[House Hearing, 116 Congress]
[From the U.S. Government Publishing Office]


                PERSPECTIVES ON ARTIFICIAL INTELLIGENCE:
                       WHERE WE ARE AND THE NEXT
                       FRONTIER IN FINANCIAL SERVICES

=======================================================================

                                HEARING

                               BEFORE THE

                 TASK FORCE ON ARTIFICIAL INTELLIGENCE

                                 OF THE

                    COMMITTEE ON FINANCIAL SERVICES

                     U.S. HOUSE OF REPRESENTATIVES

                     ONE HUNDRED SIXTEENTH CONGRESS

                             FIRST SESSION

                               __________

                             JUNE 26, 2019

                               __________

       Printed for the use of the Committee on Financial Services

                           Serial No. 116-37
                           
 [GRAPHIC NOT AVAILABLE IN TIFF FORMAT]    
 
                               __________
                               

                    U.S. GOVERNMENT PUBLISHING OFFICE                    
39-737 PDF                  WASHINGTON : 2020                     
          
--------------------------------------------------------------------------------------


                 HOUSE COMMITTEE ON FINANCIAL SERVICES

                 MAXINE WATERS, California, Chairwoman

CAROLYN B. MALONEY, New York         PATRICK McHENRY, North Carolina, 
NYDIA M. VELAZQUEZ, New York             Ranking Member
BRAD SHERMAN, California             PETER T. KING, New York
GREGORY W. MEEKS, New York           FRANK D. LUCAS, Oklahoma
WM. LACY CLAY, Missouri              BILL POSEY, Florida
DAVID SCOTT, Georgia                 BLAINE LUETKEMEYER, Missouri
AL GREEN, Texas                      BILL HUIZENGA, Michigan
EMANUEL CLEAVER, Missouri            SEAN P. DUFFY, Wisconsin
ED PERLMUTTER, Colorado              STEVE STIVERS, Ohio
JIM A. HIMES, Connecticut            ANN WAGNER, Missouri
BILL FOSTER, Illinois                ANDY BARR, Kentucky
JOYCE BEATTY, Ohio                   SCOTT TIPTON, Colorado
DENNY HECK, Washington               ROGER WILLIAMS, Texas
JUAN VARGAS, California              FRENCH HILL, Arkansas
JOSH GOTTHEIMER, New Jersey          TOM EMMER, Minnesota
VICENTE GONZALEZ, Texas              LEE M. ZELDIN, New York
AL LAWSON, Florida                   BARRY LOUDERMILK, Georgia
MICHAEL SAN NICOLAS, Guam            ALEXANDER X. MOONEY, West Virginia
RASHIDA TLAIB, Michigan              WARREN DAVIDSON, Ohio
KATIE PORTER, California             TED BUDD, North Carolina
CINDY AXNE, Iowa                     DAVID KUSTOFF, Tennessee
SEAN CASTEN, Illinois                TREY HOLLINGSWORTH, Indiana
AYANNA PRESSLEY, Massachusetts       ANTHONY GONZALEZ, Ohio
BEN McADAMS, Utah                    JOHN ROSE, Tennessee
ALEXANDRIA OCASIO-CORTEZ, New York   BRYAN STEIL, Wisconsin
JENNIFER WEXTON, Virginia            LANCE GOODEN, Texas
STEPHEN F. LYNCH, Massachusetts      DENVER RIGGLEMAN, Virginia
TULSI GABBARD, Hawaii
ALMA ADAMS, North Carolina
MADELEINE DEAN, Pennsylvania
JESUS ``CHUY'' GARCIA, Illinois
SYLVIA GARCIA, Texas
DEAN PHILLIPS, Minnesota

                   Charla Ouertatani, Staff Director
                   
                 TASK FORCE ON ARTIFICIAL INTELLIGENCE

                    BILL FOSTER, Illinois, Chairman

EMANUEL CLEAVER, Missouri            FRENCH HILL, ARKANSAS, Ranking 
KATIE PORTER, California                 Member
SEAN CASTEN, Illinois                BARRY LOUDERMILK, Georgia,
ALMA ADAMS, North Carolina           TED BUDD, North Carolina
SYLVIA GARCIA, Texas                 TREY HOLLINGSWORTH, Indiana
DEAN PHILLIPS, Minnesota             ANTHONY GONZALEZ, Ohio
                                     DENVER RIGGLEMAN, Virginia
                            
                            
                            C O N T E N T S

                              ----------                              
                                                                   Page
Hearing held on:
    June 26, 2019................................................     1
Appendix:
    June 26, 2019................................................    33

                               WITNESSES
                        Wednesday, June 26, 2019

Buchanan, Bonnie, Head of Department of Finance and Accounting, 
  Full Professor of Finance, Surrey Business School, The 
  University of Surrey...........................................     6
McWaters, R. Jesse, Financial Innovation Lead, World Economic 
  Forum..........................................................    10
Merrill, Douglas, Founder and CEO, ZestFinance...................     8
Turner-Lee, Nicol, Fellow, Center for Technology Innovation, 
  Brookings Institution..........................................     4

                                APPENDIX

Prepared statements:
    Buchanan, Bonnie.............................................    34
    McWaters, R. Jesse...........................................    46
    Merrill, Douglas.............................................    54
    Turner-Lee, Nicol............................................   109

              Additional Material Submitted for the Record

Budd, Hon. Ted:
    GAO report entitled, ``Insurance Markets: Benefits and 
      Challenges Presented by Innovative Uses of Technology,'' 
      dated June 2019............................................   127
Hill, Hon. French:
    ZestFinance article entitled, ``Clarifying Why SHAP Shouldn't 
      Be Used Alone''............................................   170

 
                       PERSPECTIVES ON ARTIFICIAL
                       INTELLIGENCE: WHERE WE ARE
                        AND THE NEXT FRONTIER IN
                           FINANCIAL SERVICES

                              ----------                              


                        Wednesday, June 26, 2019

             U.S. House of Representatives,
             Task Force on Artificial Intelligence,
                           Committee on Financial Services,
                                                   Washington, D.C.
    The task force met, pursuant to notice, at 10 a.m., in room 
2128, Rayburn House Office Building, Hon. Bill Foster [chairman 
of the task force] presiding.
    Members present: Representatives Foster, Casten, Adams, 
Garcia of Texas, Phillips; Hill, Loudermilk, Budd, 
Hollingsworth, Gonzalez of Ohio, and Riggleman.
    Also present: Representative Himes.
    Chairman Foster. The Task Force on Artificial Intelligence 
will now come to order.
    Without objection, the Chair is authorized to declare a 
recess of the task force at any time.
    Also, without objection, members of the full Financial 
Services Committee who are not members of this task force are 
authorized to participate in today's hearing, consistent with 
the committee's practice.
    Today's hearing is entitled, ``Perspectives on Artificial 
Intelligence: Where We Are and the Next Frontier in Financial 
Services.''
    The Chair will now recognize himself for 5 minutes for an 
opening statement.
    Thank you, everyone, for joining us today at the first 
hearing of the House Financial Services Committee's Task Force 
on Artificial Intelligence. And I would like to begin by 
thanking Chairwoman Waters and Ranking Member McHenry for 
working to establish this important task force and reaffirming 
this committee's commitment to understanding technological 
innovation in the financial services sector.
    It is an exciting time to be on this committee. Today, the 
financial services sector is facing a period of rapid 
disruption and innovation, and artificial intelligence (AI) is 
at the heart of these changes.
    AI is transforming the way Americans live, work, and 
interact with each other. As members of this committee, it is 
incumbent upon us to engage with and to understand more deeply 
how it works, how it is designed and operated, and how it 
affects and may affect consumers.
    When done right, AI can mean innovative underwriting models 
that allow millions more people access to credit and financial 
services. And at a time when there are still over 50 million 
unbanked or underbanked Americans, this is a big deal. 
Companies are also using AI to execute trades, manage 
portfolios, and provide personalized services to customers.
    AI can be used to better detect fraud and money laundering, 
and regulators are using AI to improve market surveillance and 
policing of bad actors. This is important, because AI is also 
giving criminals more ways to impersonate customers and steal 
their assets and sensitive financial information.
    Last year, there were almost 15 million victims of identity 
fraud, costing Americans billions of dollars. Social security 
numbers, credit card numbers, and other personal identity 
factors can be stolen and sold on the dark web or used by 
criminals for quick and easy profit.
    That is why it is imperative that we come up with better 
ways of protecting and securing our digital identities online. 
In fact, I was just, in the last hour, giving a keynote speech 
at the Identiverse conference, where thousands of people come 
together each year to understand what technologies can be 
applied to allow both individuals and organizations to protect 
themselves from often AI-enabled identity fraud.
    And now, as the name of this hearing suggests, the other 
part of this equation that we need to explore is, where is this 
technology going and what are the next frontiers?
    To truly reach its potential to change the face of 
financial services, there are some questions we need to 
address. First, how can we be sure that AI credit underwriting 
models are not biased? Second, who is accountable if AI 
algorithms are just a black box that nobody can explain when it 
makes a decision? And third, AI runs on an enormous amount of 
data. Where does this data come from? How is it protected? Do 
customers know where it is being held, under what legal regime?
    Also, AI works far better with large datasets. Will these 
large datasets be one more factor driving the consolidation of 
financial services sectors? I worry frequently that small 
community banks may end up going the way of small community 
newspapers.
    Another thing we will be looking at is, how many and what 
kind of financial services jobs will AI displace? A recent 
study by Deloitte indicated that 75 percent of financial firms 
are planning to displace humans with technology, and this is 
probably not a trend that will slow down. And it is not only 
going to apply to bank tellers and entry-level people; it will 
apply to some of the very highest salaried positions.
    And as I mentioned, just the question about whether small 
banks and startups will be able to compete with the big tech 
firms, particularly when everyone is going to need access to 
these very large, personally identifiable datasets.
    Over the next 6 months, we will begin to examine these 
questions to gain a deeper understanding of how this technology 
is being used in the financial services industry. It is my hope 
that today's dialogue between our diverse and bipartisan group 
of Members and the expert panel of witnesses joining us will 
lead to a better understanding of how AI is changing the 
industry, how it can lead to innovative and inclusive products 
and more personalized customer experience, and how this 
technology will shape the questions that policymakers will have 
to grapple with in the coming years.
    And so at this time, I would like to recognize the ranking 
member of the task force, my colleague, Mr. Hill from Arkansas, 
who has been a valuable asset and a trusted bipartisan partner 
as we begin this important endeavor.
    Mr. Hill. I thank the chairman. I appreciate you convening 
the hearing today and selecting this excellent panel before us. 
And I, too, want to thank our mutual leaders, Chairwoman Waters 
and Ranking Member McHenry, for their partnership in creating 
this task force.
    Over the next few months, I look forward to working with 
you and our colleagues on both sides of the aisle to find ways 
to foster innovation through the use of artificial intelligence 
for both disruptive innovators and for our incumbent financial 
players, both small and large, as well as finding ways to use 
AI successfully to enhance our compliance obligations among our 
regulatory agencies.
    The use of AI has grown exponentially in the last few 
years. AI has the potential to improve human life, economic 
competitiveness, and societal challenges.
    Recent GAO testimony identified four high-consequence 
sectors where leveraging AI will bring significant benefits: 
cybersecurity; automated vehicles; criminal justice; and 
financial services. And today's timely hearing will discuss how 
AI is impacting and influencing financial services.
    Artificial intelligence can be used to gather enormous 
amounts of data, detect abnormalities, and solve complex 
problems. Financial institutions are already experimenting 
extensively with AI strategies to enhance and streamline 
financial institutions, BSA and AML compliance, CRA 
requirements, fraud detection, and real estate valuations, all 
while reducing cost levels.
    Also, AI can create better efficiencies for underwriting 
and reaching underbanked communities. Algorithmic-driven 
lending is proliferating online and transforming everything 
from personal loans to small business credit extension. A 
recent National Bureau of Economic Research working paper found 
that online financial companies discriminate 40 percent less 
than loan officers who make decisions face-to-face.
    I know Dr. Merrill of ZestFinance, who grew up in my 
district in Arkansas, has been doing some interesting things in 
regard to AI and underwriting, and I look forward to hearing 
more from him today.
    All that to say that the use of artificial intelligence and 
machine learning is not without challenges and questions, just 
like any other technology.
    Dr. Henry Kissinger published an interesting article in The 
Atlantic recently outlining concerns about the rise of 
artificial intelligence. Dr. Kissinger argues that we are in 
the midst of a technological revolution that could culminate in 
a world ``relying on machines powered by data and algorithms 
and ungoverned by ethical or philosophical norms.'' He goes on 
to say that, ``Truth becomes relative and information threatens 
to overwhelm wisdom.'' Well, we are not into overwhelming 
wisdom in anything we do on Capitol Hill.
    While it remains to be seen whether Dr. Kissinger's 
concerns are fully proved, I think we should heed his advice. 
As policymakers, we need to ensure that we are asking the right 
questions about appropriate testing and evaluating of new 
technology, so that the ultimate benefits are, in the end, 
benefiting consumers.
    We need to ensure that AI does not create biases in lending 
toward discrimination and that prudential regulators and market 
participants have an understanding of the underlying 
technology, model validation, and how algorithmic decisions are 
being made and the manner of the audit trail. These questions 
must be analyzed.
    Lastly, I would be remiss if I didn't mention the potential 
of job losses connected with the advent of artificial 
intelligence. I am sure this topic will arise throughout our 
hearings during the Congress.
    The World Economic Forum argues that machines and 
algorithms in the workplace are expected to create 130 million 
new roles in work, but cost about 75 million jobs to be 
displaced by 2022, which means net 58 million jobs might be 
created. In my view, this will contribute positively on the 
economy and the future of work in the long run.
    People might be putting the cart before the horse on the 
number of net displacements. I start this journey in the ``cup 
half full'' camp, and I am optimistic about our future.
    I look forward to continuing to seek out answers throughout 
our work on the task force. I thank my good friend, Dr. Foster, 
for his partnership. And I look forward to finding bipartisan 
solutions to these many interesting and challenging questions 
in financial services.
    I yield back.
    Chairman Foster. Thank you.
    Today, we welcome the testimony of Dr. Nicol Turner-Lee, 
fellow at the Center for Technology Innovation, Brookings 
Institution; Dr. Bonnie Buchanan, head of the School of Finance 
and Accounting and full professor of finance at the Surrey 
Business School, University of Surrey; Dr. Douglas Merrill, 
founder and CEO of ZestFinance; and Mr. Jesse McWaters, 
financial innovation lead at the World Economic Forum.
    Witnesses are reminded that your oral testimony will be 
limited to 5 minutes, and without objection, your written 
statements will be made a part of the record.
    So, Dr. Turner-Lee, you are now recognized for 5 minutes to 
give an oral presentation of your testimony.

 STATEMENT OF NICOL TURNER-LEE, FELLOW, CENTER FOR TECHNOLOGY 
               INNOVATION, BROOKINGS INSTITUTION

    Ms. Turner-Lee. Thank you very much, distinguished members 
of the task force, and thank you for this opportunity to speak 
before you on artificial intelligence and the application of 
autonomous systems in the financial services sector.
    With a history of over 100 years, we at Brookings are 
committed to evidence-based nonpartisan research in this area, 
and my particular area of focus is on algorithmic bias. So, I 
appreciate the opportunity to speak before you.
    Increasingly, the public and private sectors are turning to 
AI and machine-learning algorithms to automate simple and 
complex decision-making processes. The mass scale digitization 
of data and the emerging technologies that use them are 
disrupting most economic sectors, including transportation, 
retail, advertising, financial services, and energy.
    These massive datasets have made it easy to derive new 
insights through computers, and as a result, machine-learning 
algorithms, which are step-by-step instructions that computers 
follow to perform a task, have become more sophisticated and 
pervasive tools for automated decision-making.
    While many of us are aware of the context in which they are 
used, from making recommendations about movies, to credit 
products, these models make inferences from data about people 
including their identities, their demographic attributes, their 
preferences, and their likely future behaviors, as well as the 
objects related to them. And from that data, it learns a model 
which then can be applied to other people and objects, making 
what they believe to be accurate predictions.
    But because machines can treat similarly situated people 
and objects differently, we are starting to reveal, much like 
has been said, some troubling examples in which the reality of 
algorithmic decision-making falls short of our expectations or 
is simply wrong.
    In the case of credit, we are seeing people denied credit 
due to the factoring of digital composite profiles, which 
include their web browsing histories, social media profiles, 
and other inferential characteristics in the factoring of 
credit models, and these biases are systematically finding 
themselves with less favor to individuals within particular 
groups where there is no relevant difference between those 
groups which justifies those harms.
    While my written testimony goes into more detail about 
this, I would just like to share in my remaining few minutes 
how we can create more fair, ethical, and just algorithmic 
models. From this perspective, if we do not do such at this 
time, we have the potential to replicate and amplify 
stereotypes historically prescribed to people of color and 
other vulnerable populations.
    Let me start with an initial truth about emerging 
technologies: Despite their greater facilitation of efficiency 
and cognition, the online economy has not resolved the issue of 
racial bias. And we see that in terms of search inquiries that 
have classified African Americans as primates in the past.
    These controversies are primarily due to the microtargeting 
of certain populations that go awry, even when they are not 
deliberate. Some of it can happen on an explicit level, where 
the algorithm may not start out being discriminatory in intent 
but adapts to the societal stereotypes and unfair profiling. In 
the case of credit, Latanya Sweeney at Harvard University has 
said that African Americans may find themselves the subject of 
higher-interest credit cards and other financial products 
simply because the computer has inferred their race.
    In the issue of implicit or unconscious bias, we simply do 
not have enough people working in this field to help us make 
the right decisions, which goes back to the inclusivity and the 
diversity and design of these models.
    Given this--and, again, in my written testimony I speak to 
the ways and the reasons of these biases, whether it is skewed 
training data, whether it is the fact that we have less 
counterfactual data that is actually going into training the 
algorithm--these issues are nonetheless troubling and 
dangerous, particularly for vulnerable populations like African 
Americans and Latinos, who have been ill-served within the 
financial services market. Most of these populations tend to be 
unbanked compared to whites, underbanked, and lack access to 
home ownership.
    If you think about the physical redlining that happens 
oftentimes offline, what does it mean, as Frank Pasquale has 
called weblining or applications discrimination, when we begin 
to look at the algorithmic economy?
    What do we do about this so that we avoid unfair credit 
rationing, exclusionary filtering, digital redlining? I would 
just like to offer just three recommendations that I would love 
to answer additional questions around that may be helpful.
    First and foremost, Congress must modernize civil rights 
laws and other consumer protections to safeguard protected 
classes from online discrimination. We have laws like the Equal 
Credit Opportunity Act, the Fair Housing Act, and other laws, 
which I feel have to be modernized in the digital age to ensure 
equity and fairness.
    We also need companies to exercise self-regulatory 
behaviors, whether it is looking at the auditing of their 
algorithms, bringing in more human content moderators, or 
finding ways to advance exclusivity.
    And finally--and I will save this again for questions--I 
think it is important that we are more deliberate in bringing 
in diverse populations, partnering with Historically Black 
Colleges and Universities (HBCUs) and other minority-serving 
institutions, to ensure that we have more people at the table 
in the design of these models.
    Thank you very much, and I look forward to questions.
    [The prepared statement of Dr. Turner-Lee can be found on 
page 109 of the appendix.]
    Chairman Foster. Thank you.
    Dr. Buchanan, you are now recognized for 5 minutes to give 
an oral presentation of your testimony.

STATEMENT OF BONNIE BUCHANAN, HEAD OF DEPARTMENT OF FINANCE AND 
ACCOUNTING, FULL PROFESSOR OF FINANCE, SURREY BUSINESS SCHOOL, 
                    THE UNIVERSITY OF SURREY

    Ms. Buchanan. Thank you, Chairman Foster.
    Distinguished members of the task force, thank you for the 
opportunity to appear before you and provide testimony to help 
inform discussion about artificial intelligence in the 
financial services industry.
    I am Dr. Bonnie Buchanan, professor of finance at the 
University of Surrey Business School, and I will provide some 
insights on artificial intelligence, its applications in 
financial services, as well as its challenges and 
opportunities. And I hope we can all work together to address 
those challenges and opportunities.
    Artificial intelligence is rapidly impacting the financial 
services industry in a profound way, through banking, 
insurance, wealth management, personal financial planning, and 
regulation. It can be broadly thought of as a group of related 
technologies, including machine learning and deep learning.
    Machine learning deals with general pattern recognition and 
universal approximation of relationships. One such example 
details teaching an algorithm to learn from past regulatory 
breaches and to predict new breaches, such as insider trading 
or cartels.
    Regulators use clustering algorithms to better understand 
trades and categorize bank business models in advance of 
regulatory examinations. Chatbots, powered by natural language 
processing algorithms, have become powerful tools which provide 
a personalized and conversational experience to users.
    Deep-learning algorithms automate routine tasks, mitigate 
risk, and help prevent fraud. It is based on neural networks, 
which are based on mimicking the way the multiple layers of the 
brain's neurons work. And neural networks have been used in 
financial distress models.
    Artificial intelligence offers the possibility of greater 
financial inclusion, but its rapid growth and an already very 
complex financial system presents major challenges regarding 
regulation and policymaking, and risk management, as well as 
ethical, economic, and social hurdles. For one, the financial 
services workplace is going to look very different in the short 
and long term, with artificial intelligence augmenting many 
positions.
    Machine-learning algorithms can also potentially introduce 
bias and discrimination. Deep learning provides predictions, 
but it does lack insight as to how the variables are being used 
to reach these predictions. Hiring and credit-scoring 
algorithms can exacerbate inequities due to biased data. 
Policymakers need to be concerned about the explainability of 
artificial intelligence models, and we should avoid black-box 
modeling where humans cannot determine the underlying process 
or outcomes of the machine-learning or deep-learning 
algorithms.
    And resolving such issues as discrimination and bias 
requires being grounded in ethics and understanding what causes 
the bias in the algorithm in the first place. When it comes to 
artificial intelligence in financial services and a fairer 
future, policymakers need to be concerned about explainability, 
accountability, and, indeed, even auditability of artificial 
intelligence modeling.
    Many artificial intelligence techniques remain untested in 
a financial crisis scenario. My written testimony discusses 
several instances where algorithms implemented by financial 
firms appeared to act in ways quite unforeseen by their 
developers, leading to errors and flash crashes.
    Cybercrime costs the global economy over $400 billion, but 
many banks have started to successfully turn to artificial 
intelligence techniques to address fraud through AI-based voice 
phishing detection apps.
    Artificial intelligence and machine learning's rapid 
development are to such an extent where it is almost 
outstripping the current regulatory framework. But if we look 
overseas, we have in the United Kingdom the introduction of 
open banking, which gives consumers the ability to compare 
product offerings and exchange data between providers in a 
secure way.
    Under the General Data Protection Rules (GDPR), EU citizens 
have the right to receive an explanation for decisions based 
solely on automatic processing. Furthermore, GDPR stipulates 
that companies must first obtain consent from an EU citizen 
before using their data, and failure to comply with GDPR rules 
can result in substantial fines.
    The European Market in Financial Instruments Directive Part 
II requires that firms that apply artificial intelligence and 
algorithmic models have a robust development plan in place.
    As big data and computing power increases, artificial 
intelligence needs to be technically robust, secure, protect 
privacy, and be ethically sound and regulation-compliant. We 
must not forget the importance of better digital and financial 
literacy, and ultimately, it needs to emphasize financial 
inclusion.
    Thank you very much for your time today, and I appreciate 
the opportunity to share my thoughts with you later. Thank you.
    [The prepared statement of Dr. Buchanan can be found on 
page 34 of the appendix.]
    Chairman Foster. Thank you.
    Dr. Merrill, you are now recognized for 5 minutes to give 
an oral presentation of your testimony.

   STATEMENT OF DOUGLAS MERRILL, FOUNDER AND CEO, ZESTFINANCE

    Mr. Merrill. Chairman Foster, Ranking Member Hill, and 
members of the task force, thank you for the opportunity to 
appear before you to discuss the use of artificial intelligence 
in financial services.
    My name is Douglas Merrill. I am the CEO of ZestFinance, 
which I founded 10 years ago with a mission to make fair and 
transparent credit available to everyone.
    Lenders use our software to increase approval rates, lower 
defaults, and to make their lending fairer. Before ZestFinance, 
I was the chief information officer at Google. I have a Ph.D. 
in artificial intelligence from Princeton University.
    The use of artificial intelligence in the financial 
industry is growing. Today, I will discuss a type of AI, 
machine learning, also known as ML, that discovers 
relationships between many variables in a dataset to make 
better predictions.
    Because ML-powered credit scores substantially outperform 
traditional credit scores, companies will increasingly use ML 
to make more accurate decisions. For example, customers using 
our ML underwriting tools to predict creditworthiness have seen 
a 10 percent approval rate increase for credit card 
applications, a 15 percent approval rate increase for auto 
loans, and a 51 percent increase in approval rates for personal 
loans, each with no increase in defaults.
    Overall, this is good news and should be encouraged. 
Machine learning increases access to credit, especially for 
low-income and minority borrowers. Regulators understand these 
benefits and, in our experience, want to facilitate, not 
hinder, the use of ML.
    But at the same time, ML raises serious risks for 
institutions and consumers. ML models are opaque and inherently 
biased. Lenders put themselves, consumers, and the safety and 
soundness of our entire financial system at risk if they do not 
appropriately validate and monitor ML models.
    Getting this mix right, enjoying ML's benefits while 
employing responsible safeguards, is very difficult. 
Specifically, ML models have a black-box problem. Lenders know 
only that an ML algorithm made a decision, not why it made that 
decision.
    Without understanding why a model made a decision, bad 
outcomes will occur. For example, a used car lender we work 
with had two seemingly benign signals in their model. One 
signal was that higher-mileage cars tend to yield higher-risk 
loans. Another was that borrowers from a particular State were 
slightly less risky than those from other States. Neither of 
these signals raised compliance concerns.
    However, our ML tools noted that, taken together, these 
signals predicted a borrower to be African American and more 
likely to be denied.
    Without visibility into how seemingly fair signals 
interact, lenders will make decisions which tend to adversely 
affect minority borrowers.
    There are purported to be a variety of methods for 
understanding how ML models make decisions. Most don't actually 
work. As explained in our white paper and a recent essay on a 
technique called SHAP, both of which I have submitted for the 
record, many explainability techniques are inconsistent, 
inaccurate, computationally expensive, or fail to spot 
discriminatory outcomes.
    At ZestFinance, we have developed explainability methods 
that render ML models truly transparent. As a result, we can 
assess disparities in outcomes and create less discriminatory 
models. This means we can identify approval rate gaps in 
protected classes such as race, national origin, and gender, 
and then minimize or eliminate those gaps. In this way, 
ZestFinance's tools decrease disparate impacts across protected 
groups and ensure that the use of machine learning-based 
underwriting mitigates rather than exacerbates bias in lending.
    Congress could regulate the entirety of ML in finance to 
avoid bad outcomes, but it need not do so. Regulators have the 
authority necessary to balance the risks and benefits of ML 
underwriting.
    In 2011, the Federal Reserve, the OCC, and the FDIC 
published guidance on effective model risk management. ML was 
not commonly in use in 2011, so the guidance does not directly 
address best practices in ML model development, validation, and 
monitoring.
    We have recently produced a short FAQ, which we have also 
submitted for the record, that suggests updates to bring the 
guidance into the ML era. Congress must encourage regulators to 
set high standards for ML model development, validation, and 
monitoring.
    We stand upon the brink of a new age of credit, an age that 
is fairer and more inclusive, enabled by this new technology of 
machine learning. However, ``brink'' can also imply the edge of 
a cliff. Without rigorous standards for understanding why 
models work, ML will surely drive us over the edge. Every day 
that we wait to responsibly implement ML keeps tens of millions 
of Americans out of the credit system or poorly treated by it.
    Thank you so much for your time.
    [The prepared statement of Dr. Merrill can be found on page 
54 of the appendix.]
    Chairman Foster. Thank you.
    And, Mr. McWaters, you are now recognized for 5 minutes to 
give an oral presentation of your testimony.

  STATEMENT OF R. JESSE MCWATERS, FINANCIAL INNOVATION LEAD, 
                      WORLD ECONOMIC FORUM

    Mr. McWaters. Thank you.
    Chairman Foster, Ranking Member Hill, distinguished members 
of this task force, I am honored to be invited to appear before 
you today to discuss this important topic.
    I would like to share with you in a personal capacity key 
insights from an ongoing research initiative that I lead at the 
World Economic Forum. These findings are drawn from 18 months 
of interviews and workshops with leading thinkers from large 
financial institutions, fintech innovators, large technology 
firms, and regulatory authorities, from all around the world.
    It is manifestly clear that artificial intelligence is 
transforming the operating models of financial institutions. It 
is being deployed to improve the speed and efficiency of 
financial processes, to improve the accuracy of financial 
predictions, to create more accessible and personalized 
advisory capabilities, and to establish entirely new business 
offerings.
    Less visible, but even more important, are the potential 
long-term impacts of AI on the competitive dynamics of the 
financial ecosystem. As AI becomes more central to the 
differentiation strategies of financial institutions, their 
appetite for deeper and broader datasets will increase, making 
access to this data a competitive imperative for all financial 
institutions.
    Over time, artificial intelligence may even redraw the map 
of what we consider the financial sector. For example, small 
and midsized financial institutions which are unable to invest 
in becoming AI leaders may instead choose to employ the AI 
capabilities of third parties on an ``as a service'' basis. The 
providers of these services could be large technology firms, 
they could be specialized fintechs, or even competing financial 
institutions.
    Moreover, the tendency of AI businesses to rapidly scale 
via the so-called AI ``flywheel effect'' means that successful 
service providers of this kind could rapidly become central to 
the operations of many financial institutions, resulting in a 
deep change to the systemic structure of the financial system.
    These seismic shifts in the landscape of financial services 
obviously create new risks. The enormous complexity of some 
advanced AI systems can make them opaque, challenging 
traditional models of regulation and compliance.
    The use of ever broader datasets introduces risks to user 
privacy, as well as to the introduction of unintended bias into 
financial decision-making. Furthermore, an inherently 
specialized and interconnected financial system creates new 
vectors for both the accumulation and the propagation of 
systemic risk.
    However, while these threats are very real and should be 
taken seriously, it is critical that we avoid knee-jerk 
reactions informed by fear.
    In my view, the advent of AI does not call into question 
the fundamental principles that inform our regulatory 
framework. Rather, it demands that we be open to using both 
existing and emerging techniques to ensure that we remain 
aligned to these principles, even against a backdrop of rapid 
technological change.
    Moreover, AI's risks must be considered alongside the 
opportunities that it creates. AI has the potential to help 
motorists get the money that they need from an insurance claim 
more quickly after an accident, to help immigrants without an 
established credit history access financing, and to make high-
quality financial advice, so needed, more accessible for 
everyday Americans.
    Moreover, the ability to outsource selected functions to 
specialized third parties has the potential to help smaller 
community banks remain digitally relevant to their customers.
    Ultimately, AI is a tool. As with all powerful tools, 
preventing misuse is of the utmost importance. But with the 
right governance and oversight, I believe that AI has the 
potential to do enormous good for the financial sector.
    Thank you.
    [The prepared statement of Mr. McWaters can be found on 
page 46 of the appendix.]
    Chairman Foster. Thank you.
    And I now recognize myself for 5 minutes for questions.
    Dr. Turner-Lee and Dr. Merrill, the National Bureau of 
Economic Research Working Paper recently published by UC 
Berkeley found that the algorithmic lending models discriminate 
in their case 40 percent less than face-to-face lenders for 
mortgage and refinancing loans.
    If that sort of result proves generally true, it is 
positive news for consumers, especially African-American and 
Latino consumers, who pay $765 million in additional interest 
costs each year.
    And it highlights the fact that the artificial intelligence 
algorithms don't have to be perfect as long as they are 
significantly better than the current procedures. That is 
obviously a moving target, because as our underwriting gets 
better and more fair over time, I think we have to continue to 
ask machine-learning techniques to continually up their game as 
well.
    And so my question is, to what extent companies should be 
required to audit these algorithms so that they don't unfairly 
discriminate? Who should determine the standards for that? What 
is the current understanding of best practices?
    Ms. Turner-Lee. Thank you, Mr. Chairman, for that question.
    I am actually also delighted to see that we are seeing 
research that is actually saying that we are levering some of 
the disparities when it comes to the use of AI. But I, too, am 
cautious, because I think the institution of auditing practices 
are really what is needed to ensure that we are not seeing 
these unintended consequences of racial or ethnic bias against 
different economic classes actually happening.
    I would say to you that we are seeing more self-regulatory 
models where companies are actually coming in and engaging in 
auditing. I would also recommend, as I said earlier, that we 
see developers look at how the algorithm is in compliance with 
some of the nondiscrimination laws prior to the development of 
the algorithm, which would also help to audit out some bias at 
the onset.
    A paper that we recently released also combines auditing 
with a bias impact statement. There is a lot more proactive 
conversation prior to the launch of the product into the public 
domain.
    Chairman Foster. How close are we to having generally 
agreed-upon metrics for things like fairness? I remember 
encountering a paper that claimed to have 15 different 
definitions of fairness.
    Ms. Turner-Lee. Right.
    Chairman Foster. So, how do we decide which one of those is 
most applicable?
    Ms. Turner-Lee. That is a question with which I think all 
of us on this panel today struggle. How do you look at fairness 
and equity tradeoffs? Where do you find that there is a product 
that is not creating more discrimination versus less? And how 
do you document what those models are?
    I think at this stage, our discussion around explainability 
and accountability is one part of it. But I think, to your 
point, getting companies as well as consumers engaged, creating 
more feedback loops so that we actually go into this together, 
I think is a much more proactive approach than trying to figure 
out ways to clean up the mess and the chaos at the end where we 
are discriminating against more people, we are incarcerating 
more people, and we are denying credit to more people. We have 
to figure out how to get ahead of this game.
    Chairman Foster. Dr. Merrill?
    Mr. Merrill. I think it is quite clear that machine-
learning models are biased. They are biased for three primary 
reasons.
    First, they are biased because historically, white men have 
dominated the credit roles in the past, so that back data is a 
bad representation of the world.
    Second, they are biased because machine-learning models 
tend to use a large number of signals of variables and there 
has to date been relatively little best practice around, how do 
you analyze those variables, because many times one or more of 
them will covary to yield a protected class.
    And third, they are biased because most ML models are 
produced by the proverbial ``white guy in a hoodie.'' I, by the 
way, own a hoodie, but I try really hard not to be biased.
    I think, absolutely, we must have an audit requirement, and 
I actually think a creation up-front requirement, in the way 
that we today have build requirements for financial services. 
FCRA produces quite striking, quite clear laws on what we are 
allowed to do.
    I would hope that either through congressional intervention 
or regulatory intervention, we would come to a world in which 
there would be a language to describe what is acceptable before 
you build models and then an agreed-upon language at the end of 
models to show if, in fact, you have a bias problem, because 
again, the odds are good you are going to.
    Chairman Foster. Mr. McWaters and Dr. Buchanan, both of you 
have worked on the issue of whether or not the access to large 
datasets is going to drive consolidation. Dr. Buchanan, you 
have written on China, where they have simply let things 
consolidate and let the access to enormous amounts of data 
result in a very small number of very large players.
    Are there policy options that we can do to lean against 
that consolidation, in my negative 2 seconds? If you could just 
say one sentence, like, read my testimony or something?
    Ms. Buchanan. I do talk about this in my written testimony 
and also my Turing report, Chairman Foster.
    But I think we also have to understand what makes China so 
different, too. Its supply of data, its online population is 
twice the size of the United States. WeChat hosts over a 
billion users. And they have also--
    Chairman Foster. Okay. Now, I will have to; I am going to 
use my power of the gavel on myself.
    Ms. Buchanan. Yes, there are. We can, yes.
    Chairman Foster. All right.
    Now, I am happy to yield 5 minutes to Ranking Member Hill.
    Mr. Hill. Thank you, Mr. Chairman.
    This is a really good discussion, and I think that it is 
exactly why we have this task force, to talk through these 
issues.
    And also, we invite our regulators to be full participants. 
All of you have made that suggestion. And I think we saw 
yesterday that they are eager to do that as they appoint their 
own innovation officers, their own legal teams who are thinking 
through this set of issues.
    We are talking about innovation, we are talking about small 
and large, and then we are also talking about pursuing 
innovation, yet, obviously, complying with all the laws that we 
have in the country. And these are doable things, right?
    Nobody seeks to create a model with bias in it. In fact, 
they have a legal obligation not to do that. So, there is no 
group of people, hoodies or no hoodies, who are out there 
seeking to generate a credit model that has bias in it.
    But, Dr. Merrill, you make good points about this.
    This is a problem in government, too. Let's talk about the 
Consumer Financial Protection Bureau (CFPB), just a few years 
ago in their settlements with Honda and Toyota, where they used 
big data to estimate somebody who might have been a source of 
bias in auto finance--using big data, not real customer data, 
and just assumed that if your name is ``Hill'' and you are from 
``72207'', you might have a chance of getting a reimbursement 
from one of these settlements, based on bias. It was 
fallacious, and I think this committee was stunned by that a 
few years ago.
    We know in government and the private sector, this is a 
real challenge.
    Dr. Merrill, you talked about the model development and 
updating the regulatory guidance, and you have shared your 
work. How do we invite those regulators to put out for a 
rulemaking on updating that 2011 guidance? How would you 
propose that we encourage that?
    Mr. Merrill. My team who built the updates and I have spent 
a long time meeting with essentially all of the regulators, 
prudential and non-prudential. And one of the things that we 
have found is, I think if you wandered around Silicon Valley 
and asked, people would say, oh, regulators are against 
innovation. And that has not been my experience at all. The 
question has been, how do they do the changes in a way which 
serves them well, their regulated institutions well, and 
Congress well?
    For me, I think the single most important element moving 
forward is regulatory certainty. And I think it is impertinent 
of me to suggest what Congress should do, although I am 
``72032'', so--
    Mr. Hill. There we go.
    Mr. Merrill. --slightly different.
    But even a small push to the regulators to say, we believe 
ML is coming and we believe your methods of ensuring fairness, 
of validating for FCRA and ECOA, and of making the promise of 
ML win, would be a substantial step forward.
    Mr. Hill. That is why I support the sandbox idea. I think 
you all do, because you learn by doing. Of course, we are 
alleging the machines are learning by doing too. So, it is a 
way to backtest the reality, and I think sandboxes are useful. 
We would like to see sandbox uniformity among the agencies and 
a process that is open and not just--although I like the 
regulatory competition. In our society, it seems to be good. 
But we need to press on with that.
    Also, I was comforted in a recent meeting with one of the 
Federal Reserve district banks that, don't forget, we have a 
lot of depository institutions that are buying credit that is 
originated in this way on their books. This is a good market 
test right now because we are looking at that data, we are 
doing our HMDA, our fair lending analysis against those 
purchase loans. And that is a way to get grassroots data as 
well.
    Mr. McWaters, with 18 months of research focusing around 
the world on this, could you expand a little bit on why you are 
in the cup-half-full camp as well on long-term employment 
trends that we need? Give us some examples of these jobs that 
are being created that may see roles changed.
    Mr. McWaters. I can't speak to the specific methodology of 
the report that you mentioned. However, I think that it is 
actually quite useful when we think about this to reflect on 
history. The ATM was first introduced into the financial sector 
in the late 1960s, and there were some who predicted that we 
would no longer have branches, we would no longer have people 
in those branches.
    What has happened instead is that the role that the 
individuals in those branches perform is markedly different 
than it was 20 or 30 years ago. It is no longer focused on 
basic transaction processing, but instead focused on advice and 
new sales origination.
    And I think there are many examples of where we will see 
the fundamental activities of a job, the things people spend 
their time on, change and shift. That will likely require re-
skilling and retraining. But we won't see the job in and of 
itself removed.
    Mr. Hill. Thank you.
    Mr. Chairman, thank you, and I yield back my time.
    Chairman Foster. Thank you.
    The gentleman from Illinois, Mr. Casten, is recognized for 
5 minutes.
    Mr. Casten. Thank you very much. And thank you so much to 
Representative Foster for your leadership on this issue. It is 
truly a privilege to serve on this committee. And thank you to 
all the members.
    I have to start with a story. I ran an energy company for a 
number of years, and we had about 60 customers. Our biggest 
source of budget variance every year was our inability to 
predict how much energy our customers were going to use.
    And, nerd that I am, I built a big genetic algorithm. We 
tweaked it. And ultimately, we were able to massively cut the 
revenue variance in ways that scared the pants off my 
customers, because they had no idea how we did this, and 
neither did I.
    I mention that because what we found--I designed this to 
solve for a question of how to get better accuracy in our 
revenue forecast, and it did that beautifully.
    The more granular I got, the more inaccurate it was. If I 
asked what a specific customer was going to be, it was a little 
goofier. If I asked what a specific customer's consumption of 
chilled water would be, it would be goofier still. And if I 
said what a specific customer's chilled water consumption was 
in May, it was off the charts.
    Now, we knew well enough not to use it to ask those latter 
questions. But, Dr. Turner-Lee, a lot of what you described is 
that we have these tools that we built to ask one set of 
questions, which are really good. How do we improve our credit 
evaluation? How do we improve our underwriting? But then we 
have unintended consequences when we dig down to say, what does 
this say about a specific individual? And I don't know how to 
decouple that in the underwriting realm.
    But I guess my question for you is, do you see ways, 
computationally or regulatorily, to say, if we design this to 
do one set of things, let's use it for that thing and be aware 
to where the blind spots are, just because of the nature of the 
math? These could be totally unintended. But how do we 
constrain it in that way? Your thoughts?
    Ms. Turner-Lee. Yes, I think that is an interesting 
question. It is a regulatory question that we are looking at in 
the privacy discussion right now, the extent to which consumers 
give so much data that there are no start and stop points with 
the accumulation of that.
    I would echo what the panelists have said about the opaque 
nature of algorithms. And to your point, Congressman, what we 
are seeing is once it goes deeper into the ocean, the 
inferences that come out of that data are what is troubling, 
and are what lead to those unintended consequences.
    So, we have to find ways to cure that. Do we allow 
consumers to tell us when that start/stop is with regard to use 
of their data? And the comment earlier about regulatory 
sandboxes, do we permit for anti-bias experimentation the use 
of demographic information when we know it is actually going to 
help us curb bias in ways that would be detrimental to certain 
populations?
    I think, as you are talking about, the more granular we 
get, the less accurate we are, because there are certain data 
blind spots, as you suggested, that we are just not getting at. 
And the way that the technology works with machine-learning 
algorithms is, it assumes because a person or subject or object 
has engaged in that way, that that is who they are.
    And that is where we find ourselves replicating and 
amplifying the stereotypes externally, because it is not the 
algorithm that is saying to itself, ``I am going to be biased 
today.'' It is who we are as a society and who is actually 
inputting that data to create what has been considered the 
``garbage-out'' variables.
    Mr. Casten. The second question is for Dr. Merrill or 
McWaters, you guys can arm wrestle over who gets to answer this 
one.
    None of you mentioned algorithmic trading. Some friends and 
colleagues who are in that space have described it to me as 
being: number one, awesome; and number two, completely 
unhedgeable, because it is totally blind to black swan events, 
because of the conversations that you mentioned. It overweights 
recent data, it overweights success, and, therefore, is both 
blind to black swans and, as my friend who shall remain 
nameless said, potentially creates some really bizarre social 
outcomes. Because if you are managing a socially responsible 
fund, and all of a sudden your algorithm is trading on a bet 
that we are going to invade Crimea next week, you know, weird 
things happen.
    How do you think we should be regulating algorithmic 
trading in terms of the underlying risk, how much can we let it 
penetrate the market, and what do you do with an algorithm that 
is trading in a way that people may not actually understand 
what the bet is?
    Mr. McWaters. I think that this is an excellent point and 
one that requires further investigation. We have seen in this 
space a tendency for machine-to-machine interactions to lead to 
feedback loops that have damaging impacts.
    We have also seen that the innate foreignness that you have 
referred to in terms of the way that an AI-enabled model thinks 
can create confusion between fast-moving AI and slow-moving 
individuals, where people effectively freeze in response to an 
unexpected event. And that freezing is then interpreted as a 
further negative signal by the AI, driving things to an even 
more difficult situation.
    Core to addressing this, in my mind, is scenario-based 
modeling and the types of stress-testing approaches that we 
have used in the past.
    Mr. Casten. I am out of time, so I thank you.
    And I yield back.
    Chairman Foster. The gentleman from Ohio, Mr. Gonzalez, is 
recognized for 5 minutes.
    Mr. Gonzalez of Ohio. Thank you, Mr. Chairman.
    And thank you, everybody, for being here.
    I am really excited about the direction of this task force 
and the leadership on both sides of the aisle from Dr. Foster 
and my colleague French Hill, and just really excited. And 
thank you for convening this.
    One of my big priorities here on the committee has always 
been finding ways to expand affordable credit to low- and 
moderate-income borrowers. I think that has been one of the 
more difficult challenges that we have faced as a society, 
certainly in the financial services sector, for a very long 
time.
    And part of why I am excited about machine learning is 
what, Dr. Merrill, you suggested, which is that we can do this. 
This is something that is attainable. But there are certainly 
questions.
    In your testimony, you talked about how there are 
``explainability models'' that aren't really doing a great job, 
but at ZestFinance you have developed one or you have developed 
methods that render ML models truly transparent, to directly 
quote you.
    My question is more on the technical side. Technically 
speaking, how difficult is it to create a proper explainability 
model, knowing that, from my time in tech--I used to work in 
tech, not at your level--an A-plus engineer is kind of worth 
about 10 midlevel engineers, if you will.
    Talk to me about the technical side of this, if you would?
    Mr. Merrill. Thank you for that question.
    I think the way to think about it is to just kind of draw 
some broad boundaries about the question at first. One of the 
techniques that differs in machine learning from traditional 
underwriting is you use a bunch more data, and data is 
sometimes called signals.
    And when you are going to do explainability, conceptually, 
the hard part isn't actually comparing the inputs and the 
outputs. The hard part is understanding what things inside the 
models moved together to produce that output.
    That essentially means you have to compare all pairs of 
signals. If you have 100 signals in a model--which, by the way, 
would be a very small model--you would have to compare all 100 
to all other 100, which sounds easy, except that turns out to 
be more computations than there are atoms in the universe, 
which is a bad outcome. Well, it is a bad outcome, if you want 
an answer.
    The tricky part is you have to figure out how do you 
optimize that in a way which guarantees correctness, but 
doesn't require you to be computing until the sun burns out. 
And what the mathematicians on our team have figured out a way 
to do is to make those optimizations, but to do it in a way 
that they can still prove the answer and we can demonstrably 
answer the question of are we, in fact, accidentally 
discriminating against African Americans or women.
    And that is our view, is that the two things that an 
explainability model must do: one, it has to successfully 
optimize across the space; and two, it has to be directly 
inquirable as to what do you do with respect to whatever 
classes are relevant.
    Mr. Gonzalez of Ohio. Thank you.
    And then one thing we have talked about a lot is the data 
itself. But we haven't covered as much about--Dr. Buchanan, you 
mentioned it--privacy and who ultimately owns the data. I think 
that is an outstanding question for sure.
    And so I guess my question is for Dr. Buchanan and anybody 
else who wants to take a stab at this, how should we think 
about balancing the innovation that we all agree can have a 
positive impact on society if we are good about it, with 
protecting consumers and empowering consumers with their 
individual data?
    Ms. Buchanan. Thank you, Congressman.
    I absolutely agree with this. And I have been very 
encouraged by what I have seen in the European Union regarding 
consumer protection on data and the right to own the data and 
what happens with your data.
    I think one thing I would like to stress to you throughout 
today is, I keep hearing the term ``big data'', but I think, 
moving forward, what we also need to distinguish when we are 
getting down to that granular level is that big data is not the 
same as strong, robust data.
    Mr. Gonzalez of Ohio. Right.
    Ms. Buchanan. When we are thinking about privacy, we need 
to think about using strong, robust data.
    And I think I would also draw your attention to my written 
report where I look at China. Look at what they have been doing 
with their Sesame Credit model with Ant Financial, which is not 
the same as the government social credit scoring model, where 
basically every data point ever collected about you goes into a 
model to measure what is called ``trustworthiness.'' Not 
creditworthiness, trustworthiness.
    And my thoughts on this is, at the end of the day, if I am 
going to look at getting a loan for a house, the data I really 
want to use and protect is my loan repayment history, not my 
subway fare usage, for example.
    Mr. Gonzalez of Ohio. Right. Thank you.
    Ms. Buchanan. And context is very important, too.
    Mr. Gonzalez of Ohio. Yes, ma'am. Thank you. We will follow 
up.
    And I yield back.
    Chairman Foster. The gentlewoman from North Carolina, Ms. 
Adams, is recognized for 5 minutes.
    Ms. Adams. Thank you, Mr. Chairman.
    First of all, let me, before I begin my questions, I want 
to thank you for the opportunity to serve on this task force. 
And I am looking forward to it, along with you, and my friend, 
Congressman Hill.
    To the witnesses today, thank you so much for your 
testimony.
    As technology becomes more and more commonplace, it is 
critical that we proactively address issues that could 
positively and negatively impact our constituents and our 
financial institutions.
    Algorithms have become a part of everyday life, even though 
most Americans have limited awareness or understanding of these 
systems and their impact. Increasingly, public and private 
enterprises have turned to artificial intelligence software and 
machine-learning programs to help increase the effectiveness of 
the services rendered.
    Let me begin by addressing this question to Dr. Turner-Lee. 
There have been concerns about bias in AI systems, such as the 
potential of historical biases in datasets to be perpetuated or 
amplified in AI systems. How do firms ensure that AI systems 
are not having a disparate impact on vulnerable communities? 
And what safeguards should regulators and Congress put in place 
to protect consumers?
    Ms. Turner-Lee. Thank you, Congresswoman, and thank you for 
that question.
    I am going to just give three points that I think need to 
be injected into this debate.
    One is diversity in the workforce. The developers who sit 
at the table in the design of algorithms are not representative 
of the colorful spectrum of people who actually are using these 
algorithms. And, as a result, I think that we miss 
opportunities to have a seat at the table to mitigate issues 
related to gender or race or even background. I am a 
sociologist sitting among computer scientists. We need more 
perspectives with regards to that.
    And I think to push for inclusion, we also need diversity 
in design. We wrote a paper at Brookings that is really about 
sitting at the table and thinking through what may become the 
intended and unintended consequences of these models. How are 
they replicating stereotypes that we see? In what ways should 
companies be trying to put in best practices that avert those 
types of discriminatory actions?
    People of color, in particular, have not come this far to 
have technology become one of the major elements of further 
discrimination and amplified bias. And so, we have to be 
proactive in increasing the number of data scientists who are 
engaged in this, who come from diverse backgrounds, and also 
creating, I think, a standard, particularly in the sensitive 
use cases like financial services, employment, and housing, 
where people of color have already been historically 
disadvantaged, that we have to ensure that these sensitive use 
cases are not open for business with regards to doing further 
damage.
    Ms. Adams. Great. Thank you.
    Dr. Buchanan, within the context of financial services, 
have you seen the potential for bias in the use of AI? And how 
are various countries handling this issue? What should 
policymakers do to ensure the use of AI doesn't discriminate 
against vulnerable communities?
    Ms. Buchanan. Some of the more notable examples that I 
highlight in my report, Congresswoman, relate to how algorithms 
are used in the peer-to-peer lending industry. And so, just to 
follow on from Dr. Turner-Lee's comments, I can refer you to a 
paper where I found that peer-to-peer listings where African 
Americans provide their pictures on the lending site are 
roughly 3 percent less likely to be funded and receive a loan 
and are more likely to pay higher basis points than white 
people with similar credit profiles. The examples I detailed in 
my reports are particularly pertinent in the debt 
consolidation.
    Ms. Adams. Okay. Let me ask a yes-or-no question: Would it 
be useful for Congress to fund algorithmic bias research 
through NSF, NIST, and other Federal agencies, to develop 
tools, methods, and programs to resolve bias in artificial 
intelligence systems? If I can get a yes or no?
    Ms. Buchanan. Absolutely, yes.
    Ms. Adams. Okay. Dr. Turner-Lee?
    Ms. Turner-Lee. Yes.
    Ms. Adams. Dr. Merrill?
    Mr. Merrill. Yes.
    Ms. Adams. Mr. McWaters?
    Mr. McWaters. Yes.
    Ms. Adams. Okay, very good. Thank you very much.
    Dr. Merrill--and I know we don't have a lot of time--what 
steps should companies and policymakers take to address this 
concern? Can you give me one?
    Mr. Merrill. I think the most important thing that 
regulators and policymakers should do is provide clarity. Even 
clarity that is not perfect is better than uncertainty to get 
companies to innovate in a good way.
    Ms. Adams. Great.
    Thank you, Mr. Chairman. I yield back.
    Chairman Foster. Thank you.
    The gentleman from North Carolina, Mr. Budd, is recognized 
for 5 minutes.
    Mr. Budd. Thank you, Chairman Foster. I want to commend you 
and my friend Ranking Member Hill for all your work on this 
task force.
    I am excited that you all are here today.
    And I want to start my time by highlighting the potential 
impact that machine learning and AI can have in our insurance 
market for institutions and their customers. But before I do 
so, I want to ask permission, Mr. Chairman, to enter into the 
record this report from the GAO. It is entitled, ``Insurance 
Markets: Benefits and Challenges Presented by Innovative Uses 
of Technology.''
    Chairman Foster. Without objection, it is so ordered.
    Mr. Budd. Thank you, Mr. Chairman.
    This report highlights how AI and machine learning benefit 
insurance markets and the consumer. I am excited to explore how 
this technology can improve underwriting accuracy, facilitate 
stronger communication with customers, make the claims 
processes easier to navigate for the consumer, and combat 
insurance fraud, among many other things.
    Let me just highlight one specific provision from the GAO 
report, that is found on page 11. The report highlights 
telematics, which is the combination of telecommunications and 
information processing to send, receive, and store information 
related to specific items such as automobiles and water 
heaters. And I happen to have one of those water heaters, and 
it never knows when the in-laws are coming and when all the 
kids are home from college.
    Telematics allows sensors in an automobile to provide data 
on a driver's behavior such as speed, hard braking, and turning 
radius. Now, according to the GAO report, insurers can then use 
that information to determine the driver's risk profile and 
help determine the premium rate for that driver, if a driver so 
chooses.
    So, I encourage my colleagues to read this report that was 
requested by Ranking Member McHenry as we move forward with 
this task force with any potential policy proposals. Thank you.
    I am sure we all agree that the U.S. must stay at the 
forefront of this new technology in the financial sector, like 
artificial intelligence and machine learning.
    And here is the question. It is for Mr. McWaters: What 
challenges are companies facing that inhibit them from 
achieving the full potential of these emerging technologies? 
How are overly burdensome regulations stunting growth in this 
area? And how can our committee ensure that proper controls are 
in place to protect customers while also fostering growth in 
AI?
    Mr. McWaters. Thank you very much.
    I think that one of the most significant instances of where 
we see challenges to responding to this on the part of 
particularly incumbent financial institutions are the legacy IT 
systems that are in place.
    Typically, data is heavily siloed, making it difficult for 
that data to be ingested and used by conventional machine-
learning methods, and the systems themselves, while extremely 
robust and resilient, are not as adaptable as modern and 
particularly cloud-based computing methodologies.
    Interestingly, one of the things that we have seen in this 
space--and this pertains to some degree to Chairman Foster's 
question about consolidation--is that there is an opportunity 
for third-party service providers to play a helpful role in 
enabling financial institutions to leapfrog forward, in terms 
of their capabilities.
    By plugging into specialized fintech or regtech firms, into 
large tech firms which might offer, for example, machine vision 
as a service, you might as an insurance entity be able to use 
that machine vision to accelerate the processing of minor 
automotive claims, for example.
    I think that, in terms of the discussions that I have 
internationally, one of the perceptions of the United States in 
this space is that the regulatory environment is extremely 
complex to navigate and that the large number of regulatory 
entities creates challenges to deploying new innovations 
effectively.
    I don't have a specific remedy for that, but it certainly 
is one of the contributors to the challenge of deploying these 
technologies here in the United States.
    Mr. Budd. I appreciate that, Mr. McWaters. And continuing 
on with you, besides lower cost of financial products and 
services, what are some other ways in which a consumer stands 
to benefit from adoption of these technologies in the financial 
services?
    Mr. McWaters. I think one of the particular items here is 
the opportunity to provide valuable advice and intervention for 
clients. So, if you pursue the example of insurers that you 
gave, telematics has an opportunity to, on one hand, support 
more accurate and more personalized underwriting, but it also 
increasingly has the potential to give drivers valuable 
feedback on how they might be safer drivers.
    The water heater that you mentioned might be able to alert 
you if there was a leak, allowing you to minimize the damage to 
your home in a way that is beneficial both to you and to the 
insurer who has provided that cover.
    Mr. Budd. It sounds like a lot of opportunities.
    With that, I yield back. Thank you.
    Chairman Foster. Thank you.
    And after consultation with the ranking member, I would 
like to inform Members that we are going to have time for a 
second round of questions, subject to the fact that we have to 
be done here by 11:30. So, we should at least have a partial 
second round here.
    I now recognize the gentlewoman from Texas, Ms. Garcia, for 
5 minutes.
    Ms. Garcia of Texas. Thank you, Mr. Chairman, and thank you 
for having this hearing. And I thank Chairwoman Waters for 
really focusing on this issue, because it is so important as we 
move forward.
    However, I think it is one that is kind of confused, and I 
wanted to just start with a question. I was trying to figure 
out which professor to ask, so I am going to go ahead and go 
with a woman. I, too, have some biases.
    Dr. Buchanan, for those who are watching who are not in the 
financial industry, who don't know what artificial intelligence 
means, they hear the word, ``intelligence'', and they think it 
is some really super big-brother secret stuff. Can you in just 
plain English, in 25 words or less, tell the average viewer 
what the heck we are talking about?
    Ms. Buchanan. First of all, there is no generally agreed 
upon definition of ``artificial intelligence.''
    Ms. Garcia of Texas. You are using up your 25 words now. 
You are talking straight to the average consumer in the United 
States.
    Ms. Buchanan. Okay. I would say it is a group of 
technologies and processes that can look at determining general 
pattern recognition, universal approximation of relationships, 
and trying to detect patterns from noisy data or sensory 
perception.
    Ms. Garcia of Texas. I think that probably confused them 
more.
    Ms. Buchanan. Sorry.
    Ms. Garcia of Texas. With all due respect, but I think that 
is one of the challenges that we have. I wanted to do that, not 
to make light, but just to accentuate the problem that we are 
facing, because I think there is an idea that now all these 
robots are going to take over all the jobs and everybody is 
going to get into our information, this whole balance that one 
of my colleagues mentioned between privacy and the markets. So, 
I think it is important.
    Ms. Turner-Lee, one of the things that would help us better 
understand it, I think, are some of the things you pointed out, 
in terms of diversity of the people at the table who are 
developing the software, the people who are the workforce 
involved.
    If you could name the single one thing that Congress could 
do, I mean, we can't change attitudes. We probably can't change 
some of the criteria that the folks who are putting this 
together are looking at. What would you suggest that one thing 
be?
    Ms. Turner-Lee. Yes. That is such an interesting question, 
because I think the tech diversity issue has been one that 
Congress, as well as civil society actors and others, have 
really grappled with. And as we see technology evolve in the 
way that it is to a point where it is confusing, I would 
suggest that we have a lot more to do as these become much more 
ubiquitous and widespread.
    On your question, I think what Congress can do first to 
quell algorithmic bias is to create guardrails. I think it has 
been mentioned that we need to ensure the tech companies know 
that they have to be in compliance with antidiscrimination 
laws. I think we start there. We create guardrails for best 
practices in design and development.
    With regards to creating more diversity at the table, these 
are companies that are not necessarily regulated or in any way 
required to report diversity, in terms of who they serve and 
who is sitting there. But I think we should reward best 
practices where we are seeing demonstrations of companies 
wanting to bring more actors to the table.
    What does that mean? Years ago, when we had the ENERGY STAR 
standard imposed on appliances, most of us who go into a big 
box store know this appliance is going to save us money and it 
is going to be safe.
    I think we should push in the algorithmic economy a gold 
standard: What is the Energy Star rating for what consumers 
understand of how their data is being used? And how will 
companies pushing the bar, raising the expectation that they 
are going to be in compliance, not only with those 
nondiscrimination laws, but they are going to be good stewards 
of our information and they are going to have environments 
where diversity is encouraged?
    Ms. Garcia of Texas. Is there anything that we can do in 
terms of the criteria that they are using? Because I know one 
of the examples you gave on gender bias was just the word 
``woman'' being on their resume somewhere caused to trigger the 
gender bias.
    What can we do with regard to the criteria being used? For 
example, if you looked at my resume, I graduated from a 
Historically Black College, and I would hope that there is no 
assumption that I am African American, but a computer could do 
that, right?
    Ms. Turner-Lee. That is right.
    Ms. Garcia of Texas. But I also go to a women's college, 
so, obviously, that is going to peg me in that. But then they 
look at me, and I don't look like I am Latina.
    Ms. Turner-Lee. That is right.
    Ms. Garcia of Texas. I am going to have one confused 
computer.
    Ms. Turner-Lee. That is right. And you are going to have a 
double or triple jeopardy, right?
    Ms. Garcia of Texas. But is there any way that we can do 
anything about what gets in the computer?
    Ms. Turner-Lee. Yes, as a policymaker myself at Brookings, 
it is so challenging to figure out how do we get companies to 
sort of adhere to a standard without overregulating them? And 
that is why I think those guardrails are particularly 
important.
    But I also think it is important for us to continue this 
discussion on what does disparate impact mean when collective 
groups of people are denied loans or denied credit or denied 
some form of equitable opportunity in this country simply 
because the computer was wrong. Who is liable for that? Is it 
the developer?
    I actually agree with what was said earlier. I don't think 
developers necessarily walk around in a hoodie saying, ``Today, 
I am going to discriminate against people.'' I think it is the 
nature of what is in the black box that is not understood, 
which is why explainability models matter.
    People need to understand what is going into this ocean. 
And for the layperson, I will give you this example that I use. 
It is like swimming in the ocean. At the top, you can see my 
legs and my hands, but when you go down, you begin to not see 
my body because the water becomes really cloudy.
    I am okay if I actually search for camping gear for my son 
on one site and it shows up on another site. I am not okay if I 
am profiled because I am an African-American woman or a woman 
who went to a Historically Black College, et cetera. Those are 
things that I can't see how you even got there to understand 
that from just my hand sticking out.
    And so, we have to figure out what are those guardrails 
that will protect people, where are there pressure points to 
institute some other consumer protection, what is the role of 
privacy in terms of the data that is collected on people?
    And I would suggest to you, where in the process can I 
recurate my identity and let them know that, ``Hey, I am not 
this person that you keep thinking I am just because I buy 
camping gear. It is not me going out; it is my son.''
    Ms. Garcia of Texas. It is a good point. Thank you.
    Chairman Foster. Thank you.
    Ms. Garcia of Texas. I yield back. Thank you, Mr. Chairman.
    Chairman Foster. This is a wonderful discussion that could 
go on forever.
    The gentleman from Virginia, Mr. Riggleman, is recognized 
for 5 minutes.
    Mr. Riggleman. Thank you, Mr. Chairman, and thank you to 
Ranking Member Hill, and thank you to all of the witnesses for 
being here.
    I would like to start by saying I am proud to be a member 
of the inaugural Artificial Intelligence Task Force. And I was 
going to send my avatar today, but it kept going in circles and 
bumping into walls, so I said, I am going to come here myself. 
That was a bad, bad joke.
    But, anyway, my background experience with data analytics 
has taught me a lot, especially about the evolution I 
personally witnessed since 2002. And to get to my questions, I 
just want to talk really quickly about what I have done. My 
experience might be a little bit different than everybody up 
here.
    I have been trying to aggregate big data and analyze big 
data for predictive analysis to go after actually network 
centers of gravity and critical touchpoints for a long time in 
the nonkinetic space on the military side.
    And back in 2002, I want to tell you guys, that the big 
thing about the military--we have this incredible saying, that 
we try to solve today's problems with yesterday's technology 
tomorrow.
    I think what I saw in 2002, there was never a statement of 
AI or machine learning. We were using these just really kludgy 
relational databases, trying to build arbitrary translators to 
try to make sure the nodes and attributes actually made sense 
for unproductized data, productized data, but mostly data that 
just didn't make a lot of sense to us in 2002.
    What we have seen in the last 5 years, and I know this is 
crazy because sometimes the DOD is a little bit behind, but it 
is our work with places like Johns Hopkins University's 
Federally Funded Research and Development Centers (FFRDCs), 
working with the physics labs. And now you see a lot of not 
only private-public partnerships, but you see a lot of 
commercial and government partnerships in big data.
    And what we have seen going forward is, that 5 years ago we 
might have been using relational databases, but now we are 
using graph databases and dynamic translators we could have 
never foreseen in the future. We had about 40 people working 
with us trying to find every touchpoint and every critical node 
in a network. So, I went from dropping bombs to actually 
dropping nonkinetic bombs, right, in specific types of 
networks, is pretty much what we did.
    And it is just amazing to me, listening to all of you, that 
my background is so different, just based on trying to work 
with data, and the fact that machine learning and artificial 
intelligence, even up until 2010, 2011, in the military space, 
and big data with my companies, we really didn't talk about it 
much. We just really didn't. But now we can.
    And what we see now is that now we are getting 
unproductized data. We are getting disparate data, multiple 
datasets. I am getting natural language processing. We are 
getting tons of unstructured data. We are able to go into 
dynamic translators we can put into graph databases, and now we 
are actually coding to what people are thinking when they are 
looking at a specific problem set. We are coding to an 
analyst's brain serially in parallel. Now, we have machine-
learning templates.
    And here is what happened after all that incredible stuff: 
It failed miserably the first time, because we were missing so 
much data.
    The thing that I am going to ask, because I have my own 
reasons about this, and I will ask Mr. McWaters first, when you 
look at AI and ML, when you are looking at ML templates, 
machine-learning templates, when you are looking at what 
artificial intelligence is, the difference between templating 
and the difference between rules, where do you think the split 
is? And I want to ask some of you, where do you think the split 
is because definitions of machine learning and AI?
    I know I have my own, but I would love to hear from you, 
because sometimes I even get sort of wrapped around the axle in 
trying to figure out where that split is and where we can 
actually look at some of the safeguards to make sure that we 
make the right jump from ML to AI.
    Mr. McWaters. There is an old joke that artificial 
intelligence is whatever a computer can't do yet.
    Popularly, our definitions of this have tended to move over 
time. Twenty years ago, you might have said that a computer 
would be intelligent if it could beat a grandmaster at chess. 
Today, we sort of think of that as being a relatively trivial 
case of intelligence. We think of it as being programmatic.
    So, I think our definition of artificial intelligence tends 
to move over time. And, as Dr. Buchanan said, I don't think 
there is a clear articulation of exactly which techniques--ML, 
deep learning, and others--are specifically rested under the 
umbrella of that definition.
    Mr. Riggleman. Dr. Merrill?
    Mr. Merrill. I think we can spend a lot of time trying to 
get our heads around the different definitions. When I started 
in the field, which is a long time ago now, AI was generally 
thought to be machines that tried to actually reason, that 
tried to start with an initial point and take steps to get to 
an end point, whereas ML was viewed more as just rote math, 
just like throw a computation at the problem.
    Mr. Riggleman. Right.
    Mr. Merrill. You can still sort of throw that distinction 
out, but it just turns out to be a little bit unhelpful at the 
end, because AI failed when I started and it is roughly still 
failing, because it is just a really hard problem. People turn 
out to be really, really complicated beings.
    And stuff which we said could never get done until AI 
worked is now relatively trivial in ML. To wit, your car's 
brakes are better than you are. And that is a case of ML that 
we said could never be done. You could never compute friction, 
but it turns out you can.
    Ultimately, I think the most important class is maybe not 
whether it is AI or ML, but rather what are the characteristics 
of the problem you are trying to solve? AI-based techniques are 
trivial to explain. ML techniques are quite a bit harder to 
explain, but quite a bit more powerful. And so I guess I would 
encourage us to think less about the technique and more about 
the category of problem.
    Mr. Riggleman. Thank you.
    And that is why I am so excited about this. Thank you, Mr. 
Chairman. Because I think we have a chance to really solve some 
problems here, and I am happy to be here. Thank you, sir.
    Chairman Foster. Thank you.
    The gentleman from Georgia, Mr. Loudermilk, is recognized 
for 5 minutes.
    Mr. Loudermilk. Thank you, Mr. Chairman.
    I appreciate the panel being here. It is a very intriguing 
discussion we are having here today, especially as I spent 30 
years in the information technology industry, as my good 
colleague, Mr. Riggleman, also spent time in the intelligence 
community in the Air Force in the earlier days where we were 
using analytics of massive amounts of data. And what is 
happening in that arena today is light years beyond anything 
that we were able to do with rooms full of main processing 
systems, mainframes back in the time.
    And I am really interested in this field today, in what we 
can do with our artificial intelligence. I think it is also as 
important to understand our limitations of what we can't do and 
draw our boundaries around that, but yet on the periphery of 
that boundary having the sandboxes to where we can test and we 
can implement what we may be able to do in the future once we 
stabilize that.
    One of the things I am interested in is what can we do 
today with artificial intelligence and fraud detection and 
prevention, because that is something that is really important 
in the industry, especially as we move more in the fintech 
arena.
    My line goes back to the chip card industry. Since I have 
been in Congress, when I first started here, my debit card and 
my credit card had a chip, but I could only use it when I 
traveled overseas.
    Once we implemented that ability here, the fraud went down 
by 76 percent. But criminals being criminals, all they do is 
shift their focus, and that focus has gone over into the 
digital payments arena, which is where we have a lot of 
challenges today.
    And, Dr. Buchanan, I appreciate your discussion that you 
brought up in your testimony about how one of the payment card 
networks is using AI to help financial institutions reduce 
their fraud by $25 billion annually. Can you tell us more 
detail about how payment processors-- financial institutions, 
insurance, retail, and others are using AI to combat the 
digital payment fraud?
    Ms. Buchanan. When we are thinking about AI's automating 
simple and complex decisions--actually, that is my 10-word 
definition, so I think I have redeemed myself, Congressman.
    One area that I can address to you is that 50 percent of 
phishing detections are now finance-related. And so what I 
detail in my report are some very encouraging examples around 
the world where financial services companies have tried to 
reduce phishing attacks.
    There is a really good example in my report, IBK, a 
phishing voice detection app, and it is really a coordinated 
effort between regulators in South Korea and the financial 
services industry.
    Basically what this app looks at is--and phishing in South 
Korea accounts for millions of dollars a year--a phone call is 
made, and it looks at picking particular keywords in the phone 
call. And if it meets a particular threshold, then an alert 
signal is sent that this is a potential voice phishing scam, 
and a significant financial transaction is halted.
    In Estonia, Monese is using artificial intelligence in this 
arena as well, particularly when they are trying to on-board 
customers in the first place. So, they are looking at matching 
documents with video selfies in order to detect fraudulent IDs 
and fight identity theft.
    Mr. Loudermilk. I traveled to Estonia last year, and what 
they are doing in the fintech industry is really a model for a 
lot of other nations. It is surprising, especially being an 
Eastern Bloc country, the suppression that they had during 
communism, to be able to come out to where they are now.
    Regarding the things you just explained to us, payments.com 
showed that less than half of financial institutions use AI for 
fraud prevention. Why are we not seeing more use in the 
industry for fraud prevention?
    Ms. Buchanan. That is an interesting question, Congressman. 
I think really it is because detecting fraud in the first 
place, we think about fraud as really being a latent variable. 
I mean, it is not necessarily directly observable, and so it is 
more challenging to machine-learning algorithms.
    Actually, in some sense, you have a little bit of a self-
defeating goal here. You could have the case of falsely 
declining transactions as fraudulent, okay. That actually costs 
the industry a lot in lost customer loyalty each year.
    And apart from this erosion of customer loyalty and loss of 
retail losses, the machine-learning algorithms to detect fraud, 
as I said, they are more latent, in the sense that it is easier 
to track someone's shopping history directly. You see what they 
purchase. You see what they buy. But fraud is just another 
layer. It is not as directly observable. And I think that 
presents a complexity to the process.
    Mr. Loudermilk. Thank you.
    Chairman Foster. Given the time constraints on our 
occupancy of this hearing room, it looks like we will have time 
for only 5 minutes of questioning by the ranking member and the 
Chair. So, I would now like to recognize the distinguished 
ranking member for 5 additional minutes of questions.
    Mr. Hill. I thank the chairman.
    I thank, again, the panel for being here today. I 
appreciate your contributions to this important beginning of 
the task force work for this Congress.
    Mr. McWaters, I wanted to start with you and just talk 
about some of the ways today that you are seeing AI being used 
in the financial services industry.
    So, if you would talk about two or three of the biggest 
ways you are seeing artificial intelligence being used by the 
financial industry in customer acquisition, extension of 
credit, regulatory compliance costs? Name two or three or four 
specific elements in each of the main areas, if you would.
    Mr. McWaters. I think we are seeing four key ways in which 
this is being deployed in financial services.
    The first is driving increased efficiency, being able to do 
the same thing faster and with less manual input. And that can 
be a benefit both to the organization, obviously, in their 
bottom line, but also to the consumer, who is able to get an 
answer to their question or to their request more quickly.
    Second, we are seeing an improvement in outcomes. Dr. 
Merrill made reference to this in terms of being able to 
originate more loans, accept more applications without a 
significant increase in defaults.
    Third, we are seeing entities build out entirely new 
businesses. By virtue of some data flow that exists, is 
propagating through already, you may be able to create new 
value propositions. So, a payment network might be able to 
create a business of macroeconomic forecasting based on the 
data that flows through their network and monetize that 
separately.
    And then finally, advice. Americans struggle to access the 
financial advice that they need to make good financial choices 
in the moment to plan for retirement. That advice traditionally 
has needed to be delivered by expert individuals and can be 
very expensive.
    We are at the very beginning, I believe, of the opportunity 
to provide high-quality advice to individuals in real time that 
will help to address that issue. It is nascent today, but the 
opportunity is quite significant.
    Mr. Hill. On that point, I believe in making sure that we 
have an economy that offers choices to consumers from the whole 
spectrum of the most machine-led robo-adviser to the most 
sophisticated one-on-one consultation. I don't think that 
government policy should bias towards that, and we have had 
some debates over the last 4 years where I think government 
policy actually directed people away from advice to machine-
driven robo-advisers.
    If I go through a sharp downturn in my portfolio and it has 
been dependent on a robo-adviser, who am I holding responsible 
for that? Who can I go talk to about that?
    Mr. McWaters. I think that is an open question.
    Mr. Hill. I don't like open questions. That is why we are 
here today. We need to make sure that those consumers know the 
risks of that. And that may be the trend of the moment or the 
trend of the time or it may be, in the short run, more 
affordable, but those are the kinds of things I think we have 
to talk about here in this, in our work.
    Mr. McWaters. I would also note that I think that you will 
see in this space that even amongst some of the sort of highest 
echelons of private banking, what we now see is an appetite by 
those consumers to have a mix of both automated and in-person 
mediated items.
    The other thing that I would note in response to your 
earlier question about consolidation in the marketplace is that 
these technologies can also provide an interesting opportunity 
for small and midsized financial institutions to rapidly catch 
up to large entities.
    Mr. Hill. I do share your optimism there. All through the 
technology cycle, going back from a mainframe to a business 
size computer to the cloud, small broker-dealer competitors and 
small financial services competitors have had access to scaled-
up technology through a vendor platform that in some ways helps 
them do a better job of being in full compliance of risk.
    Data privacy, if each of you would just quickly answer, do 
you support the use of APIs when it comes to protecting 
customer service, customer data interfaces between aggregators 
or individual companies?
    Dr. Turner-Lee, do you want to start?
    Ms. Turner-Lee. Yes, I do.
    Mr. Hill. Dr. Buchanan?
    Ms. Buchanan. Yes, I do.
    Mr. Hill. Dr. Merrill?
    Mr. Merrill. Yes, I do.
    Mr. Hill. Mr. McWaters?
    Mr. McWaters. Yes, I do.
    Mr. Hill. Good. Thank you. I yield back.
    Chairman Foster. Thank you. And I guess as a follow-up on 
the API question, what do you think the state of the art is for 
authenticating yourself for access to those APIs?
    Because one of the scariest things that I see about 
artificial intelligence is just the very impressive high-
quality tools being used for phishing. Things, for example, 
where they will listen to your voicemail response, use that to 
synthesize your voice, and fake a phone call to one of your 
friends in your contact list saying, ``Hey, Joe, I just sent 
you an email with an attachment, can you have a look at the 
attachment and call me back?'' And everyone clicks on that 
attachment. And that is not even mentioning the video that is 
now available.
    I think one very valuable thing the government can do is to 
at least provide citizens who are interested in having a high-
quality way of digitally authenticating themself online very 
much in the way Estonia has been leading the way.
    And my closing question, I guess to each of you is, we have 
about 1 minute for each, if you look forward at the competitive 
environment, you see all of the giant banks trying to--they all 
have 10-year plans to turn themselves into tech firms. All of 
the tech firms are getting into banking as rapidly as you can 
imagine.
    And so looking forward a decade, what do you think about 
the competitive landscape? Will there be any difference between 
giant financial institutions and tech firms, as we know them 
now?
    Just march down the line.
    Dr. Turner?
    Ms. Turner-Lee. I think we are going to go in this era of 
converged services, and it is going to be very challenging for 
regulators and Congress to discern what guardrails apply to 
whom. And right now, we have strong sectoral policies that 
affect the financial services sector, and we have loosely 
regulated policies that may apply to tech companies.
    I think going forward we are going to have to figure out, 
particularly on behalf of consumers, where do those protections 
lie and where do we again place pressure for regulatory 
frameworks that allow for innovation while at the same time 
putting some stresses around the fact that we cannot have 
permissionless forgiveness in areas that have huge consequence 
for consumers.
    And so, I completely agree with you. I think at some point, 
the lines are going to be so blurred we are not even going to 
know.
    But keep in mind it has been consumers who are driving that 
demand for these services. So, I agree with you as well, we 
have to do--
    Chairman Foster. And in Congress it is, obviously, a big 
issue, because I think there are seven committees that claim 
they are doing some part of IT, information technology, which 
means, of course, no one is doing it.
    So, Dr. Buchanan, any thoughts on this?
    Ms. Buchanan. The landscape I see moving forward, Chairman 
Foster, is more mergers and partnerships between banks, 
financial institutions, and big tech companies.
    I do agree with Dr. Turner-Lee about drawing this line 
about how data is used. And I am very concerned, moving 
forward, that I want to make sure we don't give up privacy at 
the expense of convenience.
    Chairman Foster. Thank you.
    Dr. Merrill? And also, if you could comment on the role of 
the startup in this, where they may or may not have access to 
these giant datasets that seem to be essential for success in 
AI?
    Mr. Merrill. I guess I will be a little bit of an outlier 
here amongst my distinguished colleagues.
    I think there is essentially no chance that in a decade we 
will see mergers and material consolidation between technology 
companies and big banks, because the cultural differences will 
be so great that the mergers will blow up.
    I was responsible for a variety of our financial products 
when I was still at Google, all of which were carefully 
regulated really, because we were a bit weird about that. And 
it was clear that that was the wrong place to do those, those 
products, not because anyone had the wrong intent, but just 
because it just didn't fit.
    I think ultimately, startups are at material risk, and I 
think that is very dangerous for the U.S. economy. We are at 
risk because it is hard to get data. We are at risk because a 
brief sideswipe by a large company, let alone the government, 
will crush any of us.
    And I think over the last 20 years, for good or for ill, we 
have seen a lot of the development in this economy coming from 
startups. So, my biggest worry is that.
    Chairman Foster. Mr. McWaters?
    Mr. McWaters. I would argue that we need to think outside 
the bank, if you will, that we think about financial services 
in a heavily verticalized and siloed fashion. We need to think 
about it in a more modular way.
    And so when I look forward to the 10-year landscape, I 
would predict a world in which customer experiences for 
financial services increasingly trend towards the best of what 
big tech can offer, whether that is offered by a traditional 
financial entity or a technology entity, but that the products 
that the consumer accesses, the loans, the insurance, they need 
to fundamentally remain regulated.
    And the data that is used to inform the entire experience 
needs to become more secure, the customer needs to have more 
control, and we need to really enfranchise the customer within 
a regulated framework.
    Chairman Foster. Thank you.
    And I would like to thank all of our witnesses for their 
testimony today.
    The Chair notes that some Members may have additional 
questions for this panel, which they may wish to submit in 
writing. Without objection, the hearing record will remain open 
for 5 legislative days for Members to submit written questions 
to these witnesses and to place their responses in the record. 
Also, without objection, Members will have 5 legislative days 
to submit extraneous materials to the Chair for inclusion in 
the record.
    This hearing is hereby adjourned.
    [Whereupon, at 11:36 a.m., the hearing was adjourned.]

                            A P P E N D I X

                             June 26, 2019
[GRAPHICS NOT AVAILABLE IN TIFF FORMAT]