The Impact of Emerging Technology on Industry, Government and Society

The Impact of Emerging Technology on Industry, Government and Society


now Stuart we’ve just heard from you
Wendy would you like to introduce yourself okay there we go Nick professor of AI at Imperial College
London I’ve been a professor I’ve been a AI researcher for about 30 odd years now
and I also spend time in government as the chief scientist for national
security for six years yeah to pass some would you post your mind
for second thank you sir good morning I’m Terah Lyons the executive director of
the partnership on AI or non profit organization and a multi-stakeholder
initiative with over 100 organizations involved thank you and Stuart we’re
gonna come to you shortly on on some of the points that you raised in your
presentation but tier if you would take the microphone I’d love to start with
you and just have you explain a little bit about your organization some of the
challenges you guys have identified and the corporate founders and they were
joined by organizations like the American Civil Liberties Union and the
partner foundation and a lot of other civil society and academic interests
actually so our coalition includes organizations like the International we
are essentially trying to build best practices for the response the
partnership was founded on the premise that government to industry frankly
where a lot of the power is situated right now my personal background is in
policy anything in the United States government so I worked the Obama
administration working on AI policy issues for about the last three years
the second term of that presidency at the time the private sector was
investing in AI researcher development at eight times and so it’s very clear
with the levers I’d love you to talk about what a specific example of the
kind of challenges that you’ve taken on because that’s an extraordinary group of
people to put together i mean i’d i mean that getting i mean getting events like
this together is hard i think actually you probably got the the tougher job
there but what kind of problems you specifically jumping on and how do you
get everybody two together aligned so we recently just released a report and this is a big issue especially quite
broken’ frankly and fairly racist among other dynamics and the state of
California just taking one local policy sample just has to fill that abolish the
money bail system and replaced it with a mandatory per county requirement to use
algorithms these rules are nowhere close to being what are the issues what’s the
issue with the systems to more structural challenges like whether or
not judges can actually interpret so we released a set of requirements and
standards which are statistical and also structural and really hopefully you know
local governments start adopting policy measures of the sort like the one that I
described they will listen I mean God couldn’t get something more
central than the criminal justice system to to make sure we get these decisions
right Nick I know you’ve spent a lot of time thinking about and applying
yourself to how a AI is gonna impact government and you spent time in
government you know what what’s your perspective on where we are and the
challenges and the work to be done left it and actually what I I think
government faces many challenges that many companies also face so I think
there’s sort of a lot of interest and a lot of excitement about AI I would say
there’s a real dearth of expertise within government to be able to make
sense of what AI can do so I see lots of lots of departments thinking about well
this intractable problem that we’ve had for twenty years we’re gonna sprinkle a
bit of magic AI dust or machine learning dust over it and lo and behold it will
solve all of our problems and sort of there’s a lot of fear of missing out
FOMO something should be done but it’s not necessarily something that is being
done so so I think there’s that lack of it lack of expertise we started I think
I’ve seen it in the last couple of years of sort of era of small-scale pilots and
prototypes taking a small bit of an organization and sort of seeing what I
can do but to sort of see major government departments looking at and
you wrote a recent paper on machine behavior and it covered a lot of really
important issues around how we you know bring machine and humans together in a
in a positive and cooperative sense I think you know if you could you know
pick one area that you think government should be cracking onwards and really
applying AI to what what would you single out and I think the way this is going to
happen is through different understanding mechanisms on the details
of yeah fascinating when diem I know that there’s a lot of parts of the
government that you’ve been involved in are not least authoring the the
industrial strategy paper and you know obviously a lot of investment has gone
in time energy and and obviously you’ve made a huge contribution it would be
great to tee off that conversation around government and where we are
heading but also I know that you know you’ve got quite a global perspective
that you’ve been developing and I’d love to dig into that as well twice how best to long way to go but we
really need to using stuff in the public and private sector to stay ahead of the
game but the other thing I want to point perspective is that AI today which is
effectively machine learning deep learning requires on requires dates a
huge amounts of data that’s why we can do what we do but the access that days
is really crucial and so data governance is as important as in Internet
governance it’s the subset of it in a way and we see I’ve written the paper
recently The Cure now harvest our hampson called for Internet we’re
talking about the fragmentation of the Internet what does that mean it’s just
to break it down in simple terms in the future that might not be possible
because of the way the internet is fragmented they want us they’ve got a big enough
market locally not to feel they need to do two hands typing I’d about China’s a
complete different model again if you do anything to do a date on the inside
China you have to assume make that date accessible to the government Chinese
government and there’s all sorts of consequences that we could discuss it’s
not all that they solve it is to save a place to be give up freedom speech all
sorts of things and Russia is going to same way as China India the same sort of
size of populations China could go the same way it’s a democracy but it could and we just went to come on the internet
and they’re largely in rural China rural India so the shape of the incent the way
works and operates could change really quite significantly in the few years
which will affect our freedom to develop products and services and do you think
that’s a trend that’s gonna continue or do you think it’s something that you
know we should there should be a call to action to try and address that
fragmentation I mean is it inevitable at this point the world so talking about but I do and I’ll finish by saying that
the end of May and I do think I think that makes
tremendous sense and and the question certainly on my mind is always you know
what body is that all bodies is that that has that conversation because you
know see we have things like the UN we have lots of government initiatives you
know in every single country absolutely I want to come back to Stewart and give
you a chance to comment on on these points because I think this this for
Internet’s concept is is such a big one and data obviously fuels AI but so would
you like to comment what’s your perspective on this this point around
governance and and and data behind AI there are governance issues that and
Wendy mentioned not wanting to be killed they are by AR and we’re actually trying
through the UN General Assembly to get the resolution saying that you should
not design choose to kill humans that seems like
something that’s pretty common sense at present the UK government is opposed
to that principle so oddly enough even though the UK government itself has
declared that it will not use algorithms to kill human beings it is opposed to a
treaty where other countries would agree to do the same thing so we have a sort
of a contradiction otherwise known as the UK wanting to commit military
suicide and just I think that’s a really that’s a huge point right so why leaving
the UK’s current policy whatever it may be aside for a second why would anybody
disagree with the idea that we would ban leaflets on immerse weapons I mean
that’s it that’s a big topic so let civilian casualties we have seen fewer
civilian casualties from semi-intelligent precision weapons and I
think you could argue perhaps that that might continue but I think actually this
is a more or less a red herring the real issue is that autonomous
weapons would be weapons of mass destruction and we would see attacks on
a vast scale because you’re basically putting something that’s more powerful
than a nuclear weapon much much cheaper and can all can also just for fun the
ethnically targeted yeah no just I wouldn’t let the rest of the panel talk
about that point just for a minute before we go into the the beneficial
side of things but obviously in the partnership for AI I don’t know whether
you have a specific policy on this but on government and the work
you’ve been doing Wendy and Nick I mean this is a I mean you know this is a huge
point you know you know do you regulate you not regulate and and and how does
the UK or any other country you know make a decision on this and and decide
whether they participate or not it is a complex issue in general about sort of
where we regulate for computers and algorithms making choices I think it’s
important the one disentangled sort of AI from computing in general I hear lots
of things about sort of AI shouldn’t be used to to control these autonomous
weapons and that’s nothing no computer algorithm should be used for
that whether it’s using AI or not is is more or less irrelevant tournament so I
I hear lots of things about that a pageant targeted at AI so biasing
algorithms there should not be biasing algorithms me that’s irrespective of
whether those are AI based algorithms or whether they’re not so I think it’s
important to sort of distinguish and disentangle the piece about what’s
specific to err on what we apply to algorithms and and digital systems in
general I think it’s important I think in some cases it’s reasonably clear cut
that I would agree with Stuart that sort of be inappropriate in general to have
algorithms making choices about what about who and where missiles are fired
for example I I think in general that some of the some of the other decisions
are more complicated and involved there’s a lot of interesting things
around autonomous systems and I think we need to have a proper societal debate
about where what we’re willing to allow within our society and and part of the
Machine behavior paper was actually arguing that it is that societal debate
it’s not shouldn’t just be left of people doing the tech or just absolutely
a philosophic and that’s very much in line with what
we try and do here by building an event that’s inclusive and and broadens the
discussion I mean wendy is it is it even possible
to get agreement to ban something like this one I need to work at that UN level
to get these agreements but I can also see why the British government appears
to be schizophrenic about it because at one level you can say we’re not going to
do this but when you’re actually negotiating UN level you can’t blink
before other people blink or you’re not going to agree until you know other
people are going to agree it’s the it’s the nuclear bomb type thing or the but
I’m actually better students made this point I don’t know today or previously I
definitely I know he made this point the chemical weapons agreement that has held
since the first world war that more or less apart from you know things going on
in the Middle East is on but generally with the big powers that’s held because
it was so awful in the first world war they agreed they didn’t want to do that
again what we don’t want to have to do is have that sort of calamity caused by
AI and then agree we need to agree it before that calamity or that disaster good well I think that’s I think it
feels like every is an agreement that this would be a good thing to do and and
it sounds like there’s a lot of work underway but and you know with the only
the only challenge that I’ve heard to that whole debate is you know are people
able to create lethal autonomous weapons themselves today anyway by buying things
off the shelf and programming them and how would you
control that and then obviously what happens if there isn’t universal
agreement you go and buy them and you put them
together and make chemical weapons but the chemical manufacturers are part of
the treaty and so they are required to know their customer they’re required to
account for a large scale purchases and we also have a big intelligence
operation to make sure that the wrong people are not are not buying chemicals
and because of the treaty it’s completely stigmatized so when Syria
uses chemical weapons there’s an international outcry and the u.s. feels
entitled and is supported by the rest of the world community in sending cruise
missiles to to blow up half a Syrian Air Force for example and there’s absolutely
no doubt that a treaty would have a very very significant effect in reducing the
nefarious uses of chemical weapon you know just because it’s possible for
people to go out and shoot somebody or stab them with the kitchen knife doesn’t
mean we shouldn’t have laws against murder right we have laws against murder
even though sometimes people break those laws still a good idea to have that laws
and the argument that we shouldn’t have a treaty because someone might break the
treaty I just don’t understand it yeah okay well listen it’s great to have had
that discussion and I do want to move on to to the optimistic and exciting future
that we see ahead for AI and maybe actually what you know you gave a great
talk earlier and and you know the trillions of dollars that we have ahead
of us some of the challenges we still need to address to make sure we can
truly realize that I think you know work that you’ve been involved in recently
leading some of around beneficial AI could you talk a little bit about that
Stuart yeah so I mentioned that I’m talking about this tomorrow morning in
more depth but in a nutshell the reason why the standard model of AI which is
optimizing a fixed objective doesn’t quite work is that it’s often impossible
to by the objective correctly so for
example if you say I’d like to cure cancer as quickly as possible we’d all
like to cure cancer as quickly as possible if you take that go literally
however horribly the fastest way would be to induce tumors in the entire world
population so that you can run millions of medical trials in parallel and find
the cure most quickly that way and of course that’s not what you meant and we
call this the King Midas problem because he said I want everything I touch to
turn to goal that wasn’t what he meant and when he got it of course that
included his food in his drink and his family and that was a big mistake
and so in AI you know since the early 80s or even the late 70s we’ve been
completely comfortable with the idea that there’s uncertainty in the world
that we can’t have certain knowledge of our sensory information we can’t have
certain knowledge of the dynamics of the world how things are going to unfold
what’s in other people’s minds uncertainty is completely pervasive but
we totally forgot about the objective we just continued to assume that there’s
perfect knowledge of the objective and that was just a mistake I don’t really
know how it happened except that some twit wrote a textbook where he forgot to
mention that there should be uncertainty in that in the objective and when you
when you think about that seriously when you think okay a machine is supposed to
optimize an objective but doesn’t know what it is right that actually makes the
problem completely different from optimizing fixed objective in particular
it means that the machine is intimately coupled to the human being because the
human being is the source of the objective and everything the human does
provides evidence about their underlying preferences and so the kind of AI I see
going forward is actually a very different discipline it’s not one where
the AI knows what it’s supposed to be doing and is completely oblivious to the
human it’s one where the AI is actually tied to the human because the human it
could for example say no I don’t like that or do it this way or you know what
stop and a classical AI system under the
standard model would just ignore that because it already knows the true
objective and it knows that whatever it’s doing is optimal and you just
saying stop well so what that’s just hot air
right but a system that knows that it doesn’t know what the true objective is
has to pay attention to the human and I think that so the idea of beneficial AI
is that we build systems this way and you can actually prove under certain
assumptions that at the moment are a bit too strong but we will weaken them over
time you can prove that those systems are actually going to be beneficial to
humans and it’ll be a completely different theory and we have to sort of
rewrite the textbook unfortunately from scratch well it’s interesting because
obviously it feels like there’s so much momentum behind AI development right now
and you’re pointing it like what feels like a really fundamental you know area
of work which is only just beginning so we already talked about the idea of
classifying images using machine learning and so when you when you train
those algorithms you’re minimizing what’s called a loss function and the
loss function is supposed to say what’s the cost of misclassifying an object to
type a as an object of type B for example what’s the cost of
misclassifying a human being as a gorilla
well Google did that Google photos did that in a very public way and the answer
is hundreds of millions of dollars or maybe even billions of dollars of
goodwill and bad public relations for Google as a result of one single Mis
classification error now I bet you that they train their classification
algorithm with what we call the uniform loss function meaning that all mistakes
are equally bad so misclassifying and Norwich Terrier as a Norfolk Terrier is
just as bad as misclassifying a person as a gorilla well clearly that wasn’t
the true loss function but that was a loss function that Google used in their
algorithm and if you actually sat down and asked yourself do I know what the
loss function is you’d say no I haven’t thought about it they
twenty thousand categories so that’s four hundred million possible types of
errors that I could make and each of them has its own cost I don’t know what
that’s going to be so I have to be doing machine learning under uncertainty about
the loss function and then you get completely different algorithms even a
completely different workflow for machine learning it’s fascinating and
you know in some ways it’s comforting to know that we’re obviously not as far
ahead as perhaps some people think we are in the development of AI and and you
know there are say general intelligence or super intelligence but it feels like
there’s a lot of work to be done Terra is is this an area that the the
partnership has looked at and how do you think about you know getting the
objectives of the AI right I think just expounding on something that Stuart was
describing earlier I think that once the greatest the deepest right now and the
bias embedded in systems a number of different stages of development and
deployment is a huge limitation to effectively producing responsible
technology right now so you know I think that one thing we haven’t talked about
yet but it’s extremely important to the field both current and future States is
for example what the developer community looks like and a big conversation in the
partnership right now is how do we bring more people more inclusively to the
table show this conversation there on governance either at the governmental or
intergovernmental level or even at the institutional level where product
managers are delivering and deploying technology in these companies every day
how do we diversify the people who are every single scale so I think and that
diversity comes back to making sure that the data going into the system is not
just the important part of diversity but also the people programming the systems
as well yeah it’s a full-stack set of interventions it starts from training
data it expands to the context into which technology is or isn’t getting
deployed and normative decisions we make as a technology community about how
understanding contextual circumstances structural inequity jurisdictional
differences as we talked about today so I wanted to switch topics just as we
move towards the the end of the panel and we’ve got just about 10 minutes left
to talk about the one of the big reasons that we’re all here and and and and one
of as Tabitha said asked a question of how we can point our tools tech and
talent of this industry at helping with the delivery of these seventeen UN
sustainable development goals and I’d love to ask each of you you know to
think about the opportunities you think are in front of us both now and in the
future to really move the needle there and Wendy would you care to go first for
that across in our part of the world anyway there are many Malaysia Singapore
the B Stephen I go into a computer science classroom in a university and
see more than 2% the class of women which is fantastic amazing yeah when I
first started in computing AI was not a subset of computer science AI was
philosophy and psychology and alternative building building things
that worked like the human brain it has and the a I were in now is a deep subset
of computer science and it’s machine learning and so we are taking a pipeline
that’s very not diverse on gender particularly I want diversity across the
board including nowadays age right but an accessibility with a stick and
getting up it’s not that easy charlie but the for me in gender where now we’ve
got a pipeline that’s not diverse in gender and we’re taking that and taking
a smaller it’s even worse than AI and this is so important because of the bias
issues we’re all biased anyway but if you’ve got a very small subset of
society building the algorithms then it’s not fit for society and we’ve
absolutely absolutely got to make this part of the ethical framework and the
way that we regulate AI diversity SP right up there in terms of if your
company is not taking a diverse approach or has an interdisciplinary team even in
a small start-up you can surround yourself with with people testing in
other ways in which you can actually make sure that you’re thinking about
your product right up the top and we can do that now even before you know because
one of the UM sustainability goals is gender gender equality yeah when you see
what’s happening in India as the internet rolls out on top of that the
women are not getting the phones yeah how’d you put some numbers around
how big is it right now yeah it was all mostly young but white males
right there wasn’t they could not there was not one woman in that fuselage right
it is huge huge and this conference having mobility we are such a big focus so before we move on from that point cuz
that that it is a huge one in an attempt that was here she would be saying
exactly the same words which is so what can everybody hear listening and in the
room do to contribute to that what’s the most important thing they can do to move
that forward brilliant thank you Nick sustainable
development goals is that is the kind of why didn’t know no I know what diversity
is hugely important and this is all we’re gonna have time for in terms of
the closing question but would love you to you know talk to your perspective on
that the one where I think AI can have a really strong impact is around health
and well-being I think sort of the it is an area both physical and mental health
where where it’s very much right for innovation we have lots of lots of data
and lots of opportunity all around the world so I’m not just talking about
Western health care systems I think there’s a the we’re on the cusp of being
able to do and prescribe things and treatments and social treatments at a
level that we’ve not been able to do previously I think we can do it right
down to the individual level and I think that will really start to shift the dial
around the world in terms of what we can do around health and brilliant are there
any specific examples you want to bring I think sort of if I go back to the the
partnership model so sort of the humans and AI working together I think the
medical domain is a really strong example of that I mean computer
algorithms are very good at recognizing images they they’re able to train on on
more images than your average clinician or radiologist is going to see in the
whole of their lifetime that just lets them see much more but you don’t want to
just leave it to the to the algorithm you want to be able to work with it and
sort of that’s I think one of the one of the real big challenges is not just
getting better machine learning algorithms but understanding from both
sides of how partnerships with AI and humans work so what does it mean for an
AI to be a partner for human and for human to work with an AI system very
different to work into a standard computer system and I think work is
needed on both sides so it’s assisting doctors assisting surgeons assisting and
diagnosis him absolutely it’s a huge one terrors that an ST G you’d like to focus
on I think one of the most important challenges for this field industry
especially to come to see the human values like them and I think that
incentive structure doesn’t really currently exist right it’s something we
need to think really just give an example what do you mean by that well
the tech industry and frankly any for-profit company is motivated by
profit generation and we’ve seen an up ending of the
industry in recent years especially as coalitions of workers advocating for
certain values within those companies have pressed those values upon their
institutions and I think that it’s the start of a movement that you know we
will see more to come of that type of behavior likely but we really need to
make this a global public conversation and it needs to be again it needs to
span governments industry and other players but it it needs to be one that
really thinks about power structures and ultimately where empowerment is situated
what I think that the conversation should just be limited to places like
the UN it really needs to be thought about widely and by by the organizations
that are really in the driver because there are discussions even in the UK
about changing the companies act you know and an adapting it so that it
actually has an element of social responsibility rather than just
employees shareholders suppliers you know whether windfall causes a huge
conversation as we start see jobs change and some jobs go and obviously new jobs
come as to how how those decisions are going to get made I had a fascinating
interview the other day where one CEO was saying of a u.s. corporation they’re
actually not going as fast as they could with automation because they didn’t want
to disrupt the local community where their head office was in the u.s. I
think it said that’s an interesting point of view and it was really around
the call center and billing center that they run internal governance structures such that
they can make decisions like that based on again values about profit in some
circumstances and there are different ways of justifying that pure
shareholders but the end of the day to be able to say that you know you killed
a deal over a principle that your institutions upholding yep
fantastic sure the the goals the global goals where I know you’ve been spending
some time in this area as well so tell us I just want to respond to something
that that Tara said yeah so if you if you think about the profit motive that
that is one of these fixed objectives that in fact there are already super
intelligent AI systems that are pursuing these incorrect objectives and are
destroying the world that call corporations and so we should actually
think about that you know we really have created intelligent systems that they
are outwitting us right you wonder why can’t we get the climate right right why
can’t we you know fix co2 emissions and so on because they’re they have
outwitted us and we need to think hard about how we set up this optimization
problem so to come back to the SDGs yeah so Amir benefit Emmy who’s over there in
the front row and I and a few other people set up from something called the
AI Commons which is precisely about using AI shared data resources and so on
to address the SDGs and I’m very optimistic that we can we put
significant effort into these these kinds of applications we can make real
progress so famine prediction is one example detection of the beginnings of
deforestation illegal logging in the Amazon is another understanding the
dynamics of these growing cities in the global South where the local governments
actually have very little data about their own City they don’t even know
where the streets are they’re not centrally planned in the
same way that we traditionally do it in the West and when you think about all
these applications it’s not a mystery that you can use satellite data to to
solve a lot of these problems and prototypes for this have been developed
over decades there are hundreds and hundreds of
prototype projects showing that there are really cool things you can do with
satellite data and what’s the result of all those prototype projects a nice
feeling a report on a shelf and then a bunch of software that doesn’t work
anymore so what’s the problem the problem is that turning a prototype into
a real 24/7 global service something as available and as immediate and as global
as Google or Facebook or any other of these global platforms costs a great
deal of money probably on the order for satellite legacy or think about okay
ingesting all of the satellite data which is petabytes and information every
day storing it processing it and then making it available through some
interface this requires at least a billion dollar investment upfront and no
one project and no one NGO can’t possibly manage that just finance what
if we amortize it over hundreds of applications across all the SDGs then we
can do it so for example Planet which is right now largest launcher of global
observation satellites is working on precisely this global platform on which
you can then build applications for just a couple of million you could put
together a deforestation detector that’s 24/7 online all the time scanning the
entire globe for deforestation effort amazing amazing and I read that they put
a hundred and eighty satellites up more than anybody else in the last 12 months
or something it’s quite staggering so this would be a constant stream of
satellite imagery that other people could just tap into to create an
application may be as approach that to start with but actually could be a
production application running 24/7 and at a small marginal fantastic fantastic
now we actually have a couple of minutes left so I’m just gonna ask the panel if
there’s you know this panel is about the impact of a eye on industry government
of society and you know taps and I were talking audio you know there’s a very
positive and optimistic outlook at this conference although we do love to run at
the difficult challenges and the difficult questions and thank you for
those today what’s the most exciting use case that you have come across recently
that really has kind of stopped you and made you think you know that’s just an
awesome and magical use of AI or technology recently Tara do you want to
go first are you by all means take a minute I
think a lot of the healthcare applications are really really
extraordinary I mean that the number of clinical errors that result in
accidental deaths in hospitals for example it’s pretty astounding and that
is an issue that plagues a number of different countries and I think the
capability producing human error in the medical context diagnosis center I think
is is pretty amazing but that’s just the first one that comes to mind no words
Nick so I would have said that so let me say something else so I think some of
the things that are happening around education is really interesting at lots
of different ages I think the ability to to tailor learning and experience and
adaptive yeah advanced AI technology to to help and again give
education to two individuals in a way and a manner and a level that’s that
whole personalization and pacing of the obviou the coursework one of the main meta characteristics
that a I can bring the ability to tailor your education tailor your health tailor
your experiences to you and what you want to do is is for me one of the great
promises of AI and we have some great sessions actually about that coming up
with century tech and others Priya’s company talking about exactly
how they’re doing that today so I couldn’t agree more
Wendy my automated pod that takes me to the doctors or the pub when I am too old
to drive myself and I reckon that’s gonna be achievable just about I think
that will be fabulous levitating this may feel a long way how do I get
involved it’s all too scary or it’s too hard and picking up on what Nick said
actually the first thing I saw this weekend this weekend was somebody saying
this little online course we’ve got going on the data science and AI for
non-techies was really a life changer for this person I thought actually
everybody can get involved it is not something that is gonna be just for the
geeks or just for the pointy-headed and he’s gonna ruin all our lives
everybody can get involved because we can we can that’s my job skills champion
you gonna see a lot of programs rolling out in the UK UK where everyone can get
involved and come on Wednesday we’ve got the Girl Guides coming with it fantastic
fantastic it’s trying to give you the last word and we’re gonna wrap up all of
the things you’ve seen I think actually taking something from my talk an
application that could help the England team be better at taking penalties brilliant well thank you very much for
that thank you thank you to a fantastic panel Thank You Wendy thank you NIT
Thank You Tara Thank You Stuart you

Leave a Reply

Your email address will not be published. Required fields are marked *