From the U.S. Government Accountability Office, www.gao.gov

Transcript for: Artificial Intelligence: Watchdog Report Deep Dig
Edition

Description: Artificial intelligence or AI will improve human life and
economic competitiveness in countless ways over the next few years, but
it also raises new questions about how this technology is used and
managed. We met with GAO's Chief Scientist Tim Persons, and special
guest, the former-Federal Chief Information Officer Suzette Kent to
discuss.

Released: July 2020

[ Music ]

[Jacques Arsenault:] Artificial intelligence or AI will improve human
life and economic competitiveness in countless ways over the next few
years. For example, it could help reduce the amount of time it takes to
bring lifesaving drugs to the market and improve the speed of banking,
but it also raises new questions about how this technology is used and
managed. Welcome to this Deep Dig edition of the Watchdog Report where
we focus on larger issues. You're going to hear from people behind the
work at GAO, their efforts, and their experiences. You know, things we
could dig deeper on. This episode is taking a look at artificial
intelligence or AI.

[ Music ]

[Jacques Arsenault:] I'm Jacques Arsenault, and with me today to discuss
this fascinating subject, we're talking not only to GAO's Chief
Scientist Tim Persons but for the first time on our Watchdog Report,
we're happy to welcome in a special guest, Federal Chief Information
Officer Suzette Kent. Thank you both for joining me. 

[Tim Persons:] Thanks very much. 

[Suzette Kent:] Thank you for having us, Jacques. 

[Jacques Arsenault:] So, Tim, why is there so much hype surrounding AI?
Can you help us tell the difference between what's reality and what is
fiction? 

[Tim Persons:] Right. So, it's a great question, Jacques, because as you
look at AI in the reality versus the hype narrative, this is very
important to understand. Essentially, AI has been a vacuum that's been
filled by Hollywood for over half a century. I mean, remember back it's
now 52 years old since the "2001: A Space Odyssey" and it's a scary
movie in parts even though it captured the imagination about technology.
But, it was just scary because of the bad guy in the case was the Hal
9000 computer. I think it spoke to some of the key unsettling issues
about AI that if not properly understood can engender fear in terms of
the narrative on things. So, those are two things. One is the fear or
perception of the loss of control as well as a second one that really is
pervasive in this day and age as well. That's really the fear of the
loss of meaningful work or being replaced. And so, that's the mythology.
The reality is different in that it's a technology that's already here.
It's pervasive. It's been commoditized, and it has technology that has
its limitations, but it has come a long way over the decades, and it is
impacting work. It will impact more routine and repetitive task oriented
jobs, but on the other hand, it could enable work and hire a more
complex work for people. So, in many ways, it's much more of an
augmented type of intelligence versus replacement. 

[Jacques Arsenault:] So, we shouldn't have any fears about the
singularity coming in the next few years. 

[Tim Persons:] I don't think that's a reasonable concern in this case.
Where we are now is we can still use AI for common, everyday tasks, but
it's not something that AI's ready, sort of take control of everything
and do it without human intervention. 

[Suzette Kent:] And, Jacques, Tim hit the most important point there is
that even though there's, you know, wild and vivid speculation about
what it could do, and it  is an incredibly powerful tool, it still
requires human direction, human construction, and data that is put
together by humans. So, there's a lot of pieces in it towards that, you
know, control and change and, you know, there's also a hype that is the
silver bullet that's going to be able to do everything and, you know,
kind of wrapping into what Tim said. A lot of power there, but it still
requires deeply involved, human directors in the process. 

[Jacques Arsenault:] So, thinking about that process and the journey
that AI has taken up to this point, Suzette, can you talk about how AI
has changed over the years and what can we expect to see in the future? 

[Suzette Kent:] We have become much more insightful about the critical
link between what AI can do and those capabilities and the data that we
have available. Inside the federal agencies, we're getting much better
at defining the specific questions that we want to answer, and we learn
where AI can be most useful, and that has something to do with the
clarity of the question, the available data, how comprehensive it is,
and the types of controls that are in place and the types of things that
are aligned with values, ethics, and mission, you know, of those
particular agencies. So, I think we're going from wild speculation to on
the ground use, and as we continue to, I'll call it move up the food
chain, on what those uses look like, we learn more about the operating
environments that we have to have in place, the infrastructure, and the
human interaction that's required and what we will need to do in the
future, not only for the workforce but people who are consuming the
outputs that may come from things that are driven by AI. 

[Tim Persons:] How AI's changed over the years is really based upon the
convergence of key technologies that we've seen today which can be a
challenge from one sense, but on the other hand, there's an asset with
that in that you can now store that data, and you can compute on that
data and do communications with the data with advanced algorithms and
things that, you know, heretofore were unimaginable. So, all those
things sort of merging together has created ideal conditions for AI. 

[Jacques Arsenault:] So, knowing there are these powerful algorithms
ingesting data and using it to impact our lives, Tim, could you talk
about some of the changes you see AI having on everyday life in the
future? 

[Tim Persons:] I, for one, would be a big fan of an AI that could real
all of my email and triage and give me exactly the things I need to do
today and exactly the most important decisions. I think there are
assistants now trying to do that, but that's certainly something that I
would love to see personally for my own workflow. But, I think we will
see, again, more augmentation of tasks so that it actually replaced
things that we might have done that are more mundane or routine or
repeatable and we're going to be leveraging more and more automation
with the usage of statistical computing and data and things like that.
Now, I have my kids, for example, already learning to speak to devices
and they're capably interacting with an Internet of Things and AI
algorithms to answer questions or help them out or play music or
whatever, and they're totally unaware of the sort of the computer
science and mathematics and engineering miracles that have been behind
what we have today, and yet they don't need to. They can still use the
technology and still do things without having to have a master’s degree
in computer science, and that just gives me hope that what they're going
to be doing moving forward in their jobs is leveraging more and more of
AI. 

[Suzette Kent:] This is an area that actually gets me excited, and when
we think about the government uses, Tim and I have talked about the
things that happen in everyday life that citizens already experience
because data's being collected from so many sources. And, because they
experience these in everyday life, as we think about other applications,
maybe that's a way for us to get more comfortable. And, what I mean by
that, searching and marketing and entertainment, and many of us may have
something suggested to you based on something else that you did. And,
that's, you know, using different types of capabilities and inputs to
suggest or predict a likely behavior. 

[Jacques Arsenault:] What about the ability of AI to take all of this
data and allow us to be more prepared for things that as humans we're
just unable to predict? 

[Tim Persons:] Thinking about these systems from where they're well
known, or we had them being able to do a baseline amount of what's
called descriptive analysis. So, you can use them, and they can say,
well, here's a snapshot of an as is, kind of, scenario. So, that's
great. Now, with the power of AI and moving from descriptive into a
predictive scenario where you're talking about what you think will be.
And, AI's going to help out with that just like Hurricane tracking and
prediction and those sort of things that we see in everyday life. But,
even more exciting than beyond just the predictive part is then the
prescriptive, right. Given what you think will happen, here's what the
AI might recommend should be done based upon massive amounts of data and
calculations and things that no human, no matter how smart we may be,
would be able to ingest and make sense of at one time. 

[Suzette Kent:] Some of those things that we've had these great
conversations inside the government about is think of something simple
with the city, tree maintenance. I don't go just every third month and
cut the trees. What if I understood the data from rain, wind, aerial
views, you know, city maintenance, and I knew when to go take that
preventative action so that a person, a building, or transportation
wouldn't be disrupted. Or, if we saw a storm coming and we know in our
weather model, you know, can predict what flooding might look like, and
we know what it takes to recover from that. At the same time that we're
taking actions to save people, we're already forward on bringing
supplies for, you know, food, lumber, construction goods so we can get
back to normal as quickly as possible. 

[Jacques Arsenault:] Personally, if I find myself watching Netflix or
Amazon Prime, listening to Spotify or just seeing adds online, it gets
me wondering about other ways that AI might be impacting me. When we
come back from the break, Tim and Suzette will talk about something
called natural language processing and how that impacts not only
everyday customer services but also government services. 

[Tim Persons:] Customers have bought what you just bought, also buy
this.

[ Music ]
You're mentioning, like, movie suggestions or one thing that's been a
big impact on our lives, but even in a similar way, just going to a
major shopping site or online presence. You know, if I buy something, as
soon as I buy it, they'll say, well, you know, customers that bought
what you just bought also buy this. Or, if I'm buying a Lego set for
what looks like a certain age range, they'll assume, like, well, okay,
well, you must have a daughter of this age range and let me suggest
books of this author or something like that. So, certainly we see that
in our everyday lives. I've been thankful in recent years being able to,
when I've been traveling, when we were traveling on airlines, I've been
able to change flights by calling an agent that sounded a lot like a
person where just a few years ago, it would sound very robotic. You
could absolutely tell that you were dealing with a machine. It was
frustrating. You would have to repeat yourself. You would have to revert
to pushing buttons. And, now, with what's called natural language
processing, you can interact with an agent, so to speak, that's actually
a computer AI at the other end and get very pristine service. 

[Jacques Arsenault:] Suzette, what are ways federal agencies are using
AI to improve government services? 

[Suzette Kent:] One overarching one touches on kind of where Tim was
going, but with use of natural language processing, image, or collected
locational data, we can do things to make interactions more secure. So,
many agencies are using AI as they look at threats. They look at access,
and they use it both to protect access to federal networks and data as
well as interpret, you know, what they're seeing. We have experiments in
hydroponics. So, how do you use AI to maximize food production where,
you know, light and water and contents of the soil. There's been a lot
of discussion across the Department of Transportation and others about
autonomous vehicles of many types. And, think about the amount of data
that you have to have, you know, for an effective autonomous vehicle.
We're seeing in HHS, you know, right now, we're seeing AI as well as our
high performance compute environments used to accelerate the work to
find vaccines and to look at relationships and correlations. We've also
used it inside that same department for some operational activities to
help with regulations and things like that. We have a community of
practice around AI inside the government that's operated between GSA and
the CIO counsel on the executive branch side so that agencies can both
share resources and insights and most importantly use cases as we are
moving down this path of AI. 

[Jacques Arsenault:] Suzette, what do you think is the most important
component of AI? 

[Suzette Kent:] I want to ensure that data stays top of mind. In many
cases, we can imagine a lot of uses, but having the data to power that
is where we find ourselves in need and spending time and ensuring that
not only do we have the volume, that it is free from bias. And, just
because it's historical, there still may be things that we need to
address. And, that's why we have side by sides of technology and data
initiatives in many of the areas of the agencies, but also as we talk
about controls and oversights. 

[Jacques Arsenault:] Now, Tim, there have to be enormous challenges in
many aspects in getting buy in at the federal level. 

[Tim Persons:] There's a benefits and challenges narratives with AI
everywhere the federal government may engage with it, right? It could be
a regulatory function. It could be a research and development function.
It could be operational entities that are trying to solve real world
problems, and everywhere AI goes into, there will be a benefits
challenges narrative, and understanding that quickly and being able to
adapt in an agile manner more toward the benefits and recognizing and
mitigating the challenges is the proper narrative across the board. 

[Jacques Arsenault:] So, then, what is the federal role in overseeing
AI? 

[Tim Persons:] So, it's a simple answer, and yet there's complications
behind it, meaning it's simple in that the federal government absolutely
has a role of oversight in AI. The oversight role, however, is context
dependent. If you're talking about let's take the Food and Drug
Administration, our key regulatory entity, about determining medical
device safety or drug safety or Suzette mentioned vaccine development's
a key and exciting area where AI's developing it, but it also help with
de-risking the vaccine candidates through the chemical trials process,
and so on. And so, it has a key oversight role in that regulatory sense.
However, if you shift over just to a different part of HHS and look at
NIH, what they're going to want to oversee is it might be the data
quality issues and things that go into the research that are going to
inform what we hope are those disease curing and vaccine discovering
kind of research activities around the institutes. And so, let me just
mention one more, a third, in just in the HHS umbrella which is Medicare
and Medicaid services. So, CMS, as it's called, is where you're talking
about billing and so on, and they are going to want to have, and they do
have machine learning algorithms that try and make sure that claims that
are legitimate for medical care, billing to Medicare are properly and
timely responded to or paid. But, at the same time, to protect taxpayer
dollars, they want to be able to select out fraudulent or erroneous
things and so on, and so you're going to want to have oversight into
that process to make sure that the legitimate claims for urgent medical
needs and things are timely responded to by CMS, but at the same time
money is finite and so on. And so, you want to protect the taxpayer
resource so that more benefit is paid out properly and not improperly or
fraudulently. 

[Suzette Kent:] When you think about the federal government's role, we
have a role both as supporting American industry and that's some of the
external things that we do around AI, as well as we have an important
commitment to citizens about how we use it inside the federal government
in those purposes. And, you know, as part of the industries of the
future and the President's management agenda, there have been a clear
marker stated that, you know, we want to be a world leader in this
space. And, we want that leadership and the way that we approach AI to
be in alignment with American values. And, that means that, you know, we
still protect privacy and we don't, you know, discriminate, and we act
in a lawful way, and that things that we do are traceable, repeatable,
and explainable. And, those all sound like great words, but we have to
be able to prove those things. So, we have some responsibilities around
how we think of it in our context as a world leader and, you know, what
that looks like in our interactions in American industry. You know, in
the federal government, a lot of the AI that we use, just like our
conversation about being a citizen, we're actually requiring form
somewhere else. So, we have to understand it, and those things that we
build, we have to build with very specific intention to ensure that the
context it fits the outcomes, and it is in the alignment of the
commitments that we've made to serving, you know, that particular
mission or set of citizens. 

[Jacques Arsenault:] So, context is key. One size clearly does not fit
all in approaching oversight of AI and the way that technology is
applied to AI in the safety of our automobiles, say, is different than
how it's applied to the criminal justice system. But, make no mistake,
AI is reliant on data, and many of us are worried about the personally
identifiable information or PII in that data. When we come back, we'll
talk about how policy makers can think strategically about the use of
personal data in government programs. 

[Suzette Kent:] What is appropriate in this context, and do I have
appropriate, you know, permission from wherever the data came from. If
it came from a citizen, is this in line with the purpose for which I
collect it?

[ Music ]

One of the areas where there's a lot of debate across the federal
agencies is the balance between protecting privacy and the accuracy of
the data in the outcomes. And, an example of if you anonymize data, in
doing that, you miss some of the information like age or gender that may
be very important to the outcome. If you deidentify data, you know, you
have to create, you know, tools and structure depending on what you're
going to do with the data on the other end to put it back together. You
have to authorize uses in different ways. So, as policy makers, we have
to think about back to that context question. What is appropriate in
this context, and do I have appropriate permission from wherever the
data came from? If it came from a citizen, is this in line with the
purpose for which I collected it? If it has to do with, you know,
health, safety, and welfare, do I have other reasons for using it? And
then, in many cases, in what we've been currently doing with the virus
is a litmus test there. When the data changes very frequently, how do I
ensure the integrity, ensuring that the data is accurate, has not been
manipulated, and is appropriate is another set of works. 

[Jacques Arsenault:] So, what about the flipside of that? You know, not
having enough data. 

[Suzette Kent:] It's the right balance, you know, the right use, in the
context, then those are the types of things where agencies are still
going through examination, experimentation, and building the
perspective. 

[Jacques Arsenault:] So, Tim, what can policy makers plan for now in
order to be able to adapt and balance opportunities and challenges that
AI presents? 

[Tim Persons:] What we've heard the community or experts talk about this
or even practitioners in their areas is of course baseline, you know,
picking up on what Suzette was just saying is more research, clearly, is
needed. We are awash in data. It's not always the right data. AI is,
would still suffer from the risk of, you know, garbage in and garbage
out, so we need to have a data-centric strategy in terms of research.
Whether you're a conventional research entity like a National Science
Foundation or things like that, but you also need to think about it if
you are in an operational space or if you're a regulator or things like
that. And so, more research is needed, is true for AI across the board,
regardless of your mission. The second issue is, again, looking at data.
It is the lifeblood of AI. Just like we have blood in our veins that
gives us life and so on, so to, you can think about an AI system with,
you know, having good data is, leads to a healthier system and healthy
operations. And so, you want to have it, not only have the quality of
the data but the access to that kind of data. What's the right time, the
right size, the right place kind of approach to data access at times
where, you know, frankly in the federal government, we don't think in
terms of that kind of way of a responsibility to provide versus when we
might think automatically about need to know. And, a lot of that is
completely legitimate. It's just that with the AI usage adding into
data, increasing access can yield to better performance and outcomes,
albeit again, you have to identify the risk you're trying to avoid,
let's say, you know, to personally identifiable information or violating
privacy and civil liberties. The other thing that I think the government
needs to do is essentially have a standardization kind of approach. This
is often something that's left off. We do need more research, but we
want to be able to have AI operate in its context, but can we think
about standardization of data and uniform standards for data an
algorithms that can help drive solutions. And, there are science
entities or, you know, science societies and things, medical societies
that could help do this as well. This is a, I would say, a [inaudible]
thing. But, I think that would be important. And then, another thing
policy makers can do, really a lot of this is about the human side of
things. It's not strictly getting the algorithm and the machine right,
but we need to invest in human capital. We need to rethink how we're
training the next generation to, again, use these technologies to
augment their various appeal to practice in their vocations, not to
necessarily replace, but to do so, so that their productivity, their
outcomes, and so on can be far beyond what their predecessors could even
imagine. And then, I think another thing the market is going to need in
terms of the private sector because that's going to be important, but I
think bringing around regulatory certainty is critical with these
technologies if you're in a regulatory context. That could be something
that's done in a process where it's, again, cross-sectoral[] usage of AI
to help drive and bring more regulatory certainty could drive better
outcomes in terms of American competitiveness and the innovation
systems. 

[Suzette Kent:] Last year, towards the end of June, we updated the AI
national research and development strategy, and that was issued directly
by OSTP, but it was a representation of input from many of the agencies,
and it made commitments around what the government will do for private
sector. One of those was the point Tim just made was being very careful
about partnering with the private sector in developing regulations and
not jumping ahead before we have, you know, understanding. It was
commitment to long term investment. It was collaborative discussions
about the ethics and, you know, legal, societal implications. Had a lot
of elements of reexamining safety, and it also made some pretty bold
commitments around those public datasets and environment that we could
understand and examine, and then one of the ones that I think is really
important for how we go forward, a commitment around the workforce
needs. And, that was looking at the context of the workforce in the
light of some of the comments as to how we started the conversation of
what is the turning point when we need different types of pathways for
people to become professionals and operate in this space. I know as we
formed the select committee on AI about two years ago, a few of the
universities were the very first ones to actually even have a major in
AI, and we know, you know, based on, you know, that that's not the only
path, you know, into technology. But, I thought it was an interesting
indication of where we are in preparing the workforce that it's not
something that is completely covered through traditional academia. So,
we've made commitments around regulations, commitments around funding,
commitments around supporting the workforce, but we're going to continue
to iterate on that to ensure that we are building out the technologies,
the workforce, and those protocols that we need to be a world leader in
this space. 

[Jacques Arsenault:] And, Tim, what is GAO doing to ensure artificial
intelligence accountability? 

[Tim Persons:] From GAO's perspective, we are or have put together a
comptroller general forum on artificial intelligence accountability. The
key goal is to develop an accountability framework around AI systems
that can be used not only by GAO, but by our federal partners or others
to try and get a sense for how one does conduct oversight on AI and
machine learning systems. 

[Jacques Arsenault:] When we come back from our last break, we'll talk
to Suzette about her recent announcement that she's leaving her position
as Federal Chief Information Officer and her next steps. 

[Suzette Kent:] I will share that my next endeavor's going to be deeply
focused on many of the topics that we talked about today.

[ Music ]

This has been an incredible opportunity, and the things that we talked
about today are square in the center of one of the reasons that, you
know, I came out of the private sector to spend time in the federal
government because it is so important that we have perspectives from
both private sector, you know, and government to advance the ball. And,
in the near term, not doing anything definitive, but I will share that
my next endeavor's going to be deeply focused on many of the topics that
we talked about today and how we do continue to leverage technology to
drive not only U.S. leadership in the world economy, but we improve the
lives for citizens in the U.S. Whether it's recovery from a hurricane or
ways that we look at the dynamics of different types of support
structures, to creating meaningful lives for citizens. Those are the
things that are really exciting, and I look forward to continuing to
work with Tim and team because it's going to take a unified effort
across the Executive Branch and Congress and deep interaction with
citizens for us to get this right. 

[Tim Persons:] I look forward to working with you, as well, Suzette.
Thanks very much for that, and thanks for, again, all your leadership
and service as our Federal CIO and certainly leadership in this AI area. 

[Jacques Arsenault:] Well, my final question I'll pose to each of you in
turn, and I'll start with you, Suzette. What would you say is the bottom
line when it comes to artificial intelligence? 

[Suzette Kent:] I think we are still very early in the journey, and
there are amazing possibilities that we can imagine. We have to move
down that journey, you know, with caution, but in light, with
excitement, but bounded by American values, the mission of agency,
appropriate context, and focused on the outcomes, you know, that we're
trying to achieve. And, ensure along the way, as we're getting excited
about a new technology and we have to ensure that the technology, the
businesses process, the data, and the workforce and people receiving
those services are all at the table as a part of this journey. Unlike
many other technologies, this has the ability to move faster, and we
have to be very vigilant in ensuring that, you know, everyone has a
voice in the discussion about how we use those powers. 

[Jacques Arsenault:] And, Tim, what would you say is the bottom line? 

[Tim Persons:] This is not the robot apocalypse. We're not going to lose
control, and it's not going to destroy the world and take over things
and so on. I think that's important because this is really about
augmentation of human function and intelligence and not a complete
replacement. Which leads to the second part of it is that this isn't the
job-pocalypse either. It's not going to outsource us all. We just have
to think differently. You think on the way we do our jobs, break things
down into tasks, and are those things that are more mundane or
repeatable and so on, we should embrace whole heartedly to completely
augment ourselves to that we can be grander and better version of
ourselves in whatever it is that we do. And then, related to that is
that there's no vocation that won't be touched directly or indirectly by
AI. I think that when we look back at this timeframe, we can't imagine
not having AI around to do things. It's just, again, just like
electrification where we can't imagine not having outlets in our walls
that when we plug in our devices that, you know, electrons will flow
through, and we'll be able to do things. So, AI will be that kind of
same kind of incredibly transformative technology. I think we should
embrace it. I do think that we need to be cautious about it. So,
skepticism is warranted, and that's where I think we just need to work
on managing those risk around it in the proper context. 

[Jacques Arsenault:] So, the risks of artificial intelligence are there,
but the tremendous improvements in our lives and the economic advantages
that AI provides make the management of these risks, in the proper
context, so critical. Thank you to Tim Persons, Managing Director of
GAO's Science Technology Assessment and Analytics Team and GAO's Chief
Scientist. And also, to Suzette Kent, the Federal Chief Information
Officer for their work in this area and for sharing their efforts and
experience on the subject of artificial intelligence technology. And,
thank you for listening to this Deep Dig edition of the Watchdog Report.
To hear more podcasts, subscribe to us on Apple Podcasts. Leave a review
while you're there and let people know about the work we're doing.
Follow us on Twitter, Facebook, LinkedIn, and now also on Instagram, and
for all things GAO, visit us at gao.gov.

[ Music ]