This group reviewed data and measurements indicating the positive potential of AI to serve Sustainable Development Goals (SDG’s). Alongside these optimistic inquiries, this group also investigated the risks of AI in areas such as privacy, vulnerable populations, human rights, workplace and organizational policy. The socio-political consequences of AI raise many complex questions which require continued rigorous examination.
Handwritten Text Recognition for manuscripts and early printed texts
Breakout 3. AI for Sustainable Development and Human Rights: Inclusion, Diversity, Human Dignity.
1. The Problem with Metrics
Rachel Thomas, PhD
USF Center for Applied Data Ethics & fast.ai
@math_rachel
2. Goodhart’s Law: “When a measure becomes a target, it
ceases to be a good measure.”
• Any metric is just a proxy for what you really care about
• Metrics can, and will, be gamed
• Metrics tend to overemphasize short-term concerns
• Many online metrics are gathered in highly addictive environments
• Metrics are most likely to be useful when they are treated as one
piece of a bigger picture
3. Any metric is just a proxy for what you really
care about
Papers published/citation count vs. genuine impact
# of students taught vs. what they understood/learned
Diversity stats vs. how people were treated/what opportunities
they were given/if their feedback was listened to
Benchmark tasks vs. suitability of those benchmarks
4. “How many people, species, countries, or square meters would be impacted by a
solution to the problem? What level of performance would constitute a meaningful
improvement over the status quo?”
Machine Learning That Matters, by Kiri Wagstaff
5. Metrics tend to overemphasize short-term
concerns
It is much easier to measure short-term quantities: click through rates,
month-over-month churn, quarterly earnings.
Many long-term trends have a complex mix of factors and are tougher to
quantify.
6. what metrics don’t capture
“Since we can not know in advance every phenomenon users will
experience, we can not know in advance what metrics will quantify
these phenomena.”
7. Case study: Diversity & Inclusion in Tech
Needed stats:
• comparative promotion rates
• cap table ownership
• retention rates
• number of harassment victims silenced by NDAs
• rates of under-leveling
Even then, all this data should still be combined with listening to first-
person experiences of those working at these companies.
8. How to measure negative impacts of AI on human rights?
How to measure how much people from underrepresented
groups are listened to and included?
10. • Any metric is just a proxy for what you really care about
• Metrics can, and will, be gamed
• Metrics tend to overemphasize short-term concerns
• Many online metrics are gathered in highly addictive
environments
• Metrics are most likely to be useful when they are treated
as one piece of a bigger picture
11. CONFIDENTIAL AND PROPRIETARY
Any use of this material without specific permission of McKinsey & Company
is strictly prohibited
October 30, 2019
Extracts from MGI research
Applying AI for
social good
12. McKinsey & Company 2
We built a library of use cases of AI for social good
interviews and expert discussions inside
and outside of McKinsey with
▪ Social impact leaders
▪ Academics
▪ Technologists
~100
use cases mapped in depth
10 social impact domains mapped
18
~ 160
AI capabilities assessed
Social problems
AI capabilities
AI capabilities
Social problems
~160 AI for social good
use cases
Source: McKinsey Global Institute analysis
13. McKinsey & Company
Mapping AI to social good use cases
4
15
21
28
16
16
17
13
11
15
Economic empowerment
▪ Agricultural quality and yield
▪ Financial inclusion
▪ Initiatives for economic growth
▪ Labor supply and demand matching
Education
▪ Access and completion of education
▪ Maximizing student achievement
▪ Teacher and administration productivity
Environment
▪ Animal and plant conservation
▪ Climate change and adaptation
▪ Energy efficiency and sustainability
▪ Land, air, and water conservation
Equality and inclusion
▪ Accessibility and disabilities
▪ Exploitation
▪ Marginalized communities
Information verification and validation
▪ False news
▪ Polarization
Infrastructure
▪ Energy
▪ Real estate
▪ Transportation
▪ Urban planning
▪ Water and waste management
Public and social sector
▪ Effective management of public sector
▪ Effective management of social sector
▪ Fundraising
▪ Public finance management
▪ Services to citizens
Security and justice
▪ Harm prevention
▪ Fair prosecution
▪ Policing
Crisis response
▪ Disease outbreak
▪ Migration crises
▪ Natural and man-made disasters
▪ Search and rescue
Health and hunger
▪ Treatment delivery
▪ Prediction and prevention
▪ Treatment and long-term care
▪ Mental wellness
▪ Hunger
Number of use cases per domain
3Source: McKinsey Global Institute analysis
14. McKinsey & Company
UN goals AI use case breakdown
Life below water
Affordable and clean energy
Clean water and sanitation
Responsible consumption and production
Sustainable cities and communities
Gender equality
Partnerships for the goals
Zero hunger
Decent work and economic growth
Climate action
Reduced inequalities
Industry, innovation, and infrastructure
No poverty
Life on land
Quality education
Peace, justice, and strong institutions
Good health and well-being
AI use cases support the UN Sustainable Development Goals
AI
use case
key
Security and justice
Public and social
sector
Infrastructure
Info verification
and validation
Health and hunger
Crisis response
Economic
empowerment
Education
Environment
Equality and inclusion
UN Sustainable Development Goals
Number of use cases
4Source: McKinsey Global Institute analysis
15. McKinsey & Company 5
▪ While assistive techniques exist today, it
is often difficult for visually impaired
individuals to access education and
resources, especially in developing
countries
▪ Microsoft’s Seeing AI mobile application
enables increased self-sufficiency by
interpreting and narrating the
environment using numerous computer
vision capabilities to help with obstacle
recognition, reading printed text, and
labeling utilities
▪ It is accessible to users in 70 countries
around the world
AI For Social Good Example
Helping The Blind
16. McKinsey & Company 6
▪ Mobile application leveraging AI to identify
whether photographed moles are indicative
of skin cancer could be a helpful solution for
those in rural communities where there is
poorer access to dermatologists
▪ Researchers at Stanford University and the
University of Heidelberg have been able to
generate an AI model that is more accurate
than dermatologists (95% accuracy vs 86%)
▪ When models as such can be packaged into
easy to access smartphone apps, it can be a
game-changer for rural communities
AI For Social Good Example
Detecting Skin Cancer
17. McKinsey & Company 7
▪ Frequent and broad area satellite imagery
enables new AI-based systems to quickly and
accurately detect infrastructure changes that
may affect evacuation and response
▪ For example, following the passage of
Hurricane Harvey in 2017, a collaboration
between Planet Labs, provider of satellite
imagery, and CrowdAI provided an immediate
view of the greater Houston area and was
able to detect road outages due to flooding
and quantify infrastructure damage
Flooded roads
Non-flooded roads
AI For Social Good Example
Road Access During Floods
18. McKinsey & Company 8
Several bottlenecks could limit AI’s benefit to society
Data accessibility Data quality
Talent with AI-expertise “Last mile” implementation challenges
Source: McKinsey Global Institute analysis
19. McKinsey & Company 9
Four major types of risks to be mitigated
when applying AI for social good
Bias and fairness Privacy and
security
Explainability Use and misuse
considerations
Source: McKinsey Global Institute analysis
20. McKinsey & Company 10
What to measure?
▪ Which AI applications are currently being deployed? In which domains?
Where, and by which organizations?
▪ Where is there high potential but limited deployment?
Deployed applications
of AI for social good
▪ How available / accessible is data across the UN SDGs and use cases?
▪ Where is AI talent being hired? Where are experts available to work on
social impact problems?
Progress addressing
bottlenecks
▪ How might we measure fairness and detect bias?
▪ What progress has been made in development of inclusive datasets? Are
we able to measure the value of privacy?
▪ Who is committing to responsible use (e.g., signing on to AI principles)?
▪ What progress has been made in measuring explainability?
Risks and actions to
mitigate
▪ How can we use AI to measure progress towards each of the UN SDGs
and indicators?Progress towards the
UN SDGs
A
B
C
D
Measuring deforestation
through satellite imagery
Source: Planet Labs
Measuring economic activity
through flows of postal items
Source: UN Global Pulse
25. Methodology: Human & Machine
Quid Machine
Global News & Blogs
Organizations
Global Patents
Custom
Open Source Data
Client Data
HUMAN
Human Researcher Human-In-The-Loop
Curation &
Peer Review1. Determines relevant terms
for the question (e.g.
“Artificial Intelligence,”
“growth,” “funding”)
2. Arranges terms into a
Boolean search, thereby
translating human intent
into machine language
3. Determines search
parameters (timeframe,
geography, publication
type, etc.)
4. Iterates on the inputs to
match intent
1. Applies the Boolean
search to a corpus of
written documents and de-
duplicates results
2. Groups together
documents with semantic
similarity (Louvain
clustering)
3. Extracts representative
documents for each cluster
and chooses cluster names
(noun chunk methodology)
1. Researcher works in a
responsive UI to gain insights
from the data
2. Human corrects AI’s
mistakes
3. Researcher shares Quid
analysis with peers for
review and iteration on
methodology
4. “Human in the loop”
paradigm combines human
intuition and expertise with
machine scale and AI power
26. The Technology Behind the Platform
Integrated
Custom
Open Source Data
Client Data
Output: Network Map
& Derived AI Insights
Integrated:
• Global News & Blogs
(1.8 million sources*,
provided by LexisNexis)
• Organizations (2.2
million*, provided by
Capital IQ and
Crunchbase)
• Global Patents (global
coverage of past 60
years, provided by
Thomson Innovation)
Custom:
• Open Source Data
• Client Data
• Topic Detection/Modeling
• Document Clustering
• Entity Extraction
• Story Lines
• Anomaly Detection
• Sentiment Analysis
• Cross-Dataset Analysis
• Event Detection
• Data Summarization
• Machine Learning
PlatformData
2010 -2019
29. Network Map: News Coverage, 2018 -2019
News article network with 3451 stories. Colored by clusters. Sized by degree. Labeled by clusters.
Data Privacy (9.8%)
AI & Data Science
Thought Leaders (9.6%)
Ethical AI (7.6%)
Artificial Intelligence (6.7%)
Machine Learning Explainability (5.2%)
EU Commission AI Ethics Guidelines (4.7%)
Hiring Bias (4.6%)
Google Dissolves AI Ethics Council (4.5%)
Facial Recognition Bias (3.9%)
Machine Learning Bias (3.6%)
Gender Bias, Gender Diversity (3.2%)
AI in Government (3.1%)
AI Ethics Framework (2.8%)
Machine Learning Fairness (2.5%)
Health Data Privacy (2.4%)
Google Walkout (2.4%)
Unesco Defends Artificial Intelligence (2.3%)
IBM Introduces AI Explainability (2.3%)
Human Rights & AI (2.3%)
EU AI Guidelines (1.9%)
UK Government & AI (1.5%)
Oxford University AI Study (1.4%)
Google Announces AI Ethics Council (1.2%)
AI 's Diversity Crisis (0.99%)
Facebook Backs TUM's AI Ethics Institute (0.96%)
AI Fairness (0.96%)
Public Trust (0.67%)
Apple Acquirers Silk Labs (0.67%)
30. Companies That Are Driving The Conversation Around Ethical AI In The News
News article bar chart with 910 stories. Colored by segment.
0 100 200 300 400 500
Number of Stories
Google
Amazon
Facebook
Microsoft
IBM
Blackstone Group
Twitter
26% of Google’s news coverage is on
their short-lived “AI Ethics Board.”
Bias is a significant portion of
Amazon’s news coverage. This is
predominately attributed to Amazon’s
recruiting algorithm that was deemed
discriminatory towards female
applicants.
Google is the dominate company in
the news when it comes to policy
and AI
Color Key: Policy 31%, Governance 29%, Ethics 20%, Bias 18%
31. Granular Conversations: Stories With High Social Engagement*
News article scatter plot with 6 stories. Colored by clusters. Labeled by story title.
A global ethics study aims to help AI solve the self-driving “trolley problem”
Hundreds of Google employees have signed a petition calling for the removal of a member of Google's new AI
ethics board over her comments on immigrants and trans people
US billionaire gives Oxford University £150m
Female-voice AI reinforces bias, says UN report
Artificial intelligence has been weaponized in China. That should be a wake-up call for the world
Saying ‘no’ to gender bias Google Translate will now show gender-specific translations for some languages
Published Count
SocialEngagement
Gender Bias/Gender Diversity, Ethical AI, Google Dissolves AI Ethics Council, Human Rights & AI, Oxford Uni AI Study
1 2 4 7 13
4112
8958
19515
42512
92612
MIT Media Lab’s viral experiment, “Moral Machine,” one of the
largest studies done on global moral preferences.
This study is indicative of people’s desire to be included in the
conversation around the morality of AI, especially through the lens
of something tangible like autonomous driving.
Gender bias is highly
engaged within the
context of AI.
The removal of Kay Cole Jones
from Google’s AI Ethics Board was
a key storyline in the news
33. S p e a k e r :
• D a r i a M e h r a ( d m e h r a @ q u i d . c o m )
M a t e r i a l s C r e a t e d b y :
• J u l i e K i m ( j k i m @ q u i d . c o m )
• E m i l y R a l s t o n ( E m i l y . R a l s t o n @ q u i d . c o m )
Quid Contacts For Further Questions
34. The algorithm uses textual similarities to identify documents that are similar to each other. It
then creates a network based on these similarities so that the user can visualize these
similarities as a network of clusters. To do this, Quid leverages proprietary NLP algorithms
and unsupervised machine learning to automate the topical generation.
The dots, or nodes, represents individual companies (or articles), and the links represent
semantic similarity between two nodes, with the clusters (groupings) differentiated by colors
representing the topics.
Readers can refer to Olivia Fischer, Jenny Wang (2019). Innovation and
Convergence/Divergence: Searching the Technology Landscape for more details on the
methodology.
Network Methodology
35. 1.8 M news sources are indexed in order to search across the database with Boolean logic. Meta data such as
publication data, publication count, source demographics, social engagement, and more can be utilized as filters and
tags.
News and Blogs data is updated on a 15 minute basis. Trends are based on reading any text to identify key words,
phrases, people, companies, and institutions; then compare different words from each document (news article) to
develop links between these words based on similar language.
This process is repeated at an immense scale which produces a network that shows how similar all documents are.
Search, Data Sources, and Scope
36. News & Blogs data from LexisNexis is one of three integrated datasets in the Quid platform. These sources include “top
tier”(e.g. New York Times, Washington Post, etc.), “mid tier” (e.g. Florida Post, Herald Chronicle, etc), and “other” (e.g.
blogs).
Data sourced from:
Quid indexes 1.8 million news sources allowing you to search the entirety of a news article, while filtering by included
metadata ranging from social engagement to source statistics such as count, location, demographics, and more. News
information is updated on a 15 minute basis. 10,000 articles can be analyzed within one network.
Data
37. Cluster Color Key
● Data Privacy 9.8%
●
AI & Data Science
Thought Leaders
9.6%
● Ethical AI 7.6%
● Artificial Intelligence 6.7%
●
Machine Learning
Explainability
5.2%
●
EU Commission AI
Ethics Guidelines
4.7%
● Hiring Bias 4.6%
●
Google Dissolves AI
Ethics Council
4.5%
●
Facial Recognition
Bias
3.9%
●
Machine Learning
Bias
3.6%
●
Gender Bias, Gender
Diversity
3.2%
● AI in Government 3.1%
● AI Ethics Framework 2.8%
● China Tech Talk 2.7%
●
Machine Learning
Fairness
2.5%
● Health Data Privacy 2.4%
● Google Walkout 2.4%
●
Unesco Defends
Artificial Intelligence
2.3%
●
IBM Introduces AI
Explainability
2.3%
● Human Rights & AI 2.3%
● EU AI Guidelines 1.9%
● UK Government & AI 1.5%
●
Oxford University AI
Study
1.4%
●
Google Announces
AI Ethics Council
1.2%
●
Google Announces
AI Ethics Council
1.2%
● AI 's Diversity Crisis 0.99%
● AI Fairness 0.96%
●
Facebook Backs
TUM's AI Ethics
Institute
0.96%
● Pope + Brad Smith 0.75%
●
Apple Acquirers Silk
Labs
0.67%
● Public Trust 0.67%
● Amazon's Role in
Fairness in AI
0.52%
● Employers & AI 0.52%
●
IAG Invests in Ethics
of AI
0.41%
●
Gartner Report: AI
Misconceptions
0.35%
●
Singapore Wins For
AI Governance &
Ethics
0.35%
● Temenos Acquires
Logical Glue
0.32%
●
Twitter Acquirers
Fabula AI
0.26%
38. D I G E S T L A N G U A G E ,
N O T J U S T N U M B E R S
Easily leverage our proprietary data
science for AI-driven insights
V I S U A L I Z E D A T A A T
S C A L E F O R C O N T E X T U A L
U N D E R S T A N D I N G
Quickly gain context around any topic with
long-form content analysis
Q U I D H E L P S Y O U
S U R F A C E S P E C I F I C
I N S I G H T S
Understand the data through unique
views for deep research and discovery
S H A R E Y O U R
A N A L Y S I S
Quid helps you build reports to influence
strategic decisions
H o w D o e s
Q u i d W o r k ?
Quid software reads
millions of documents and
offers immediate insight
by organizing that content
visually.
We power human intuition
with machine intelligence,
enabling organizations to
make decisions that matter.
39. HOW TO
READ A
QUID
NETWORK
Nodes - Each node
represents one data point
Isolated nodes - Not
clustered due to lack of
available text or very niche
languageClusters – Nodes that
are bound together by
unique keywords
(semantic clustering
based on unstructured
text data)
Far proximity – Cluster is quite niche
Nodes in this cluster have few
connections to other clusters
Link – A link exists between
nodes if they share common
keywords in their text
Close proximity – Displays high
degree of relevancy and
integration between two themes
Node with potential
fresh new crossover
across multiple themes
40. Traditional search returns lists Quid identifies themes and organizes them visually
Enterprise IOT
(15%)
IoT Driving Change
(13%)
IoT Startups
(11%)
Cyber Security
(10%)
Performance & Design
(10%)
International Innovation
(10%)
CES Previews
(4.9%)
IoT Opportunities
(4.6%)
Network Implications
(4.4%)
183,000,000 results
Events & Conferences
(5.4%)
41. WHAT IS
QUID USED
FOR?
USE CASE USE CASE DESCRIPTION Sample Question
Emerging Tech
Identify new technology trends and “hot spots” emerging within
them
What does the technology landscape
for digital retail look like?
Adjacency Opportunities Understand how specific technologies may be applied in new ways
What are the applications of my
autonomous driving technology and
where can they be applied?
Competitive Intelligence
Understand where competitors are active to identify key
battlegrounds
What payments offerings are my
competitors providing?
Trend Analytics Identify novel trends that will shape the future at a macro-level
What are the key conversations
around ageing?
M&A Screening Identify acquisition or partnership targets within a sector of interest
How can I leverage data from my
competitors’ acquisition strategy to
inform my M&A strategy?
Partnerships Analysis
Identify subtle corporate partnerships to get a better pulse on
industry initiatives
Who is Tesla partnering with and are
they expanding beyond the automotive
industry?
Brand Perception
Map media conversations and understand how competitors are
differentiated
Which brands are leading the current
narrative around skin care?
Consumer Sentiment
Analyze online forums to see what customers are really saying or
thinking
What are the top concerns of diabetic
patients as discussed on online forums?
Survey Analysis
Map open-text survey responses to understand the key themes and
novel insights
How can I leverage the free-text survey
responses from my customers to
improve my service offerings?
KOL Analysis Identify Key Opinion Leaders in the news media
Is discussion of our brand influenced in
the media primarily by internal or
external stakeholders?
42.
43.
44.
45.
46.
47.
48.
49. Count LinkedIn members
with AI skills who added a
new employer.
Normalize to the count of
LinkedIn members in the
country.
Index to the average month
in 2015-2016 to allow for fair
cross-time comparisons.
An index of 1.05
indicates a hiring rate
that is 5% higher than
the average month in
2015-2016.
50. Implications Workforce adaptability and resilience through, for instance, career transitions.
Source: LinkedIn’s AI Hiring Rate, Stanford AI Report (2019).
51. Financial Analysis,
Finance, and
Forecasting are the top
most representative
skills for economists.
Frequency-based Skills Genome
for an Economist
Management Msft Excel Msft Word
TF-IDF Skills Genome for an
Economist
Financial
Analysis
Finance Forecasting
Count all skill adds by
LinkedIn members in a given
entity (job, geo, industry, etc).
Re-weight skill frequency
using a TF-IDF model to get
most representative skills.
Compute the share of skills
that belong to AI (or other skill
groups) out of the top 50 skills
in the selected entity.
53. Build the Skills Genome of
an entity (incl. inclusion
dimension such as gender).
Pick a relevant benchmark
(eg. global average).
Compute relative
penetrations as the ratio
between a country
penetration rate and the
benchmark penetration
rate, controlling for
occupations.
A relative AI Skills
Penetration of 1.5
indicates the average
occupation in that
country uses 1.5 times
the intensity of the
average occupation in
the average country.
All Occupations on
LinkedIn
Occupations in
Benchmark
Occupations in
country A
60. Key Themes
▪ Everyone Wants to Innovate and Maybe
Disrupt
▪ Early Days of Deference To or Faith in
New Tech Are Over
▪ New Concerns and Old Responsibilities
▪ How Can We Create New Value AND
Maintain Trust and Success?
2
61. Agenda
▪ Gain an overview of how Law and Data &
Algorithmic Driven Innovation Intersect
▪ Use Data & Algorithmic Innovation Issues to Drill
Into the Legal Issues
▪ Propose Industry and Government Actions to
Address Problems of Data and Algorithmic
Innovation
3
62. ABOUT ME
▪ Currently: Associate Professor Law and Ethics, Georgia
Institute of Technology, Schller College –
deven.desai@scheller.gatech.edu
▪ Berkeley, A.B., Yale, J.D.
▪ Litigated early Internet and Tech Law at Quinn Emanuel,
LLP
▪ Worked at start-up tech firm, and then Mattel
▪ Cory Booker for Mayor, Policy and Finance
▪ Research Fellow, Princeton Center of Information
Technology Policy
▪ Google, Inc. Academic Research Counsel – Policy team
▪ Identify upcoming law, ethics, policy issues
▪ Liaise among academics, government, and industry
4
67. Data and Algorithmic
Innovation Seems Empty
▪ FB – sharing data with Spotify and Netflix
▪ UBER and VW – software to evade
government
▪ Voting Machines – Are they secure?
▪ Employment decisions, Government
Benefits, etc. determined by machine
▪ Google – People want to know how it
works
▪ Jobs lost to algorithmic industries
WHAT DO THE REST OF US GAIN?
9
70. Melinda Gates on Innovation
[S]ometimes the tech world thinks the solution is to
give somebody an app. Well, that’s not going to
change everything. I would also love to see more
tech innovation on behalf of the world. “Let’s
create the next thing that tracks my dog” — that’s
fun and nice, but come on, there are people dying.
-NY Times Magazine, April 15, 2019
12
71. AUTONOMOUS VEHICLES –
DIFFERENT SOCIAL VALUE
SAME PROBLEMS
Source: WikiCommons; Author Dllu; License: Creative Commons
Attribution Sharealike 4.0 International
72. GREAT EMPTINESS
“But We Increase
CONSUMER WELFARE”
is
No longer a good or
full answer
Deven R. Desai, The Chicago School Trap in Trademark:
The Co-Evolution of Corporate, Antitrust, and Trademark Law
37 CARD. L. REV. 551 (2015)
73. New Frontier for Social,
Political, and Commercial:
Data, Privacy, and AI/ML
GDPR –
EU Law that is quite strict about:
▪ Use of data,
▪ Right to explanation,
▪ Right to erasure,
▪ Data portability, and
▪ Rectification (or correction)
15
74. WHY?
▪EU is weak at technology
innovation
▪Revenge and trade-war
style mindset
INCORRECT EXPLANATION
75. New Frontier for Social,
Political, and Commercial:
Data, Privacy, and AI/ML
▪ GDPR – EU Law that is quite strict about
use of data AND
▪ California Consumer Protection Law
▪ Right to see what is collected; request
deletion, access about to whom data is
sold; and demand companies not sell
data
What’s odd here?
17
76. GDPR AND CA LAW
LESS OR NO DATA
And maybe
ALGORITHMIC
TRANSPARENCY?
78. CURRENT MINDSET OR CHALLENGE
FOR A COMPANY
▪ Evolve with Data and Technology or
Die
▪ WHAT TO DO? (Note Marriott taking
on Airbnb) 🡪 COMPETE
▪ Seems to Require Breaking Laws to
Disrupt and Compete
79. BETTER MINDSET OR CHALLENGE
FOR A COMPANY
▪ Evolve with Data and Technology
AND
▪ BUILD TRUST INTO THE SYSTEMS
80. Trust But Verify,
A Guide to Algorithms and the Law
Deven R. Desai & Joshua A. Kroll
31 Harvard Journal of Law & Technology 1 (2017)
81. EXAMPLE:
“Good” and “Bad” Discrimination
▪ OK to use credit scores, or other measure
correlated with race/gender, for other purposes?
▪ ”Good” discrimination
▪ Efficient scoring, tied to risk
▪ “Bad” discrimination
▪ Bad Data to Start can lead to bad outcomes
(garbage in/garbage out)
▪ Example: City of Boston’s pothole detection mobile app –
only detected problem in wealthy part of town because
that area had more cell phones
▪ Feedback loops: Send more police to precincts with
more crimes last year?
Slide adapted from and courtesy of Prof. Peter Swire
82. EXAMPLE:
“Good” and “Bad” Discrimination
▪ The more analytics is important, the
more people will care about
▪ DATA
▪ PRIVACY and
▪ THE ALGORITHMS BEHIND THE SCENES
▪ Ignoring the problem not a good
coping strategy
Slide adapted from and courtesy of Prof. Peter Swire
83. Ways To Mitigate Risks on
Algorithmic Driven Innovation:
INDUSTRY ACTIONS
Test the outcomes to see if there are
statistically large variations based on
sensitive categories – race, national origin,
gender, other things
25
Slide adapted from and courtesy of Prof. Peter Swire
84. ▪ Have “ethics review boards” on analytics
projects, like Institutional Review Boards
for medical research
▪ Examine practices for lending, housing,
and employment effects so that analytics
does not lead to clearly disparate
outcomes
26
Slide adapted from and courtesy of Prof. Peter Swire
Ways To Mitigate Risks on
Algorithmic Driven Innovation:
INDUSTRY ACTIONS
85. Ways To Mitigate :
INDUSTRY and GOVERNMENT ACTIONS
▪ Technical Accountability – show that methods
are OK – Industry can offer; Government can
require
▪ PROBLEM– current software was not built with
transparency or accountability in mind
▪ SOLUTION:
▪ BUILD all new software for Technical Accountability –
show that the software meets technical specifications
of law or other interest
▪ REMEMBER – SOME, but not ALL, ML Techniques
and Deep Learning are difficult, possibly not able,
to be explained
27
See Deven R. Desai & Joshua A. Kroll, Trust But Verify, A Guide to Algorithms and the Law,
31 Harvard Journal of Law & Technology 1 (2017)
86. ▪ Whistleblower Statutes
▪ ISSUE: Testing software to show that it will not give a
prohibited answer quite difficult if not possible
▪ ISSUE: Person or company builds software to
perform banned act
▪ SOLUTION:
▪ Pass a law to encourage and protect Whistleblowers
▪ Insiders can reveal that bad behavior was built into
software but whistleblowers need protection
▪ Reward money encourages telling AND addresses
loss of income for doing the right thing (often cannot
get a new job)
28
See Deven R. Desai & Joshua A. Kroll, Trust But Verify, A Guide to Algorithms and the Law,
31 Harvard Journal of Law & Technology 1 (2017)
Ways To Mitigate Risks:
GOVERNMENT ACTIONS
87. ▪ Public Interest Cause of Action
▪ Example Industries – Automotive and Voting
Machines – State sets civil penalties for violations
(maybe $2,500 per violation, maybe more if needed)
▪ SOLUTION:
▪ Private citizens police industry and notify the State
Attorney of a possible violation
▪ Private Citizen submits Certificate of Merit, swearing
and explaining good faith belief of violation
▪ State has time (e.g., 60 days) to chose to pursue
▪ If State does not pursue, private citizen is allowed to
sue
29
See Deven R. Desai & Joshua A. Kroll, Trust But Verify, A Guide to Algorithms and the Law,
31 Harvard Journal of Law & Technology 1 (2017)
Ways To Mitigate Risks:
GOVERNMENT ACTIONS
88. SUMMARY
▪ DEFERENCE TO INDUSTRY IS ENDING
▪ WE NEED TO HAVE TRUSTED SYSTEMS
▪ INDUSTRY CAN ACT TO ADDRESS
GOVERNANCE BY BUILDING SYSTEMS
PEOPLE CAN
▪ UNDERSTAND and
▪ TEST FOR PROBLEMS
▪ GOVERNMENT CAN PASS LAWS TO ENSURE
BAD ACTORS ARE DISCOVERED AND
PROSECUTED
89. In Simplest Terms – Co-Create Value
▪ “Treating a person like a resource is an
error.
▪ Simply saying you signed this contract, and
we can do what we like is an error.
▪ Better practices on these fronts from the
government, university researchers, and …
industries would allow people to trust—trust
that science [and industry] is interested in
progress [for all] and not only in profit.”
Deven R. Desai, Privacy? Property?: Reflections on the Implications of a
Post-Human World, 18 Kansas Journal of Law and Public Policy 174 (2009)
93. Definitions:
What is a killer robot?
“A fully autonomous weapon system that
selects & engages targets without meaningful
human control over the use of force.”
94. Risks of fully autonomous weapons
▪ Lack judgement for proportionality & distinction
▪ Unintended + unpredictable algorithmic interactions
▪ Proliferation to non-state actors
▪ Sparking accidental conflicts
▪ Prolonging conflicts & increasing their scale
▪ Vulnerable to hacking
▪ No justice or recourse for victims
105. “It is DoD policy that autonomous and
semi-autonomous weapon systems shall
be designed to allow commanders and
operators to exercise appropriate levels
of human judgment over the use of force.”
106.
107. Public attitudes towards work &
emerging technologies
Monica Anderson
Associate Director, Internet & Technology
Research
109. Americans generally express more worry than enthusiasm
about four types of tech developments
Source: Survey conducted May 1-15, 2017.
110. Americans anticipate that widespread job automation will
lead to greater income inequality
Source: Survey conducted May 1-15, 2017.
111. Publics across the world are generally skeptical about
workforce automation
Source: Spring 2018 Global Attitudes Survey and survey of U.S. adults conducted May 1-15, 2017.
112. Back in the U.S., Americans are split on who should be
responsible for displaced workers
Source: Survey conducted May 1-15, 2017.
113. But there is somewhat stronger support for limiting the
number of jobs that companies can replace
Source: Survey conducted May 1-15, 2017.
114. Certain occupations viewed as more at risk than others –
but “my own job” is seen as relatively safe
Source: Survey conducted May 1-15, 2017.
% of U.S. adults who think
it likely that the following
jobs will be replaced by
robots or computers in
their lifetimes
115. But even as the public thinks their job is less likely to be at risk–
a small share of Americans have already felt the impact
Source: Survey conducted May 1-15, 2017.
% in each group who say they have ___ because of automation of their job duties by their employer
Age
Household
income
Employment
status
116. 10
Thank you! Questions?
Monica Anderson
Associate Director, Internet and Technology Research
manderson@pewresearch.org
@MonicaRAnders
119. Human Capital
We study & work to
influence the manner
and extent to which
changes to the enabling
environment can
improve policy towards
practical
evidence-informed
actions for
development.
Enabling Environment
We study & work to
influence investments and
practices that can result in
improvements to the civil
service to strengthen
formulation and delivery
through evidence-informed
approaches.
We study and work to influence
strategic increases in sustainable
investments for human capital
and the enabling environment to
strengthen government ability to
formulate, implement and
review practical
evidence-informed actions for
development.
Sustainable Financing
Our Focus Domains
122. Our Use Case
Name: Food & Nutrition Security Early Warning System
Type: Decision Support
Data Sources: Satellite Data (Sentinel-2, MODIS, GPM), Local
commodity prices, food composition tables, DHS, Census data and
weather data, crop calendars
Outputs: Email alerts, policy maker dashboard, data journalist’s portal,
SMS alerts
Users: Government, smallholder farmers, journalists
Stage: Concept
126. Dietvorst, et al., 2014
Dzindolet, et al., 2002
Keeffe et al., 2005
Promberger & Baron, 2006
Sinha & Swearingen, 2001
Yeomans, et al., 2018
Empirical Evidence
Dijkstra et al.,1998
Dijkstra, 1999
Prefer AlgorithmPrefer People
127. No Deception
Incentivized
Judge Advisor System
Link to pre-registrations, materials, & data posted on OSF,
found at:
www.jennlogg.com/papers.html
Initial
Estimate
Advice
Final
Estimate
(Sniezek & Buckley, 1995)
128. Make initial, numeric estimate
0 30 60 90 100
Initial
Estimate
Advice
Final
Estimate
Initial
134. Same advice labeled as from…
Other People
“The average estimate of
participants from a past study
was: 163 pounds.”
Algorithm
“An algorithm ran calculations
based on estimates of participants
from a past study. The output that
the algorithm computed as an
estimate was: 163 pounds.”
135. 0
0.1
0.2
0.3
0.4
0.5
0.6
Weight Age Songs Matchmaker Researcher
Predictions
Weight on
Advice (0-1)
N = 202 N = 671 N = 215 N = 286 N = 199
Human Condition
Algorithm Condition
Participants relied more on algorithmic advice,
which improved their accuracy
*********** ***
*** p < .001
** p < .01
136. What is this person's age in the photograph?
0
0.1
0.2
0.3
0.4
0.5
0.6
Weight Age Songs Matchmaker Researcher
Predictions
Weight on
Advice (0-1)
N = 202 N = 671 N = 215 N = 286 N = 199
Human Condition
Algorithm Condition
*********** ***
*** p < .001
** p < .01
137. What rank will ‘Perfect’ by Ed Sheeran place on the
Billboard Magazine ‘Hot 100’ this week?
0
0.1
0.2
0.3
0.4
0.5
0.6
Weight Age Songs Matchmaker Researcher
Predictions
Weight on
Advice (0-1)
N = 202 N = 671 N = 215 N = 286 N = 199
Human Condition
Algorithm Condition
******** ***
*** p < .001
** p < .01
***
138. ***
0
0.1
0.2
0.3
0.4
0.5
0.6
Weight Age Songs Matchmaker Researcher
Predictions
Weight on
Advice (0-1)
N = 202 N = 671 N = 215 N = 286 N = 199
How funny…
How attractive…
How much enjoy dinner date….
******** ***
*** p < .001
** p < .01
Human Condition
Algorithm Condition
139. N = 199
Read What
Turkers Did
Click
Through
Materials
Read
Description
of DV
Predict
Turkers
Responses
Initial
Estimate
Advice
Final
Estimate
People Alg
IncentivizedAverage Response
To Advice
Matchmaker Study
Researchers predicted how participants responded to
advice
140. Results from 1A – 1D: Algorithm Appreciation
Researchers predicted: Algorithm Aversion
0
0.1
0.2
0.3
0.4
0.5
0.6
Weight Age Songs Matchmaker Researcher
Predictions
Weight on
Advice (0-1)
N = 202 N = 671 N = 215 N = 286 N = 199
Human Condition
Algorithm Condition
*********** ***
*** p < .001
** p < .01
141. People show “algorithm appreciation”
Visual Estimates: Weight, Age
Predictions: Popularity of Songs, Attraction
Researchers predict “algorithm aversion”
Summary of Findings
Logg, Minson, Moore, 2018 (OBHDP)