://ebookmeta.com/product/aptitude-learning-and-instruction-
volume-3-conative-and-affective-process-analyses-richard-e-snow-
editor/
ebookmeta.com
Practical Trend Analysis Applying Signals and Indicators
to Improve Trade Timing 2nd Edition Michael C. Thomsett

https://ebookmeta.com/product/practical-trend-analysis-applying-
signals-and-indicators-to-improve-trade-timing-2nd-edition-michael-c-
thomsett/
ebookmeta.com

Terrorism in the Late Victorian Novel 1st Edition Barbara


Arnett Melchiori

https://ebookmeta.com/product/terrorism-in-the-late-victorian-
novel-1st-edition-barbara-arnett-melchiori/

ebookmeta.com

The Work of Life Writing Essays and Lectures 1st Edition


G. Thomas Couser

https://ebookmeta.com/product/the-work-of-life-writing-essays-and-
lectures-1st-edition-g-thomas-couser/

ebookmeta.com

Weather Studies Investigations Manual 2021 22 Summer 2022


American Metorological Society

https://ebookmeta.com/product/weather-studies-investigations-
manual-2021-22-summer-2022-american-metorological-society/

ebookmeta.com

Naruto: A Mythical and Religious Analysis 2nd Edition


Anonyous

https://ebookmeta.com/product/naruto-a-mythical-and-religious-
analysis-2nd-edition-anonyous/

ebookmeta.com
Hijab The Islamic Commandments of Hijab Dr Mohammad Ismail
Memon Madani

https://ebookmeta.com/product/hijab-the-islamic-commandments-of-hijab-
dr-mohammad-ismail-memon-madani/

ebookmeta.com
Building
Responsible AI
Algorithms
A Framework for Transparency,
Fairness, Safety, Privacy, and
Robustness

Toju Duke
Building Responsible
AI Algorithms
A Framework for Transparency,
Fairness, Safety, Privacy,
and Robustness

Toju Duke
Building Responsible AI Algorithms: A Framework for Transparency,
Fairness, Safety, Privacy, and Robustness
Toju Duke
London, UK

ISBN-13 (pbk): 978-1-4842-9305-8 ISBN-13 (electronic): 978-1-4842-9306-5


https://doi.org/10.1007/978-1-4842-9306-5

Copyright © 2023 by Toju Duke


This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or
part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way,
and transmission or information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed.
Trademarked names, logos, and images may appear in this book. Rather than use a trademark
symbol with every occurrence of a trademarked name, logo, or image we use the names, logos,
and images only in an editorial fashion and to the benefit of the trademark owner, with no
intention of infringement of the trademark.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if
they are not identified as such, is not to be taken as an expression of opinion as to whether or not
they are subject to proprietary rights.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal
responsibility for any errors or omissions that may be made. The publisher makes no warranty,
express or implied, with respect to the material contained herein.
Managing Director, Apress Media LLC: Welmoed Spahr
Acquisitions Editor: Jonathan Gennick
Development Editor: Laura Berendson
Editorial Assistant: Shaul Elson
Copy Editor: Kezia Endsley
Cover image by Valery Fedotov on Unsplash (www.unsplash.com)
Distributed to the book trade worldwide by Springer Science+Business Media LLC, 1 New York
Plaza, Suite 4600, New York, NY 10004. Phone 1-800-SPRINGER, fax (201) 348-4505, e-mail
orders-ny@springer-sbm.com, or visit www.springeronline.com. Apress Media, LLC is a
California LLC and the sole member (owner) is Springer Science + Business Media Finance Inc
(SSBM Finance Inc). SSBM Finance Inc is a Delaware corporation.
For information on translations, please e-mail booktranslations@springernature.com; for
reprint, paperback, or audio rights, please e-mail bookpermissions@springernature.com.
Apress titles may be purchased in bulk for academic, corporate, or promotional use. eBook
versions and licenses are also available for most titles. For more information, reference our Print
and eBook Bulk Sales web page at http://www.apress.com/bulk-sales.
Any source code or other supplementary material referenced by the author in this book is
available to readers on GitHub (https://github.com/Apress). For more detailed information,
please visit http://www.apress.com/source-code.
Printed on acid-free paper
To Emma and Alex, for teaching me what it means to be
human and responsible.
Table of Contents
About the Author���������������������������������������������������������������������������������xi

About the Technical Reviewer�����������������������������������������������������������xiii

Introduction����������������������������������������������������������������������������������������xv

Part I: Foundation������������������������������������������������������������������������1
Chapter 1: Responsibility���������������������������������������������������������������������3
Avoiding the Blame Game�������������������������������������������������������������������������������������4
Being Accountable������������������������������������������������������������������������������������������������6
Eliminating Toxicity�����������������������������������������������������������������������������������������������9
Thinking Fairly����������������������������������������������������������������������������������������������������11
Protecting Human Privacy�����������������������������������������������������������������������������������12
Ensuring Safety���������������������������������������������������������������������������������������������������13
Summary������������������������������������������������������������������������������������������������������������14

Chapter 2: AI Principles����������������������������������������������������������������������15


Fairness, Bias, and Human-Centered Values������������������������������������������������������16
Google�����������������������������������������������������������������������������������������������������������17
The Organisation for Economic Cooperation and Development (OECD)��������17
The Australian Government���������������������������������������������������������������������������18
Transparency and Trust���������������������������������������������������������������������������������������20
Accountability�����������������������������������������������������������������������������������������������������22

v
Table of Contents

Social Benefits����������������������������������������������������������������������������������������������������23
Privacy, Safety, and Security�������������������������������������������������������������������������������31
Summary������������������������������������������������������������������������������������������������������������35

Chapter 3: Data�����������������������������������������������������������������������������������37
The History of Data���������������������������������������������������������������������������������������������38
Data Ethics����������������������������������������������������������������������������������������������������������39
Ownership�����������������������������������������������������������������������������������������������������40
Data Control���������������������������������������������������������������������������������������������������41
Transparency�������������������������������������������������������������������������������������������������42
Accountability������������������������������������������������������������������������������������������������42
Equality����������������������������������������������������������������������������������������������������������43
Privacy�����������������������������������������������������������������������������������������������������������43
Intention��������������������������������������������������������������������������������������������������������43
Outcomes������������������������������������������������������������������������������������������������������44
Data Curation������������������������������������������������������������������������������������������������������44
Best Practices�����������������������������������������������������������������������������������������������������45
Annotation and Filtering��������������������������������������������������������������������������������45
Rater Diversity�����������������������������������������������������������������������������������������������48
Synthetic Data�����������������������������������������������������������������������������������������������49
Data Cards and Datasheets���������������������������������������������������������������������������49
Model Cards��������������������������������������������������������������������������������������������������50
Tools��������������������������������������������������������������������������������������������������������������������56
Alternative Datasets��������������������������������������������������������������������������������������������57
Summary������������������������������������������������������������������������������������������������������������58

vi
Table of Contents

Part II: Implementation��������������������������������������������������������������59


Chapter 4: Fairness����������������������������������������������������������������������������61
Defining Fairness������������������������������������������������������������������������������������������������62
Equalized Odds����������������������������������������������������������������������������������������������64
Equal Opportunity������������������������������������������������������������������������������������������64
Demographic Parity���������������������������������������������������������������������������������������64
Fairness Through Awareness������������������������������������������������������������������������64
Fairness Through Unawareness��������������������������������������������������������������������65
Treatment Equality�����������������������������������������������������������������������������������������65
Test Fairness�������������������������������������������������������������������������������������������������65
Counterfactual Fairness��������������������������������������������������������������������������������65
Fairness in Relational Domains���������������������������������������������������������������������65
Conditional Statistical Parity�������������������������������������������������������������������������66
Types of Bias�������������������������������������������������������������������������������������������������������67
Historical Bias�����������������������������������������������������������������������������������������������68
Representation Bias��������������������������������������������������������������������������������������69
Measurement Bias�����������������������������������������������������������������������������������������69
Aggregation Bias�������������������������������������������������������������������������������������������70
Evaluation Bias����������������������������������������������������������������������������������������������70
Deployment Bias��������������������������������������������������������������������������������������������71
Measuring Fairness��������������������������������������������������������������������������������������������71
Fairness Tools�����������������������������������������������������������������������������������������������������73
Summary������������������������������������������������������������������������������������������������������������75

Chapter 5: Safety��������������������������������������������������������������������������������77
AI Safety��������������������������������������������������������������������������������������������������������������78
Autonomous Learning with Benign Intent�����������������������������������������������������78
Human Controlled with Benign Intent������������������������������������������������������������79

vii
Table of Contents

Human Controlled with Malicious Intent�������������������������������������������������������79


AI Harms�������������������������������������������������������������������������������������������������������������80
Discrimination, Hate Speech, and Exclusion�������������������������������������������������81
Information Hazards��������������������������������������������������������������������������������������82
Misinformation Harms�����������������������������������������������������������������������������������83
Malicious Uses����������������������������������������������������������������������������������������������83
Human-Computer Interaction Harms������������������������������������������������������������84
Environmental and Socioeconomic Harms����������������������������������������������������85
Mitigations and Technical Considerations�����������������������������������������������������������87
Benchmarking�����������������������������������������������������������������������������������������������������90
Summary������������������������������������������������������������������������������������������������������������93

Chapter 6: Human-in-the-Loop�����������������������������������������������������������95
Understanding Human-in-the-Loop��������������������������������������������������������������������95
Human Annotation Case Study: Jigsaw Toxicity Classification���������������������������96
Rater Diversity Case Study: Jigsaw Toxicity Classification���������������������������������98
Task Design���������������������������������������������������������������������������������������������������99
Measures�����������������������������������������������������������������������������������������������������100
Results and Conclusion�������������������������������������������������������������������������������100
Risks and Challenges����������������������������������������������������������������������������������102
Summary�����������������������������������������������������������������������������������������������������103

Chapter 7: Explainability������������������������������������������������������������������105
Explainable AI (XAI)�������������������������������������������������������������������������������������������106
Implementing Explainable AI�����������������������������������������������������������������������������107
Data Cards���������������������������������������������������������������������������������������������������107
Model Cards������������������������������������������������������������������������������������������������109
Open-Source Toolkits����������������������������������������������������������������������������������110
Accountability����������������������������������������������������������������������������������������������111

viii
Table of Contents

Dimensions of AI Accountability������������������������������������������������������������������������112
Governance Structures��������������������������������������������������������������������������������112
Data�������������������������������������������������������������������������������������������������������������112
Performance Goals and Metrics������������������������������������������������������������������112
Monitoring Plans�����������������������������������������������������������������������������������������113
Explainable AI Tools�������������������������������������������������������������������������������������113
Summary�����������������������������������������������������������������������������������������������������116

Chapter 8: Privacy����������������������������������������������������������������������������117
Privacy Preserving AI����������������������������������������������������������������������������������������118
Federated Learning�������������������������������������������������������������������������������������119
Differential Privacy��������������������������������������������������������������������������������������121
Summary����������������������������������������������������������������������������������������������������������123

Chapter 9: Robustness����������������������������������������������������������������������125
Robust ML Models��������������������������������������������������������������������������������������������126
Sampling�����������������������������������������������������������������������������������������������������126
Bias Mitigation (Preprocessing)�������������������������������������������������������������������127
Data Balancing��������������������������������������������������������������������������������������������127
Data Augmentation��������������������������������������������������������������������������������������128
Cross-Validation������������������������������������������������������������������������������������������128
Ensembles���������������������������������������������������������������������������������������������������128
Bias Mitigation (In-Processing and Post-­Processing)���������������������������������129
Transfer Learning����������������������������������������������������������������������������������������129
Adversarial Training�������������������������������������������������������������������������������������129
Making Your ML Models Robust������������������������������������������������������������������130
Model Challenges����������������������������������������������������������������������������������������������132
Data Quality�������������������������������������������������������������������������������������������������132
Model Decay������������������������������������������������������������������������������������������������132

ix
Table of Contents

Feature Stability������������������������������������������������������������������������������������������133
Precision versus Recall�������������������������������������������������������������������������������133
Input Perturbations��������������������������������������������������������������������������������������134
Summary����������������������������������������������������������������������������������������������������������134

Part III: Ethical Considerations������������������������������������������������135


Chapter 10: AI Ethics������������������������������������������������������������������������137
Ethical Considerations for Large Language Models������������������������������������������139
Prevalent Discriminatory Language in LLMs�����������������������������������������������139
Working with Crowdworkers�����������������������������������������������������������������������140
Inequality and Job Quality���������������������������������������������������������������������������140
Impact on Creatives�������������������������������������������������������������������������������������141
Disparate Access to Language Model Benefits�������������������������������������������141
Ethical Considerations for Generative Models��������������������������������������������������142
Deepfake Generation�����������������������������������������������������������������������������������143
Truthfulness, Accuracy, and Hallucinations�������������������������������������������������143
Copyright Infringement��������������������������������������������������������������������������������145
Ethical Considerations for Computer Vision������������������������������������������������������145
Issues of Fraud��������������������������������������������������������������������������������������������146
Inaccuracies������������������������������������������������������������������������������������������������146
Consent Violations���������������������������������������������������������������������������������������146
Summary����������������������������������������������������������������������������������������������������������147

Appendix A: References��������������������������������������������������������������������149

Index�������������������������������������������������������������������������������������������������179

x
About the Author
Toju Duke with over 18 years experience
spanning across Advertising, Retail, Not-For
Profit and Tech, Toju is a popular speaker,
author, thought leader and consultant on
Responsible AI. Toju spent 10 years at Google
where she spent the last couple of years as
a Programme Manager on Responsible AI
leading various Responsible AI programmes
across Google’s product and research teams
with a primary focus on large-scale models
and Responsible AI processes. Prior to her time
spent on Google’s research team, Toju was the
EMEA product lead for Google Travel and worked as a specialist across a
couple of Google’s advertising products during her tenure. She is also the
founder of Diverse AI, a community interest organisation with a mission
to support and champion underrepresented groups to build a diverse and
inclusive AI future. She provides consultation and advice on Responsible
AI practices worldwide.

xi
About the Technical Reviewer
Maris Sekar is a professional computer
engineer, senior data scientist (Data Science
Council of America), and certified information
systems auditor (ISACA). He has a passion
for using storytelling to communicate about
high-risk items in an organization to enable
better decision making and drive operational
efficiencies. He has cross-functional work
experience in various domains, including risk
management, oil and gas, and utilities. Maris has led many initiatives for
organizations, such as PricewaterhouseCoopers LLP, Shell Canada Ltd.,
and TC Energy. Maris’ love for data has motivated him to win awards, write
articles, and publish papers on applied machine learning and data science.

xiii
Introduction
I’ve always been a huge fan of technology and innovation and a great
admirer of scientists and inventors who pushed the boundaries of
innovation, some trying 99 times, 10,000 times, and more before achieving
their goals and making history. Take the famous inventor of the lightbulb,
Thomas Edison, or the brilliant Grace Hopper, who invented the
computer. And before computers were transformed into machines, we had
human computers, such as the super-intelligent “hidden figures,” Mary
Jackson, Katherine Johnson, and Dorothy Vaughan of NASA (National
Aeronautics and Space Administration). We also have amazing geniuses
like Albert Einstein, whose theories on relativity introduced many new
ways of evaluating energy, time, space, gravity, and matter. Or the likes of
Alexander Graham Bell, who introduced the telephone, and Josephine
Cochrane, who we should thank for saving us from washing dishes by
hand and invented the ubiquitous dishwasher!
These are just a few innovators and inventors who contributed
greatly to technology, made our lives better and easier, and shed light on
unknown phenomena. And there are many other sung and unsung heroes
who contributed greatly to the world of science and technology.
Fast forward to today, to an ever-changing and evolving world:
Humans are still inventing, creating, and introducing breakthroughs,
especially in the field of technology. Many recent inventions are driven
by artificial intelligence (AI), which is made up of deep learning networks
(a form of AI based on neural networks, designed to mimic neurons in
the human brain). For example, ChatGPT (a conversational AI built on a
large language model), which is designed to provide intelligent answers
to questions and solve many difficult tasks, has become the world’s fastest

xv
Introduction

growing app, with over 100 million users in just months. It still blows my
mind how “intelligent” this app, and other dialogue AI systems similar to
it, is. Another example is the various image recognition AI systems, which
are used across the healthcare, automotive, criminal justice, agriculture,
and telecommunications industries. We also have voice assistants such
as Siri, Google Assistant, and Alexa, speech recognition systems, and so
on. There’s also DeepMind’s (a UK-based AI company) Alphafold, which
predicts 3D models of protein structures, contributing immensely to
the medical field and driving further drug development and discovery.
Alphafold solved a long-standing problem in the history of biology and
medical science.
While we have these and so many more amazing use cases and
success stories of AI and machine learning (ML) applications/products,
it’s important to note that there are also fundamental issues that plague
these technologies. These range from bias, toxicity, harm, hate speech,
misinformation, privacy, human rights violations, and sometimes the loss
of life, to mention a few. Although AI technologies are great and highly
beneficial to society in various ways, AI sometimes produces harm due
to the lack of representative and diverse training data, lack of data ethics
and curation best practices, less than optimal fine-tuning and training
methods, and the sometimes harmful ways these AI applications are used.
In this book, I cover some examples where AI has gone drastically
wrong and affected people’s lives in ways that had a ripple effect on various
groups and communities. Despite these various downfalls, I believe that
AI has the potential to solve some of the world’s biggest problems, and it
is being used in various ways to tackle long-standing issues like climate
change, as an example, by a good number of organizations. While we have
many well-meaning individuals developing these highly “intelligent”
machines, it’s important to understand the various challenges faced by
these systems and humanity at large and explore the possible ways to
address, resolve, and combat these problems.

xvi
Introduction

When tackling these issues across different dimensions, there’s a


potential framework for responsible AI that's worth adopting. This begins
with understanding the gravity and importance of being responsible
and accountable for models, products, and applications that could harm
many people and communities, whether psychologically, emotionally, or
physically; drive further inequality, leading to more poverty; violate human
rights and privacy, and so on. A responsible AI framework can be built
from the definition of principles that guide AI development. These include
AI principles that address issues with data, incorporate fairness and its
metrics, consider AI safety, take into account privacy, and build robust
ML models.
In this book, I propose a simple, fundamental responsible AI
framework that can lead to the development and deployment of ML
technologies that have the safety and well-being of its end users, people,
and customers in mind. Implementing further research on these different
areas covered in this book can lead to less harm and bias across these
technologies.
Understandably, it’s difficult to measure the harm and damage
that has been caused by these systems; it’s far easier, on the other
hand, to measure instances where AI has been beneficial and adopted
across several industries. But given the challenges with this incredibly
impressive technology, and the rapid rate of adoption in recent times,
it’s critical that we understand the benefits of AI as well as its limitations,
risks, and challenges. My hope is that this book sheds some light on
the incredibly positive potential of AI, but also on its limitations and
shortcomings. The book guides you through potential ways to address
these challenges, keeping humanity, users, consumers, and society in
mind. I’ll end with this note: being an adult is great, but with adulthood,
comes great responsibility. This responsibility should be considered in the
development of technologies that will not only be used and adopted by
billions of people across the world, but that could potentially pose risks to
these same people in many different ways.

xvii
PART I

Foundation
CHAPTER 1

Responsibility
The term responsibility is a relatable and simple term. Everyone, or almost
everyone, deems themselves to be responsible in most if not every area of
their lives. There’s a sense of fulfillment and gratification when you think
you have carried out a responsible act. Being responsible refers to carrying
out a duty or job that you’re charged with.1 Most people in positions of
authority feel a sense of responsibility to execute their jobs effectively.
This includes parents, lawyers, law enforcement officers, healthcare
professionals, members of a jury, employees, employers, and every
member of society who has reached decision-making age. This chapter
delves into AI responsibility and the need for building responsible AI
systems.
Despite the fact that we were encouraged to be responsible at a
very early age, it’s often not accounted for in technology fields, and in
particular in machine learning (ML) subfields, such as natural language
processing, computer vision, deep learning, neural networks, and so on.
Now you might argue that this isn’t entirely true. There are ethical artificial
intelligence (AI) and responsible AI practices developed every day.
Although ethics and responsibility have been long-standing conversations
that have taken place over the years, it’s only recently, within the last 2-5
years, that we’ve seen an uptick in the adoption of responsible AI practices
across industries and research communities. We have also seen more
interest, from policy makers and regulatory bodies, in ensuring that AI is
human-centric and trustworthy.2

© Toju Duke 2023 3


T. Duke, Building Responsible AI Algorithms,
https://doi.org/10.1007/978-1-4842-9306-5_1
Chapter 1 Responsibility

There are several reasons that responsible AI has been slowly adopted
over the years, the most prominent being that it’s a new field that’s slowly
gaining recognition across the various AI practitioners. It’s a bit sad that
responsible and ethical practices were not adopted at scale, despite the
66 plus years of AI introduction.3 Taking some cues from mental health
experts, let’s look at a few recommendations for acting responsibly.

Avoiding the Blame Game


Did you ever do something wrong with a sibling or a friend when you were
a kid, and then tried to blame your accomplice once you were caught? My
kids do this to me all the time. “I didn’t do it; she did!” And the younger,
not-so-innocent four year old, retorts, “No! It wasn’t me! She did it!” Many
people believe it’s an inherent part of human nature to blame someone
else when things go wrong. This could be a result of learned behavior and
a desire to avoid punishment. It might also stem from a desire to seek
parental approval. Research shows that adults who are quicker to blame
others are more likely to have experienced some form of trauma in their
lives that they haven’t dealt with.
Before we go down the rabbit hole to the causes of the “blame culture,”
it’s safe to say that people are often quick to blame someone/something else
when things go wrong. This is something people need to unlearn in order to
be and act responsible. When we have a case involving harmful ML models
or products, the first question we tend to ask is “who’s to blame?”
In site reliability engineering (SRE), it’s a known fact that failure is
part of the development process and is likely to happen quite often. A
key part of the engineering process is to have a post-mortem, which is a
written document that provides details of the incident, the magnitude of
the incident, and the actions taken to resolve it, including its root cause.
The document also includes follow-up actions/suggestions to ensure the
problem doesn’t happen again.

4
Chapter 1 Responsibility

The problem with the blame culture is that it tends to negatively focus
on people, which consequently prevents the right lessons to be learned—
what caused the problem and how to prevent it from happening again. In
light of this, a blameless post-mortem helps engineering teams understand
the reasons an incident occurred, without blaming anyone or any team in
particular. This, in turn, enables the teams to focus on the solution rather
than the problem. The key focus is to understand the measures to put in
place to prevent similar incidents from happening.4
If you’ve been working in the AI field long enough, particularly ethical
AI, you’ve heard of the infamous “trolley problem,” a series of experiments
in ethics and psychology made up of ethical dilemmas. This problem
asks whether you would sacrifice one person to save a larger number of
people.5 In 2014, an experiment called the Moral Machine was developed
by researchers at the MIT Media Lab. The Moral Machine was designed
to crowdsource people’s decisions about how self-driving cars should
prioritize lives in different variations of the trolley problem.6
In a (paraphrased) scenario where a self-driving car’s brakes fail and
it has two passengers onboard when approaching a zebra crossing with
five pedestrians walking across—an elderly couple, a dog, a toddler, and
a young woman—who should the car hit? Should the car hit the elderly
couple and avoid the other pedestrians, or should it hit the little boy? Or
should the car swerve and avoid the pedestrians but potentially harm the
passengers? In other words, should it take action or stay on course? Which
lives matter more—humans versus pets, passengers versus pedestrians,
young versus old, fit over sickly, higher social status versus lower, cisgender
versus non-binary?
In cases in which AI-related failures led to injury or death, I believe
everyone who was involved in the development of the offending vehicle
should be held accountable. That is, from the research scientist, to the
engineer, to the CTO, to the legal officer, to marketing, public relations, and
so on. It’s the responsibility of everyone involved in the development of

5
Chapter 1 Responsibility

ML technologies, regardless of their role, to think responsibly during the


development of the models and products and to think about the impact it
could have on the end users, the consumers, and society.

Being Accountable
When people accept accountability, it means they understand their
contribution to a given situation. Being accountable also means avoiding
the same mistakes over and over again. In some cases, this requires
giving an account or statement about the part the person had to play.
Not so surprisingly, accountability is a key component of responsible
AI. According to the Organisation for Economic Cooperation and
Development (OECD), companies and individuals developing, deploying,
and operating AI systems should be held accountable for their proper
functioning in line with the OECD’s values-based principle of AI.7 Chapter 2
delves further into the topic of responsible AI principles.
Let’s look at a couple of examples where AI drastically impacted the
lives of certain members of society, and the responsible companies were
held accountable. Before delving into these stories, I’d like to take a pit stop
and state that I’m a huge advocate and supporter of AI. I strongly believe
that AI has the potential to solve some of the world’s most challenging
problems, ranging from climate change, to healthcare issues, to education,
and so on. AI has also been adopted in several projects for good, otherwise
known as AI for Social Good.
For example, top tech companies such as Google, Microsoft, IBM, and
Intel are working on projects ranging from environmental protection to
humanitarian causes, to cancer diagnostics and treatment, and wildlife
conversation challenges, among others.8
AI also has many business benefits, including reducing operational
costs, increasing efficiency, growing revenue, and improving customer
experiences. The global AI market size was valued at $93.5 billion in 2021,

6
Chapter 1 Responsibility

and it is projected to expand at a compound annual growth rate (CAGR) of


38.1 percent between 2022 to 2030.9 This shows that AI has great potential
to improve businesses and people’s lives while tackling major, long-­
standing problems the world faces.
Now that you’ve learned that AI is a much needed technology, let’s
consider the story of Williams Isaac, a young man from Michigan who
received a call from the Detroit Police Department to report to a police
station while working in his office at an automotive supply company on
a Thursday afternoon. Thinking it was a prank, he ignored the summons
and drove home. Upon driving into his driveway, a police car pulled up
and blocked him in. Williams was handcuffed and arrested in front of his
wife and two young daughters. When his wife, Melissa, asked where he was
being taken, she was told by one of the officers to “Google it.”
Mr. Williams was driven to a detention center, where he had his
mugshot, fingerprints, and DNA taken; he was held overnight. Upon
interrogation the next day, after a series of conversations with two
detectives and a surveillance video shown of the criminal in question, they
discovered that they had the wrong man, due to a flawed match from a
facial recognition algorithm.
Although facial recognition systems have been used by police forces
across the world for more than 20 years, recent studies have found that
facial recognition works best on Caucasian men. The results are less
accurate for other demographics, mainly due to the lack of diversity in
the images gained from the underlying datasets.10 Facial recognition
systems are quite flawed, and people of all backgrounds have been falsely
arrested,11 although this more greatly affects people of color.
Let’s take a step back and think about the effect this false arrest could
have had on Williams, his wife, and his two little girls, who watched their
dad get arrested right before their eyes. Putting this into context, the
psychological effects of the night spent in jail could have been grave. There
are also potential emotional and physical effects. I can only imagine the
confusion and anger Williams must have felt that night, wondering what

7
Chapter 1 Responsibility

could have led to his false arrest. Knowing that the mere fact that he’s a
black man living in the United States makes him an easy target for the
police, Williams may not have been altogether surprised, but he must
have been quite saddened and anxious, hoping he’d return to his family in
good time.
In a world where racism and discrimination still very much exists,
it’s quite appalling to see these societal issues prevalent in technologies
employed and used by people in authority. These people of authority are
the same individuals who are employed to protect our communities. If
they decide to use technology and AI systems in their jobs, it’s their duty to
ensure that these systems promote fairness and equal treatment of people
from different backgrounds, cultures, ethnicities, races, disabilities, social
classes, and socio-economic statuses.
The law enforcement agency that committed the blunder apologized
in a statement,9 stating that Williams could have the case and fingerprint
data expunged. When we consider “accountability” and what it entails, an
apology and removal of his record is not enough. What the county needs
to do to is make sure this sort of life-changing error doesn’t happen again.
They need to aim for clean data, run tests across the different subgroups
for potential biases, and maintain transparency and explainability
by documenting information on the data and the model, including
explanations of the tasks and objectives the model was designed for.
Carrying out accuracy and error checks will also help ensure results are
accurate and less biased.
Some facial recognition software has been banned in certain use cases
in the United States, including in New York, which passed a law in 2021
prohibiting facial recognition at schools, and in California, which passed
a law that banned law enforcement from using facial recognition on their
body cameras. Maryland also passed a law that prohibits the use of facial
recognition during interviews without signed consent.12 Despite this

8
Chapter 1 Responsibility

progress, there has been a steady increase in states recalling their bans on
facial recognition; for example, Virginia recently eliminated its prohibition
on police use of facial recognition, only one year after approving the ban.13
I’m happy to state that a few tech companies—Google and most
recently Microsoft, Amazon, and IBM—stopped selling facial recognition
technology software to police departments and have called for federal
regulation of the technology.14
Across the globe, there are only two countries in the world that have
banned the use of facial recognition—Belgium and Luxembourg.15 In
Europe, the draft EU AI act released in April 2021 aims to limit the use
of biometric identification systems, including facial recognition.16 While
some parts of the world are still deliberating on how they’ll use facial
recognition software, it’s encouraging to see there are countries, regulatory
bodies, and organizations that recognize the dangers of facial recognition
technologies and are ready to hold businesses accountable.

Eliminating Toxicity
Distancing yourself from people who exhibit toxic traits is advice that
mental health practitioners provide to anyone seeking guidance about
responsibility. By the same token, removing toxicity from ML models is
a fundamental tenant of responsible AI. As datasets are built from the
Internet, which certainly includes human data and biases, these datasets
tend to have toxic terms, phrases, images, and ideas embedded in them.
It’s important to note that “toxicity” is contextual. What one person regards
as toxic another might not, depending on their community, beliefs,
experiences, and so on. In this context, toxic refers to the way the model is
used; it’s “toxic” when used in a harmful manner.

9
Chapter 1 Responsibility

Since ML models and products need datasets to work, the inherent


toxicity and biases prevalent in the datasets tend to creep into pretrained
and trained ML models, eventually appearing in the final products or
models. Let’s take a look at a few examples of toxicity that crept into some
recent NLP models.
Meta (formerly known as Facebook) recently launched BlenderBot 3,
an output of its open-source Large Language Model (LLM) OPT-175B17
that was released in July, 2022. BlenderBot 3 is an AI chatbot that improved
its conversational abilities by searching the Internet to learn about topics.
It conversed with people on these topics, while relying on feedback for
conversational skills and improvements.18
One of the main challenges people face with toxicity in ML models
is due to its representation of human data taken from the Internet, from
platforms such as Reddit, Twitter, Socrata, Facebook, Wikipedia,19 and so
on. If you’ve faced any form of online harassment, or witnessed it in action,
you’ve had a front seat view of how toxic humans can be to other humans.
If we’re building conversational agents using Internet data, it’s going to
be flawed. The data will be flawed, skewed, full of biases, discrimination,
toxicity, lies, and incorrect information. Within a few days of BlenderBot
3’s release, this became clear.
BlenderBot 3 didn’t have high regard for its owner Mark Zuckerberg.
In a different conversation, it stated that Zuckerberg’s company exploits
people for money. I wish I could say that’s the most toxic statement
BlenderBot 3 made. It predictably moved on to racist rhetoric and
conspiracy-based territories. Social media users saw conversations where
the bot denied the results of the 2020 U.S. election, repeated disproven
anti-vaxxer talking points, and even stated that the antisemitic conspiracy
theory that Jewish people controlling the economy is “not implausible.”20
The bot was trained off conversations from the Internet and it mimicked
some public views, many of which aren’t favorable.

10
Chapter 1 Responsibility

As George Fuechsel says, “garbage in, garbage out” (GIGO).21 GIGO is an


expression that says no matter how accurate a program’s logic is, the results
will be incorrect if the input is rubbish. BlenderBot 3 is a clear example of
GIGO. Even so, BlenderBot 3 is considered a much more successful chatbot
than the short-lived Microsoft Tay, which lasted for only 16 hours.
In early 2016, Tay was launched by Microsoft as an innocent AI
chatbot. It was immediately launched into the world of Twitter after
its release, with the aim to learn how to become a good bot. Microsoft
described Tay as an experiment in “conversational understanding.” Tay
was designed to engage and entertain people through casual and playful
conversation as they interacted online.19 It took fewer than 24 hours for
Twitter to corrupt Tay.22 Not long after the bot was launched, people
starting tweeting several misogynistic, racist, and Donald Trump remarks,
which the bot quickly adopted.
Tay’s Twitter conversations reinforced Godwin’s law, which states
that as an online discussion continues, the probability of a comparison
involving the Nazis or Hitler will emerge. There were tweets where Tay
was encouraged to repeat variations of “Hitler was right” and “9/11 was
an inside job.”23 After roughly 16 hours of existence, Tay mysteriously
disappeared from Twitter. Microsoft acted responsibly and prevented the
once-so-innocent, but now turned “toxic parrot bot,” from propagating
further harm, although the reputational damage and harm was done.

Thinking Fairly
In August 2020, hundreds of students in the UK gathered in front of the
Department for Education chanting “swear words” at the algorithm.
Thousands of students in England and Wales had received their “A-level”
exam grades, which were scored by an algorithm. Due to the pandemic
and social distancing measures, the “A-level” exams were cancelled and
the UK’s Office of Qualifications and Examinations Regulation (Ofqual)
decided to estimate the A-level grades using an algorithm.24

11
Chapter 1 Responsibility

The historical grade distribution of schools from three previous years,


the ranking of each student in their school, and previous exam results for a
student per subject were used as input to determine scores.
Experts criticized the low accuracy of the algorithm and public outrage
was on the algorithm’s unfair results. For example, if a student from a given
school didn’t have the highest grades in the past three years (2017-2019),
it was highly unlikely that any student from that school would get a high
grade. Results showed that students from smaller schools, aka private
schools, were more likely to benefit from grade inflation than those from
larger, state schools. This means that students from poorer backgrounds
and socioeconomic backgrounds were the victims of the inaccurate
algorithm.
Within days of the protests, the officials retracted the grades. Once
again, the lack of fairness and equal treatment in the algorithm led to
public confusion and outrage.
Algorithmic fairness is another key component in responsible AI
frameworks, and the algorithm developed by Ofqual failed on this front.
Prior to the release of this algorithm, fairness tests should have been
carried out to ensure all students were treated fairly and equally, with
various tests conducted on the datasets. Thinking fairly is an act of
responsibility and all practitioners working on ML models/AI systems
need to adopt this way of thinking.

Protecting Human Privacy


Products and technologies must ensure that they do not infringe on
human rights, but rather protect them. AI systems should aim to protect
human citizens, maintaining the individual rights of people to exercise
their thoughts and take action. Scientists in general understandably tend
to get carried away and excited about the amazing accomplishments a
research study uncovers, particularly if it’s a groundbreaking piece of

12
Chapter 1 Responsibility

research. This of course applies to AI researchers as well. In some cases,


this has resulted in lack of awareness of the end users and the impact the
launched product or model might have on society, including unknown/
unintended consequences. Quite unfortunately, it’s usually people from
minority groups, poor backgrounds, and poor socioeconomic statuses that
bear the brunt.
Let’s take a look at the example of the alleged Systeem Risico
Indicatie (SyRI) case from the Netherlands. SyRI was introduced by the
Dutch government as a digital welfare fraud detection system, designed
to mitigate welfare fraud. It used personal data from different sources and
allegedly uncovered fraud. In 2018, six organizations formed a coalition
to sue the Netherlands government over SyRI. SyRI was found to be too
opaque, lacked transparency, and gathered too much data where the
purposes of the data was unclear.25 SyRI also potentially had a bias toward
people living in lower-income neighborhoods, as they were clearly the
main ones seeking welfare. By 2020, a Dutch court decided the SyRI
legislation was unlawful because it did not comply with the right to privacy
under the European Convention of Human Rights.
This is the first of its kind, where policy makers were reprimanded by
a judicial court on the use of an AI system, breaching the right to privacy.
The judgement reminds everyone, especially policy makers, that fraud
detection must happen in a way that respects data protection principles
and the right to privacy. It’s an act of responsibility to recognize and
protect the human rights of people, while preserving their privacy.

Ensuring Safety
When developing algorithms, developers must ensure that AI is deployed
in a safe manner that does not harm or endanger its users. AI safety is one
of the dimensions of responsible AI and it’s important to bear in mind
when developing ML systems. As an example, there are several online

13
Chapter 1 Responsibility

applications and websites that cater to underage users. Online safety is of


paramount importance for such sites, considering the mental, emotional,
and developmental effects harmful content can have on children. Online
safety measures aim to protect children from viewing violent, sexual, and
pornographic content, exposure to false information, or the promotion
of harmful behaviors, including self-harm, anorexia, and suicide.
Oversharing personal information and involvement in bullying or online
harassment are some of the risks considered in child online safety.26
It’s important that model applications and products protect users’
personal identifiable information (PII) by employing privacy methods
such as federated learning and differential privacy. These methods help
anonymize user data and ensure a form of online safety.
Most people don’t create ML algorithms or models with the intent of
harming people, but since AI reflects society, its biases and discriminatory
actions still pose a significant risk. More often than not, AI propagates
harm in the output it produces. It’s therefore an act of responsibility to
develop ML applications with users in mind and employ methods that
promote fairness, safety, privacy, and human rights preservation, and to
ensure that the models are deemed safe, trustworthy, and fair.

Summary
This chapter laid the foundation for responsible AI by looking at the term
“responsibility” and what it means to be responsible and accountable
while protecting human rights, preserving user privacy, and ensuring
human safety. You saw various examples, from the well-known ethical
question of the “trolley problem” to several real-life examples of AI models
that displayed “irresponsible” behavior, and the detrimental effects they’ve
had. The next chapter looks at the next building block of a responsible AI
framework—principles.

14
CHAPTER 2

AI Principles
The first chapter set the foundation for responsible AI frameworks,
kicking off with responsibility and a few examples of AI and its ethical
limitations. This chapter delves into “AI principles,” which are fundamental
components of building responsible AI systems.
Any organization developing and building AI systems should base
these systems on a set of principles, otherwise known as AI principles
or guidelines. AI principles are stepping stones for all types of AI
development carried out across an organization. They are meant to be the
foundation for AI systems and describe how they should be responsibly
developed, trained, tested, and deployed.30
A good number of organizations and governing bodies have a defined
set of AI principles that act as a guiding force for these organizations, and
beyond. AI communities have seen a steady increase in AI principles/
guidelines over the past few years. While the design and outline of AI
principles are fundamental to the development of AI principles, it’s
important that governance and implementation processes are put in place
to execute these principles across an organization.
Most AI principles aim to develop ethical, responsible, safe,
trustworthy, and transparent ML models centered on the following areas:
fairness, robustness, privacy, security, trust, safety, human-­centric AI and
explainability, where explainability is comprised of transparency and
accountability. The first section of this chapter looks at fairness, bias, and
human-centered values and explores how these apply to AI principles.

© Toju Duke 2023 15


T. Duke, Building Responsible AI Algorithms,
https://doi.org/10.1007/978-1-4842-9306-5_2
Chapter 2 AI Principles

Fairness, Bias, and Human-Centered Values


I remember giving a talk on responsible AI a few months ago, and
during the question and answer (Q&A) session, someone mentioned the
following (paraphrased), “I work as a professor in computer science and
I’m tired of hearing the word fairness. It’s been used too many times now,
it’s like a buzzword.” I acknowledged his frustration but my response was
how pleased I was to hear that “fairness” was being overused, because
five to ten years ago, it wasn’t even recognized in machine learning
communities. If it’s suddenly as popular as a Hollywood celebrity, I’m
all for it!
As experts say, overcommunication is better than
undercommunication, and in many cases the saying “What you don’t
know won’t hurt you,” is incorrect, especially in the field of AI. AI
algorithms are full of potential harm, safety, bias, and privacy issues
potentially affecting consumers and users. In this case, what you don’t
know can’t only possibly hurt you, your career, and the reputation of your
organization, it could possibly hurt millions of people across the world,
and potentially future generations as well.
You’re probably wondering what fairness is and what the hoo-ha is
all about. Simply put, fairness, also known as algorithmic fairness in the
context of machine learning, is regarded as providing impartial treatment
to people, especially those of protected statuses.31 What’s fair can mean
different things depending on the context and the people in question.
It can be a confusing concept and fairness as a term has different
definitions across several disciplines. For example, fairness in law
refers to the protection of people and groups of people from any acts of
discrimination and mistreatment, with a focus on prohibiting certain
behaviors and biases. Emphasis is on ensuring that decisions are not based
on protected statuses or social group categories. In the social sciences,
fairness is viewed in the context of social relationships, power dynamics,
institutions, and markets. In quantitative fields such as math, computer

16
Random documents with unrelated
content Scribd suggests to you:
quite impervious, but becoming—at any rate in the case of the larger
and more important pair—open previous to the final ecdysis. We
have mentioned the contradictory opinions of Réaumur and Dufour,
and will now add the views of some modern investigators. Oustalet
says[341] that there are two pairs of spiracles in the nymphs; the first
pair is quite visible to the naked eye, and is situate between pro- and
meso-notum; it is in the nymph closed by a membrane. The other
pair of spiracles is placed above the posterior pair of legs, is small
and completely closed. He does not state what stage of growth was
attained by the nymphs he examined. Palmén was of opinion that
not only thoracic but abdominal spiracles exist in the nymph,[342] and
that they are completely closed so that no air enters them; he says
that the spiracles have tracheae connected with them, that at each
moult the part closing the spiracles is shed with some of the tracheal
exuviae attached to it. The breathing orifices are therefore for a short
time at each ecdysis open, being subsequently again closed by
some exudation or secretion. This view of Palmén's has been
thought improbable by Hagen and Dewitz, who operated by placing
nymphs in alcohol or warm water and observing the escape of
bubbles from the spots where the supposed breathing orifices are
situate. Both these observers found much difference in the results
obtained in the cases of young and of old nymphs. Hagen concludes
that the first pair of thoracic spiracles are functionally active, and that
abdominal stigmata exist though functionless; he appears to be of
opinion that when the first thoracic stigma is closed this is the result
of the abutting against it of a closed trachea. Dewitz found[343] that
in the adult nymph of Aeschna the thoracic stigma is well developed,
while the other stigmata—to what number and in what position is not
stated—are very small. In a half-grown Aeschnid nymph he found
the thoracic stigma to be present in an undeveloped form. On
placing a full-grown nymph in alcohol, gas escaped from the stigma
in question, but in immature nymphs no escape of gas occurred
although they were subjected to a severe test. A specimen that,
when submitted to the above-mentioned immersion, emitted gas,
subsequently moulted, and thereafter air escaped from the spiracle
previously impervious. The observations of Hagen and Dewitz are
perhaps not so adverse to the views of Palmén as has been
supposed, so that it would not be a matter for surprise if Palmén's
views on this point should be shown to be quite correct.

The number of species of Odonata or Libellulidae that have been


described is somewhat less than two thousand, but constant
additions are made to the number, and when the smaller and more
fragile forms from the tropics are collected and worked out it will
probably be found that the number of existing species is somewhere
between five and ten thousand. They are distributed all over the
world, but are most numerous in species in the warmer regions, and
their predominance in any one locality is very much regulated by the
existence of waters suitable for the early stages of their lives.

A good work on the British Odonata is still a desideratum.[344] In


Britain about forty-six species are believed to be native. They are
said to be of late years less numerous than they used to be.
Notwithstanding their great powers of flight, dragon-flies are
destroyed by birds of various kinds; several hawks are said to be
very fond of them, and Merops persicus to line its nest with their
wings. The number of Insects killed by dragon-flies in places where
they are abundant must be enormous; the nymphs, too, are very
destructive in the waters they inhabit, so that dragon-flies have no
doubt been no mean factor in maintaining that important and delicate
balance of life which it is so difficult for us to appreciate. The nymphs
are no doubt cannibals, and this may perhaps be an advantage to
the species, as the eggs are sometimes deposited in large numbers
in a limited body of water, where all must perish if the nymphs did
not, after exhausting other food, attack one another. Martin, speaking
of the Odonata of the Département de l'Indre in France, says:[345]
"The eggs, larvae, and nymphs are the prey of several fishes,
snakes, newts, Coleoptera, aquatic Hemiptera, and of some diving
birds. Sometimes the destruction is on a considerable scale, and one
may notice the dragon-flies of some piece of water to diminish
gradually in numbers, while the animals that prey on them increase,
so that a species may for a time entirely disappear in a particular
spot, owing to the attacks of some enemy that has been specially
prosperous, and also eager in their pursuit. De Selys found that from
a pond filled with carp, roach, perch, and eels, several of the dragon-
fly denizens disappeared directly the bream was introduced." On the
other hand, there can be little doubt that the nymphs are sometimes
injurious to fish; it has been recorded that in a piscicultural
establishment in Hungary 50,000 young fishes were put into a pond
in spring; in the following autumn only fifty-four fish could be found,
but there were present an enormous quantity of dragon-fly nymphs.

Odonata are among the few kinds of Insects that are known to form
swarms and migrate. Swarms of this kind have been frequently
observed in Europe and in North America; they usually consist of
species of the genus Libellula, but species of various other genera
also swarm, and sometimes a swarm may consist of more than one
species. L. quadrimaculata is the species that perhaps most
frequently forms these swarms in Europe; a large migration of this
species is said to occur every year in the Charente inférieure from
north to south.[346] It is needless to say that the instincts and stimuli
connected with these migrations are not understood.

The nymphs are capable, under certain circumstances, of


accommodating themselves to very peculiar conditions of life. The
Sandwich Islands are extremely poor in stagnant waters, and yet
there exist in this remote archipelago several highly peculiar species
of Agrioninae. Mr. R. C. L. Perkins has recently discovered that the
nymphs of some of these are capable of maintaining their existence
and completing their development in the small collections of water
that accumulate in the leaves of some lilies growing on dry land.
These nymphs (Fig. 271) have a shorter mask than occurs, we
believe, in any other Odonata, and one would suppose that they
must frequently wait long for a meal, as they must be dependent on
stray Insects becoming immersed in these tiny reservoirs. The
cannibal habits of the Odonata probably stand these lily-dwellers in
good stead; Mr. Perkins found that there were sometimes two or
three nymphs of different sizes together, and we may suspect that it
sometimes goes hard with the smaller fry. The extension in the
length of the body of one of these lily-frequenting Agrions when it
leaves the water for its aerial existence is truly extraordinary.

Fig. 271.—Under side of Agrionid nymph, with short mask, living in


water in lilies. Hawaiian Islands. × 3.

The Odonata have no close relations with any other group of Insects.
They were associated by Latreille with the Ephemeridae, in a family
called Subulicornia. The members of the two groups have, in fact, a
certain resemblance in some of the features of their lives, especially
in the sudden change, without intermediate condition, from aquatic to
aerial life; but in all important points of structure, and in their
dispositions, dragon-flies and may-flies are totally dissimilar, and
there is no intermediate group to connect them. We have already,
said that the Odonata consist of two very distinct divisions—
Anisopterides and Zygopterides. The former group comprises the
subfamilies Gomphinae, Cordulegasterinae, Aeschninae,
Corduliinae, and Libellulinae,—Insects having the hinder wings
slightly larger than the anterior pair; while the Zygopterides consist of
only two subfamilies—Calepteryginae and Agrioninae; they have the
wings of the two pairs equal in size, or the hinder a little the smaller.
The two groups Gomphinae and Calepteryginae are each, in several
respects, of lower development than the others, and authorities are
divided in opinion as to which of the two should be considered the
more primitive. It is therefore of much interest to find that there exists
an Insect that shares the characters of the two primitive subfamilies
in a striking manner. This Insect, Palaeophlebia superstes (Fig. 272),
has recently been discovered in Japan, and is perhaps the most
interesting dragon-fly yet obtained. De Selys Longchamps refers it to
the subfamily Calepteryginae, on account of the nature of its wings;
were the Insect, however, deprived of these organs, no one would
think of referring Palaeophlebia to the group in question, for it has
the form, colour, and appearance of a Gomphine Odonate.
Moreover, the two sexes differ in an important character,—the form
of the head and eyes. In this respect the female resembles a
Gomphine of inferior development; while the male, by the shape and
large size of the ocular organs, may be considered to combine the
characters of Gomphinae and Calepteryginae. The Insect is very
remarkable in colour, the large eyes being red in the dead examples.
We do not, however, know what may be their colour during life, as
only one pair of the species is known, and there is no record as to
the life-history and habits. De Selys considers the nearest ally of this
Insect to be Heterophlebia dislocata, a fossil dragon-fly found in the
Lower Lias of England.

Fig. 272.—Palaeophlebia superstes. A, The Insect with wings of one


side and with two legs removed; B, front view of head of female;
C, of male. (After De Selys.)

Numerous fossil dragon-flies are known; the group is well


represented in the Tertiary strata, and specimens have been found in
amber. In strata of the Secondary age these Insects have been
found as far back as the Lower Lias; their remains are said to exist in
considerable variety in the strata of that epoch, and some of them to
testify to the existence at that period of dragon-flies as highly
specialised as those now living. According to Hagen[347]
Platephemera antiqua and Gerephemera simplex, two Devonian
fossils, may be considered as dragon-flies; the evidence as to this
appears inadequate, and Brongniart refers the latter Insect to the
family Platypterides, and considers Platephemera to be more allied
to the may-flies.
One of the most remarkable of the numerous discoveries lately
made in fossil entomology is the finding of remains of huge Insects,
evidently allied to dragon-flies, in the Carboniferous strata at
Commentry. Brongniart calls these Insects Protodonates,[348] and
looks on them as the precursors of our Odonata. Meganeura monyi
was the largest of these Insects, and measured over two feet across
the expanded wings. If M. Brongniart be correct in his restoration of
this giant of the Insect world, it much resembled our existing dragon-
flies, but had a simple structure of the thoracic segments, and a
simpler system of wing-nervures. On p. 276 we figured
Titanophasma fayoli, considered by Scudder and Brongniart as allied
to the family Phasmidae, and we pointed out that this supposed
alliance must at best have been very remote. This view is now taken
by M. Brongniart himself,[349] he having removed the Insect from the
Protophasmides to locate it in the Protodonates near Meganeura.
There appears to be some doubt whether the wings supposed to
belong to this specimen were really such, or belonged rather to
some other species.

CHAPTER XIX

AMPHIBIOUS NEUROPTERA CONTINUED—EPHEMERIDAE, MAY-FLIES

Fam. VII. Ephemeridae—May-flies.

Delicate Insects with atrophied mouth and small, short antennae;


with four membranous wings having much minute cross-veining;
the hinder pair very much smaller than the other pair, sometimes
entirely absent: the body terminated by three or two very
elongate slender tails. The earlier stages are passed through in
water, and the individual then differs greatly in appearance from
the winged Insect; the passage between the two forms is
sudden; the creature in its first winged state is a subimago,
which by shedding a delicate skin reveals the final form of the
individual.

Fig. 273.—Ephemera danica, male, Britain.

The may-flies are well known—in literature—as the types of a brief


and ineffective life. This supposed brevity relates solely to their
existence in the winged form. In the earlier stages the may-fly is so
unlike its subsequent self that it is not recognised as a may-fly by the
uninitiated. The total life of the individual is really quite as long as
that of most other Insects. The earlier stages and life-histories of
these Insects are of great importance. The perfect Insects are so
delicate and fragile that they shrivel much in drying, and are very
difficult to preserve in a condition suitable for study.

The mouth of the imago is atrophied, the trophi scarcely existing as


separate parts. Packard says that in Palingenia bilineata he could
discover no certain traces of any of the mouth-parts, but in
Leptophlebia cupida he found, as he thought, the rudiments of the
maxillae and labium, though not of the mandibles. The antennae are
always short, and consist of one or two thick basal joints succeeded
by a delicate needle-like segment, which, though comparatively long,
is not divided. The ocular organs are remarkable for their large size
and complex development; they are always larger in the male than
they are in the female. The compound eyes of the former sex are in
certain species, e.g. Cloëon (Fig. 274), quite divided, so that each
eye becomes a pair of organs of a different character; one part forms
a pillar facetted at its summit, while the other part remains as a true
eye placed on the side of the head; in front of these compound eyes
there are three ocelli. Thus the Insect comes to have three different
kinds of eyes, together seven in number.

Fig. 274.—Front of head of Cloëon, male. a, Pillared eye; b, sessile


eye; c, ocellus.

The prothorax is small, the pronotum being, however, quite distinct.


The mesothorax is very large; its notum forms by far the larger part
of the upper surface of the thoracic region, the metathorax being
small and different in structure, resembling in appearance a part of
the abdomen, so that the hind wings look as if they were attached to
a first abdominal segment. The mesosternum is also
disproportionately large in comparison with the homologous piece
preceding it, and with that following it. The pleural pieces are large,
but their structure and disposition are only very imperfectly
understood. The coxae are small and are widely separated, the
anterior being, however, more elongate and approximate than the
others. The other parts of the legs are slender; the number of joints
in the tarsi varies from five to one. The legs throughout the family
exhibit a considerable variety of structure, and the front pair in the
males of some species are remarkably long. The abdomen is usually
slender, and consists of ten segments; the terminal one bears three,
or two, very long flexible appendages. The first dorsal plate of the
abdomen is either wanting or is concealed to a considerable extent
by the metanotum. The wings are peculiar; the anterior pair vary a
great deal in their width, but are never very long in proportion to the
width; the hind pair are always disproportionately small, and
sometimes are quite wanting. The venation consists of a few, or of a
moderate number, of delicate longitudinal veins that do not pursue a
tortuous course, but frequently are gracefully curved, and form a
system of approximately similar curves, most of the veins being of
considerable length; close to the anterior margin of the wing there
are two or three sub-parallel veins. Frequently there are very
numerous fine, short cross-veinlets, but these vary greatly and may
be entirely wanting.

Fig. 275.—Wings of Ephemera danica. (After Eaton.)

The earlier stages of the life of Ephemeridae are, it is believed, in the


case of all the species, aquatic. May-flies, indeed, during the period
of their post-embryonic development are more modified for an
aquatic life than any other Insects, and are provided with a complex
apparatus of tracheal gills. The eggs are committed to the waters
without any care or foresight on the part of the parent flies, thus the
embryonic development is also aquatic; little, however, is known of it.
According to Joly[350] the process in Palingenia virgo is slow. The
larva on emerging from the egg has no respiratory system, neither
could Joly detect any circulation or any nervous system. The
creature on emergence is very like Campodea in form, possessing
long antennae and tails—caudal setae. Owing to the organisation
being inferior, the creature in its earlier stages is called a larvule; in
its later stages it is usually spoken of as a nymph, but the term larva
is also frequently applied to it. Soon the gills begin to appear in the
form of small tubular caeca placed in the posterior and upper angles
of the abdominal rings; in fifteen days the gills begin to assume their
characteristic form, are penetrated by tracheae, and the circulation
can be seen. The amount of growth accomplished after hatching
between March and September is but small.
Fig. 276.—Nymph of Cloëon dipterum.[351] Wing-sheath of left side,
gills of right side, removed; g, tracheal gills. (After Vayssière.)

Fig. 277.—Larvule of Cloëon dimidiatum. (After Lubbock.)

The metamorphosis of Cloëon has been described by Sir John


Lubbock; he informs us that the young creature undergoes a
constant and progressive development, going through a series of
more than twenty moults, each accompanied by a slight change of
form or structure. His observations were made on captured
specimens, so that it is not certain that what he calls[352] the first
stage is really such. He found no tracheae in the earliest stages; the
small first rudiments of the gills became visible in the third stage,
when there were no tracheae; the fourth instar possessed tracheae,
and they could be seen in the gills. The wing rudiments could first be
detected in the ninth and tenth stages. The changes of skin during
the winter months are separated by longer intervals than those
occurring at other periods of the year.
Fig. 278.—Adult nymph of Ephemera vulgata. (After Eaton.) Britain.

The nymphs differ greatly in the structure and arrangement of their


tracheal gills, and display much variety in their general form and
habits; some of them are very curious creatures. Pictet[353] divides
them in accordance with their habits into four groups: (1) Fossorial
larvae: these live in the banks of streams and excavate burrows for
shelter; they are of cylindrical form, possess robust legs, abundant
gills at the sides of the body, and frequently processes projecting
forwards from the head: examples, Ephemera (Fig. 278) and
Palingenia. (2) Flat larvae: these live attached to rocks, but run with
rapidity when disturbed; they prefer rapid streams, have the
breathing organs attached to the sides of the body and not reposing
on the back; they are exclusively carnivorous, while the fossorial
forms are believed to obtain their nutriment by eating mud: example,
Baëtis. (3) Swimming larvae: elongate delicate creatures, with feeble
legs, and with strongly ciliated caudal setae: example, Cloëon (Fig.
276). (4) Climbing larvae: these live in slowly-moving waters,
especially such as have much slimy mud in suspension, and they
have a habit of covering themselves with this mud sometimes to
such an extent as to become concealed by it: example,
Potamanthus.
Fig. 279.—Nymph of Oligoneuria garumnica, France. g2 and g7, two of
the dorsal tracheal gills. (After Vayssière.)

The anatomy of the nymphs has been treated by Vayssière,[354] who


arranges them in five groups in accordance with the conditions of the
tracheal gills: (1) The gills are of large size, are exposed and
furnished at the sides with respiratory fringes: example, Ephemera
(Fig. 278). (2) The branchiae are blade-like, not fringed, and are
exposed at the sides of the body: example, Cloëon (Fig. 276). (3)
The respiratory tubes are placed on the under surface of plates
whose upper surface is not respiratory: example, Oligoneuria
garumnica (Fig. 279). (4) The anterior gill is modified to form a plate
that covers the others: example, Tricorythus (Fig. 282, B). (5) The
gills are concealed in a respiratory chamber: example, Prosopistoma
(Fig. 280). The last of these nymphs is more completely adapted for
an aquatic life than any other Insect at present known; it was for long
supposed to be a Crustacean, but it has now been shown to be the
early stage of a may-fly, the sub-imago having been reared from the
nymph. The carapace by which the larger part of the body is covered
is formed by the union of the pro- and meso-thorax with the sheaths
of the anterior wings, which have an unusually extensive
development; under the carapace there is a respiratory chamber, the
floor and sides of which are formed by the posterior wing-sheaths,
and by a large plate composed of the united nota of the metathorax
and the first six abdominal segments. In this chamber there are
placed five pairs of tracheal gills; entrance of water to the chamber is
effected by two laterally-placed orifices, and exit by a single dorsal
aperture. These nymphs use the body as a sucker, and so adhere
strongly to stones under water. When detached they swim rapidly by
means of their caudal setae; the form of these latter organs is
different from that of other Ephemerid nymphs. This point and other
details of the anatomy of this creature have been described in detail
by Vayssière.[355] These nymphs have a very highly developed
tracheal system; they live in rapid watercourses attached to stones
at a depth of three to six inches or more under the water. Species of
Prosopistoma occur in Europe, Madagascar, and West Africa.

Fig. 280.—Prosopistoma punctifrons, nymph. France. (After Vayssière.)


o, Orifice of exit from respiratory chamber.

According to Eaton,[356] in the nymphs of some Ephemeridae the


rectum serves, to a certain extent, as a respiratory agent; he
considers that water is admitted to it and expelled after the manner
we have described in Odonata, p. 421.

Fig. 281.—A, Last three abdominal segments and bases of the three
caudal processes of Cloëon dipterum: r, dorsal vessel; kl, ostia
thereof; k, special terminal chamber of the dorsal vessel with its
entrance a; b, blood-vessel of the left caudal process; B, twenty-
sixth joint of the left caudal process from below; b, a portion of the
blood-vessel; o, orifice in the latter. (After Zimmermann.)

The internal anatomy of the nymphs of Ephemeridae shows some


points of extreme interest. The long caudal setae are respiratory
organs of a kind that is almost if not quite without parallel in the other
divisions of Insecta. The dorsal vessel for the circulation of the blood
is elongate, and its chambers are arranged one to each segment of
the body. It drives the blood forwards in the usual manner, but the
posterior chamber possesses three blood-vessels, one of which is
prolonged into each caudal seta. This terminal chamber is so
arranged as to drive the blood backwards into the vessels of the
setae; on the under surface of the vessels there are oval orifices by
which the blood escapes into the cavity of the seta so as to be
submitted to the action of the surrounding medium for some of the
purposes of respiration. This structure has been described by
Zimmermann,[357] who agrees with Creutzberg[358] that the organ by
which the blood is propelled into the setae is a terminal chamber of
the dorsal vessel; Verlooren,[359] who first observed this accessory
system of circulation, thought the contractile chamber was quite
separate from the heart. The nature of the connexion between this
terminal chamber that drives the blood backwards and the other
chambers that propel the fluid forwards appears still to want
elucidation.

Fig. 282.—A, Nymph of Ephemerella ignita with gills of left side


removed; g, gills: B, nymph of Tricorythus sp. with gill cover of
right side removed; g.c, gill cover; g, g′, gills. (After Vayssière.)

The nymphs of the Ephemeridae being creatures adapted for


existence in water, the details of their transformation into creatures
having an entirely aerial existence cannot but be of much interest. In
the nymphs the tracheal system is well developed, but differs from
that of air-breathing Insects in the total absence of any spiracles.
Palmén has investigated this subject,[360] and finds that the main
longitudinal tracheal trunks of the body of the nymph are not
connected with the skin of the body by tracheae, but are attached
thereto by ten pairs of slender strings extending between the
chitinous integument and the tracheal trunks. When the skin is shed
these strings—or rather a chitinous axis in each one—are drawn out
of the body, and bring with them the chitinous linings of the tracheae.
Thus notwithstanding the absence of spiracles, the body wall is at
each moult pierced by openings that extend to the tracheae. After
the ordinary moults these orifices close immediately, but at the
change to the winged state they remain open and form the spiracles.
At the same time the tracheal gills are completely shed, and the
creature is thus transformed from a water-breather to an Insect
breathing air as usual. In addition to this change there are others of
great importance, such as the development of the great eyes and the
complete atrophy of the mouth-parts. The precise manner of these
changes is not known; they occur, however, within the nymph skin.
The sudden emergence of the winged Insect from the nymph is one
of the most remarkable facts in the life-history of the may-fly; it has
been observed by Sir John Lubbock,[361] who describes it as almost
instantaneous. The nymph floats on the water, the skin of the back
opens, and the winged Insect flies out, upwards and away; "from the
moment when the skin first cracks not ten seconds are over before
the Insect has flown away." The creature that thus escapes has not,
however, quite completed its transformation. It is still enveloped in a
skin that compresses and embarrasses it; this it therefore rapidly
gets rid of, and thus becomes the imago, or final instar of the life-
cycle. The instar in which the creature exists winged and active,
though covered with a skin, is called the sub-imago. The parts of the
body in the sub-imago are as a whole smaller than they are in the
imago, and the colour is more dingy; the appendages—wings, legs,
and caudal setae—are generally considerably shorter than they are
in the imago, but attain their full length during the process of
extraction. The creatures being, according to Riley, very impatient
and eager to take to the wing, the completion of the shedding of the
skin of the sub-imago is sometimes performed while the Insect is
flying in the air.

Fig. 283.—Lingua of Heptagenia longicauda, × 16. m, Central; l, lateral


pieces. (After Vayssière.)

The food of young Ephemeridae is apparently of a varied and mixed


nature. Eaton says[362] that though sometimes the stronger larvae
devour the weaker, yet the diet is even in these cases partly
vegetable. The alimentary canal frequently contains much mud; very
small organisms, such as diatoms and confervae, are thought to
form a large part of the bill of fare of Ephemerid nymphs. Although
the mouth is atrophied in the imago, yet it is highly developed in the
nymphs. This is especially notable in the case of the lingua or
hypopharynx (Fig. 283); indeed Vayssière[363] seems to incline to the
opinion that this part of the mouth may be looked on in these Insects
as a pair of appendages of a head-segment (see p. 96 ante), like the
labium or maxillae.

The life-history has not been fully ascertained in the case of any
species of may-fly; it is known, however, that the development of the
nymph sometimes occupies a considerable period, and it is thought
that in the case of some species this extends to as much as three
years. It is rare to find the post-embryonic development of an Insect
occupying so long a period, so that we are justified in saying that
brief as may be the life of the may-fly itself, the period of preparation
for it is longer than usual. Réaumur says, speaking of the winged fly,
that its life is so short that some species never see the sun. Their
emergence from the nymph-skin taking place at sunset, the duties of
the generation have been, so far as these individuals are concerned,
completed before the morning, and they die before sunrise. He
thinks, indeed, that individuals living thus long are to be looked on as
Methuselahs among their fellows, most of whom, he says, live only
an hour or half an hour.[364] It is by no means clear to which species
these remarks of Réaumur refer; they are doubtless correct in
certain cases, but in others the life of the adult is not so very short,
and in some species may, in all probability, extend over three or four
days; indeed, if the weather undergo an unfavourable change so as
to keep them motionless, the life of the flies may be prolonged for a
fortnight.

The life of the imago of the may-fly is as remarkable as it is brief; in


order to comprehend it we must refer to certain peculiarities of the
anatomy with which the vital phenomena are connected. The more
important of these are the large eyes of the males, the structure of
the alimentary canal, and that of the reproductive organs. We have
already remarked that the parts of the mouth in the imago are
atrophied, yet the canal itself not only exists but is even of greater
capacity than usual; it appears to have much the same general
arrangement of parts as it had in the nymph. Its coats are, however,
of great tenuity, and according to Palmén[365] the divisions of the
canal are separated by changes in the direction of certain portions
anterior to, and of others posterior to, its central and greater part—
the stomach—in such a manner that the portions with diverted
positions act as valves. The stomach, in fact, forms in the interior of
the body a delicate capacious sac; when movement tends to
increase the capacity of the body cavity then air enters into the
stomachic sac by the mouth orifice, but when muscular contractions
result in pressure on the sac they close the orifices of its extremities
by the valve-like structures we have mentioned above; the result is,
that as complex movements of the body are made the stomach
becomes more and more distended by air. It was known even to the
old naturalists that the dancing may-fly is a sort of balloon, but they
were not acquainted with the exact mode of inflation. Palmén says
that in addition to the valve-like arrangements we have described,
the entry to the canal is controlled by a circular muscle, with which
are connected radiating muscles attached to the walls of the head.
Palmén's views are adopted, and to a certain extent confirmed, by
Fritze,[366] who has examined the alimentary canal of the may-fly,
and considers that though the normal parts of the canal exist, the
function is changed in the imago, in which the canal serves as a sort
of balloon, and aids the function of the reproductive organs. The
change in the canal takes place in an anticipatory manner during the
nymph and sub-imago stages.

The sexual organs of Ephemeridae are remarkable for their


simplicity; they are destitute of the accessory glands and diverticula
that, in some form or other, are present in most other Insects. Still
more remarkable is the fact that the ducts by which they
communicate with the exterior continue as a pair to the extremity of
the body, and do not, as in other Insects, unite into a common duct.
Thus in the female there is neither bursa copulatrix, receptaculum
seminis, nor uterine portion of oviduct, and there is no trace of an
ovipositor; the terminations of the ducts are placed at the hind
margin of the seventh ventral plate, just in front of which they are
connected by a fold of the integument. The ovary consists of a very
large number of small egg-tubes seated on one side of a sac, which
forms their calyx, and one of whose extremities is continued
backwards as one of the pair of oviducts. The male has neither
vesiculae seminales, accessory glands, nor ductus ejaculatorius.
The testes are elongate sacs, whose extremities are prolonged
backwards forming the vasa deferentia; these open separately at the
extremity of the body, each on a separate intromittent projection of
more or less complex character, the two organs being, however,
connected by means of the ninth ventral plate, of which they are,
according to Palmén, appendages. We should remark that this
authority considers Heptagenia to form, to some extent, an exception
as regards the structures of the female; while Polymitarcys is in the
male sex strongly aberrant, as the two vasa deferentia, instead of
being approximately straight, are bent inwards at right angles near
their extremities so as to meet, and form in the middle a common
cavity, which then again becomes double to pass into the pair of
intromittent organs.
According to the views of Exner and others, the compound eyes of
Insects are chiefly organs for the perception of movement; if this
view be correct, movements such as those made during the dances
of may-flies may, by the number of the separate eyes, by their
curved surfaces and innumerable facets, be multiplied and
correlated in a manner of which our own sense of sight allows us to
form no conception. We can see on a summer's evening how
beautifully and gracefully a crowd of may-flies dance, and we may
well believe that to the marvellous ocular organs of the flies
themselves (Fig. 274) these movements form a veritable ballet. We
have pointed out that by this dancing the peculiarly formed
alimentary canal becomes distended, and may now add that Palmén
and Fritze believe that the unique structure of the reproductive
organs is also correlated with the other anatomical peculiarities, the
contents of the sexual glands being driven along the simple and
direct ducts by the expansion of the balloon-like stomach. During
these dances the momentary conjugation of the sexes occurs, and
immediately thereafter the female, according to Eaton, resorts to the
waters appropriate for the deposition of her eggs. As regards this,
Eaton says:[367] "Some short-lived species discharge the contents of
their ovaries completely en masse, and the pair of fusiform or
subcylindrical egg-clusters laid upon the water rapidly disintegrate,
so as to let the eggs sink broadcast upon the river-bed. The less
perishable species extrude their eggs gradually, part at a time, and
deposit them in one or other of the following manners: either the
mother alights upon the water at intervals to wash off the eggs that
have issued from the mouths of the oviducts during her flight, or else
she creeps down into the water to lay her eggs upon the under-side
of stones, disposing them in rounded patches, in a single layer
evenly spread, and in mutual contiguity." The eggs are very
numerous, and it is thought may sometimes remain in the water as
much as six or seven months before they hatch.

The number of individuals produced by some kinds of may-flies is


remarkable. Swarms consisting of millions of individuals are
occasionally witnessed. D'Albertis observed Palingenia papuana in
countless myriads on the Fly River in New Guinea: "For miles the
surface of the river, from side to side, was white with them as they
hung over it on gauzy wings; at certain moments, obeying some
mysterious signal, they would rise in the air, and then sink down
anew like a fall of snow." He further states that the two sexes were in
very disproportionate numbers, and estimates that there was but a
single female to every five or six thousand males.

Ephemeridae in the perfect state are a favourite food of fishes, and it


is said that on some waters it is useless for the fly-fisher to try any
other lure when these flies are swarming. Most of the "duns" and
"spinners" of the angler are Ephemeridae; so are several of the
"drakes," our large E. danica and E. vulgata being known as the
green drake and the gray drake. Ronalds says[368] that the term
"dun" refers to the pseud-imago condition, "spinner" to the perfect
Insect. E. danica and E. vulgata are perhaps not distinguished by
fishers; Eaton says that the former is abundant in rapid, cool
streams, while E. vulgata prefers warmer and more tranquil rivers.

These sensitive creatures are unable to resist the attractions of


artificial lights. Réaumur noticed this fact many years ago, and since
the introduction of the electric light, notes may frequently be seen in
journals recording that myriads of these Insects have been lured by it
to destruction. Their dances may frequently be observed to take
place in peculiar states of light and shade, in twilight, or where the
sinking sun has its light rendered broken by bushes or trees;
possibly the broken lights are enhanced in effect by the ocular
structures of the Insects. It has recently been ascertained that a
species of Teleganodes is itself luminous. Mr. Lewis,[369] who
observed this Insect in Ceylon, states that in life the whole of the
abdomen was luminous, not brightly so, but sufficient to serve as a
guide for capturing the Insect on a dark night. It has also been
recorded that the male of Caenis dimidiata gives a faint blue light at
night.

Nearly 300 species of Ephemeridae are known, but this may be only
a fragment of what actually exist, very little being known of may-flies

You might also like