machine learning also known as
artificial intelligence or AI for short
is spreading to every single part of our
lives but it’s only as good as the data
that it uses and the way that it learns
so about eight years ago I took a hazard
perception test in the UK this is part
of getting a driving license over there
the test is a series of short video
clips where you click when you see a
hazard coming up or the danger getting
worse I’m a video game designer and a
researcher and you know professional
game developers we’re pretty quick
seeing danger coming up on screens so
when I was taking my test for some
reason though when I got to the end of
it 2/3 of my clips had been disqualified
that was a little bit odd so I asked the
agency why this was and what they said
was we’re not saying you were cheating
but our algorithm found a strange
pattern in your clicks and so a lot of
your tests were disqualified I found
that slightly odd so when I mentioned
this around my office it turned out that
a lot of other game developers had
exactly the same problem the solution
was simple though what you did was he
watched out for when the danger was
coming and immediately you’d wait for
two seconds and then you would click so
why did that work well game developers
were clicking a little bit earlier than
that I I thought was normal and so they
we were all flagged as trying to cheat
this system we could only actually pass
through this system now by kind of
backwards engineering the code and the
data that it learned from it seemed a
little bit strange
now that test was using a very very
simple version of AI but it is one
example of how a government body is
using AI income every day societal
functions so many of us that probably
doesn’t feel like AI
and we use this kind of scientific
though sci-fi version of this which is
kind of sentient computers with full
personalities like howl in 2001 or the
replicants in Blade Runner or Janet in
the good life the good place even but
the reality of AI today is that it’s
really just a glamorous term for an
algorithm it’s just an equation it puts
data through human-made systems and it
comes out with patterns and results that
can answer very specific questions it
might be able to learn or grow a little
bit but that’s within set limits and it
doesn’t really have any intrinsic
consciousness or morals or ethics it’s
no more moral or ethical than a hammer
is we make a ice we are responsible for
the choices that they make now these AI
algorithms are incredibly useful we use
them in our everyday life all the time
and the walking directions on your phone
when you’re going somewhere that’s
powered by AI the filters on your emails
checking to see whether something spam
that’s all I I as well in your leisure
time your computer games are driven by
up by AI your social media feed is
driven by by AI and if you travel to see
someone face to face well there’s an AI
your bank which is looking at that
transaction to make sure it’s genuine
and not fraud the thing is though that
even though
AI is an equation it’s not free of human
bias it’s not entirely objective and I
think there are three steps that we can
take to improve how AI works for us in
the future to make it more responsible
towards society and when you think about
people with fast reactions being
disqualified from a reaction test for
reacting too fast we can see that the
way that AIS are built may have biases
in them so that is our first step we
must recognize that when a eyes are
created and when they spot patterns in
data that pattern is not the same as an
objective truth AI
is often accurate but we have to always
remember that it can be wrong because
the quality of results is only going to
be as good as the data that it’s using
and the code and the instructions that
it’s given to work with one of the
things that comes in here is that
minority groups whose needs are often
missed when AI is being made or a
particular risk from systemic errors now
I say that because AI is already
powering major societal problems facial
recognition systems can automatically
recognize people but they do have
numerous issues they’re particularly
poor at recognizing the faces of people
who are black and sadly due to social
and historical conditions people of
color generally are the most likely
people who facial recognition will be
used on by the police Elijah Cummings
from America’s House Oversight Committee
summarized this problem really nicely if
you’re black you’re more likely to be
subjected to this technology and the
the assumption that AI is always
accurate and always impartial Kamat
apply already deeply embedded social
prejudices in a different example in
2016 microsoft released an AI online and
it was going to be learning from
conversations with people within 16
hours of it going online it had said
gender equality equals feminism it had
so we can see that AI learning from
unfiltered data does not necessarily
produce useful results neutral results
or even logically consistent outcomes as
a last example algorithms on social
media are developed to try and keep us
engaged with their sites they are
individually choosing stories and
adverts that we each will find
compelling but I think we all know that
in the last few years we’ve seen
examples where half-truths or even lies
have become widely believed off they’ve
been spread on social media and a
made-up story
it sounds harmless enough but when
misinformation is changing the way we
interact with others or even voting in
elections and quite possibly already has
we can see that maybe AI is perhaps
contributing to undermining the very
foundations of our society so what’s the
root of all of this mathematics has no
intrinsic social viewpoints but the
datasets chosen to Train it a is the
code the instructions that are given by
human developers they do have biases
faithfully she’s a chief scientist for
Google and she’s a Stanford professor
and a couple of weeks ago she said this
be careful not be misled by the name
artificial intelligence
there is nothing artificial about it AI
is made by humans intended to behave by
humans and ultimately to impact humans
lives and human society human choices
can overlook data sources that gave
insights into different communities and
their life experiences outside the
knowledge of developers can usually with
no intentional malice limit the
usefulness of AI for all of society in
these limitations do impact on the
usefulness
of ai’s predictive ability because i is
our predictive systems they look at data
from the past and they use it to predict
the future now an AI may predict that
when a person lets go of a glass of
water it would fall to the floor and may
possibly smash now might predict that
because a programmer has told it that
the program might say this is what
happens every single time or it might
predict that it’s looked at thousands of
examples of a glass being released
falling to the floor and smashing and
then the AI says yes this is exactly
what happens when a person releases a
glass of water cool but we can see here
that either way the AI is going to be
correct
whether it’s told that or whether it
learnt it for itself this is great we
have a system that works but there are
problems here so if you imagine that I
am the AI developer and I’m standing
here in the spotlight and I’ve trained
my AI to predict this and I can rillette
go of this anytime and I will predict ly
correct correct predict correctly that
it falls to the floor you know this this
seems like it’s 100 percent correct it
seems like the objective it is
absolutely true and I might say this
will work for everyone if you let go of
this glass it will fall to the floor it
seems great but if I’m not the central
user of that AI if I’m just slightly to
the side of the spotlight the ai’s
prediction could be completely wrong
all it takes is standing a little bit
away from the center to reveal that a
glass can be released and not fall to
[Applause]
the prediction that a moment ago seemed
to be objective we probably all agreed
with is revealed to be subjective the
biases of the developer the sensuality
that they had and the data they trained
and tested it with were all limited to
make that prediction better all it
needed was adding a viewpoints away from
the mainstream so this gives us our
second step in how we can improve AI in
the future the growing business of AI is
a prime place for increasing the
diversity within the Technica tech
industry keep the people who are belong
who are already there that’s great we
have something about hiring and firing
and replacing things like that but grow
by bringing in more diverse communities
and true tans are equal in the
workplaces too because diverse
workforces they bring a new voice
v.points expert skills that have immense
value to the products and services that
we are trying to develop and provide and
of course this brings great social value
to the communities who are bringing tech
jobs into their sphere I think from
earliest education to adult careers
overcoming inequality is a good that
benefits all of us everybody in society
benefits for me it increased equality
both indirect ways and indirect ways so
for me I believe this is absolutely an
ethical duty but it is a business
necessity as well companies that speaks
the widest audiences in the most useful
dynamic and responsible ways will be the
ones with the greatest success in the AI
driven future we need everybody’s
viewpoint to make AI work and to drive
business as well so as I’ve mentioned
there are some risks in using AI and
there’s a thing I want you to remember
AI would be less dangerous for society
if it were only a predictive system the
thing is a is shaping the future too so
AI is not only predicting for us AI is
shaping us
we’ve made these algorithms and whether
it’s getting a driving license or
applying for a loan or getting arrested
based on facial recognition systems
decisions made about our future are
being based on the results of AI because
we get all our data from the past and
yes we have to get our data from
somewhere but the past was not
necessarily an equally golden age for
everyone we don’t necessarily want to
replicate that past into the future the
social risks of AI and its misuse I
think are clear I hope but I come from
the video game industry where it would
appear that AI is going to be completely
harmless because in my industry we’re
continually using AI to make our games
more engaging in the game no man’s sky
algorithms created literally billions of
creatures and worlds so that every
single player could find a new world or
species awesome in Skyrim
computer-controlled characters each had
their own priorities their own goals
which were aside from the player’s quest
so when the player met them the world
felt more alive more exciting great fun
to explore because of AI but we also
have examples where the data driving our
so when Pokemon go was released in 2016
an unintended pattern was noticed in its
data if you don’t know Pokemon go it’s a
smartphone game where you go to physical
locations in the world and when you get
there you can collect it to animals in
your phone so if you live near to these
locations then you’re less likely to
have to pay money to progress in the
game and your real world travel costs to
get to those locations is going to be
less that’s okay right the problem was
the the data used to choose these
locations was based on the most popular
locations in a previous game and that
previous game was largely played by
affluent white city dwellers
and so Pokemon go players from
marginalized rural or poor neighborhoods
were financially disadvantaged if they
wanted to get the same experience as
people who lived in richer urban
neighborhoods now this affected many
different groups the one group which was
disproportionately affected was people
of color again completely
unintentionally the developers had used
data from the past that reinforced
systemic racism that chains privilege to
skin color now I don’t believe for a
second not for a moment there was any
intentional racism within that team
if those game developers are like any of
the ones that I know they would be
horrified by that idea but the data
drive in the game was not neutral it was
not objective in how it treated it all
people social divisions from the past
men are the players with the least
privilege in the present were being
penalized again and that’s going to
stepping slightly to the side of video
game examples do women program computers
I would hope that the answer is obvious
recruitment agencies looking for a top
go-getting programmer might have
hundreds of applications to filter and
these days at their AI is that’s too
quite likely to be doing the first
filter for them so if you look at the
historic data about the last couple of
decades of who are the top programmers
you might say well men are going to have
the best career in programming and so
that a I might just disregard
applications from women and people of
other genders without being told to do
this by humans an AI could be
reproducing gender bias
this isn’t AI predicting good future
employees this is AI shaping the future
workforce countering this is going to
need awareness and it’s going to need
deliberate action by ai’s developers so
that’s going to be our third step to
improve AR in the future we need to
recognize the data current contains
Myriad’s
sample biases we need to recognize that
it does not automatically become
socially morally ethically or
economically neutral if necessary put
data ethicists at the core of your team
to make sure you’re treating your data
and your algorithms in ways that means
the whole of society is treated equally
I think this is the way that we can help
ensure that we get the best out of AI
for everybody because artificial
intelligence it is a fantastic tool
I think we’re currently too accepting of
it neutrality we’re at the beginning of
our journey with it and I believe
there’s reactions I’ve discussed today
will help us make AI better for
everybody and use it more responsibly in
the future so those ideas again are
basically we must recognize that the
patterns found by AI and not the same as
unquestionable truth we must diversify
the tech industry to get the best
products and services in their AI driven
future and thirdly we need to
consciously engage with the data and –
rather than just assuming that it’s
neutral an even big data can miss the
most marginalized of groups and that
mistake that can multiply across all of
social strata to create imbalances in
society now as I said at the beginning
iOS decisions are constrained by their
creators and the data that we train them
with social divisions that are based on
race gender ethnicity wealth and other
personal factors so many personal
factors all the reality of our past and
they are the reality of present today
and I think these dangers could be
amplified by AI not overcome AI reflects
us good and bad a digital technologies
do not automatically fix the problematic
legacies of the past AI could help us
fix this but we will only do so if we
take conscious action to make it work
for everyone thank you all for listening [Applause]
artificial intelligence or AI for short
is spreading to every single part of our
lives but it’s only as good as the data
that it uses and the way that it learns
so about eight years ago I took a hazard
perception test in the UK this is part
of getting a driving license over there
the test is a series of short video
clips where you click when you see a
hazard coming up or the danger getting
worse I’m a video game designer and a
researcher and you know professional
game developers we’re pretty quick
seeing danger coming up on screens so
when I was taking my test for some
reason though when I got to the end of
it 2/3 of my clips had been disqualified
that was a little bit odd so I asked the
agency why this was and what they said
was we’re not saying you were cheating
but our algorithm found a strange
pattern in your clicks and so a lot of
your tests were disqualified I found
that slightly odd so when I mentioned
this around my office it turned out that
a lot of other game developers had
exactly the same problem the solution
was simple though what you did was he
watched out for when the danger was
coming and immediately you’d wait for
two seconds and then you would click so
why did that work well game developers
were clicking a little bit earlier than
that I I thought was normal and so they
we were all flagged as trying to cheat
this system we could only actually pass
through this system now by kind of
backwards engineering the code and the
data that it learned from it seemed a
little bit strange
now that test was using a very very
simple version of AI but it is one
example of how a government body is
using AI income every day societal
functions so many of us that probably
doesn’t feel like AI
and we use this kind of scientific
though sci-fi version of this which is
kind of sentient computers with full
personalities like howl in 2001 or the
replicants in Blade Runner or Janet in
the good life the good place even but
the reality of AI today is that it’s
really just a glamorous term for an
algorithm it’s just an equation it puts
data through human-made systems and it
comes out with patterns and results that
can answer very specific questions it
might be able to learn or grow a little
bit but that’s within set limits and it
doesn’t really have any intrinsic
consciousness or morals or ethics it’s
no more moral or ethical than a hammer
is we make a ice we are responsible for
the choices that they make now these AI
algorithms are incredibly useful we use
them in our everyday life all the time
and the walking directions on your phone
when you’re going somewhere that’s
powered by AI the filters on your emails
checking to see whether something spam
that’s all I I as well in your leisure
time your computer games are driven by
up by AI your social media feed is
driven by by AI and if you travel to see
someone face to face well there’s an AI
your bank which is looking at that
transaction to make sure it’s genuine
and not fraud the thing is though that
even though
AI is an equation it’s not free of human
bias it’s not entirely objective and I
think there are three steps that we can
take to improve how AI works for us in
the future to make it more responsible
towards society and when you think about
people with fast reactions being
disqualified from a reaction test for
reacting too fast we can see that the
way that AIS are built may have biases
in them so that is our first step we
must recognize that when a eyes are
created and when they spot patterns in
data that pattern is not the same as an
objective truth AI
is often accurate but we have to always
remember that it can be wrong because
the quality of results is only going to
be as good as the data that it’s using
and the code and the instructions that
it’s given to work with one of the
things that comes in here is that
minority groups whose needs are often
missed when AI is being made or a
particular risk from systemic errors now
I say that because AI is already
powering major societal problems facial
recognition systems can automatically
recognize people but they do have
numerous issues they’re particularly
poor at recognizing the faces of people
who are black and sadly due to social
and historical conditions people of
color generally are the most likely
people who facial recognition will be
used on by the police Elijah Cummings
from America’s House Oversight Committee
summarized this problem really nicely if
you’re black you’re more likely to be
subjected to this technology and the
the assumption that AI is always
accurate and always impartial Kamat
apply already deeply embedded social
prejudices in a different example in
2016 microsoft released an AI online and
it was going to be learning from
conversations with people within 16
hours of it going online it had said
gender equality equals feminism it had
so we can see that AI learning from
unfiltered data does not necessarily
produce useful results neutral results
or even logically consistent outcomes as
a last example algorithms on social
media are developed to try and keep us
engaged with their sites they are
individually choosing stories and
adverts that we each will find
compelling but I think we all know that
in the last few years we’ve seen
examples where half-truths or even lies
have become widely believed off they’ve
been spread on social media and a
made-up story
it sounds harmless enough but when
misinformation is changing the way we
interact with others or even voting in
elections and quite possibly already has
we can see that maybe AI is perhaps
contributing to undermining the very
foundations of our society so what’s the
root of all of this mathematics has no
intrinsic social viewpoints but the
datasets chosen to Train it a is the
code the instructions that are given by
human developers they do have biases
faithfully she’s a chief scientist for
Google and she’s a Stanford professor
and a couple of weeks ago she said this
be careful not be misled by the name
artificial intelligence
there is nothing artificial about it AI
is made by humans intended to behave by
humans and ultimately to impact humans
lives and human society human choices
can overlook data sources that gave
insights into different communities and
their life experiences outside the
knowledge of developers can usually with
no intentional malice limit the
usefulness of AI for all of society in
these limitations do impact on the
usefulness
of ai’s predictive ability because i is
our predictive systems they look at data
from the past and they use it to predict
the future now an AI may predict that
when a person lets go of a glass of
water it would fall to the floor and may
possibly smash now might predict that
because a programmer has told it that
the program might say this is what
happens every single time or it might
predict that it’s looked at thousands of
examples of a glass being released
falling to the floor and smashing and
then the AI says yes this is exactly
what happens when a person releases a
glass of water cool but we can see here
that either way the AI is going to be
correct
whether it’s told that or whether it
learnt it for itself this is great we
have a system that works but there are
problems here so if you imagine that I
am the AI developer and I’m standing
here in the spotlight and I’ve trained
my AI to predict this and I can rillette
go of this anytime and I will predict ly
correct correct predict correctly that
it falls to the floor you know this this
seems like it’s 100 percent correct it
seems like the objective it is
absolutely true and I might say this
will work for everyone if you let go of
this glass it will fall to the floor it
seems great but if I’m not the central
user of that AI if I’m just slightly to
the side of the spotlight the ai’s
prediction could be completely wrong
all it takes is standing a little bit
away from the center to reveal that a
glass can be released and not fall to
[Applause]
the prediction that a moment ago seemed
to be objective we probably all agreed
with is revealed to be subjective the
biases of the developer the sensuality
that they had and the data they trained
and tested it with were all limited to
make that prediction better all it
needed was adding a viewpoints away from
the mainstream so this gives us our
second step in how we can improve AI in
the future the growing business of AI is
a prime place for increasing the
diversity within the Technica tech
industry keep the people who are belong
who are already there that’s great we
have something about hiring and firing
and replacing things like that but grow
by bringing in more diverse communities
and true tans are equal in the
workplaces too because diverse
workforces they bring a new voice
v.points expert skills that have immense
value to the products and services that
we are trying to develop and provide and
of course this brings great social value
to the communities who are bringing tech
jobs into their sphere I think from
earliest education to adult careers
overcoming inequality is a good that
benefits all of us everybody in society
benefits for me it increased equality
both indirect ways and indirect ways so
for me I believe this is absolutely an
ethical duty but it is a business
necessity as well companies that speaks
the widest audiences in the most useful
dynamic and responsible ways will be the
ones with the greatest success in the AI
driven future we need everybody’s
viewpoint to make AI work and to drive
business as well so as I’ve mentioned
there are some risks in using AI and
there’s a thing I want you to remember
AI would be less dangerous for society
if it were only a predictive system the
thing is a is shaping the future too so
AI is not only predicting for us AI is
shaping us
we’ve made these algorithms and whether
it’s getting a driving license or
applying for a loan or getting arrested
based on facial recognition systems
decisions made about our future are
being based on the results of AI because
we get all our data from the past and
yes we have to get our data from
somewhere but the past was not
necessarily an equally golden age for
everyone we don’t necessarily want to
replicate that past into the future the
social risks of AI and its misuse I
think are clear I hope but I come from
the video game industry where it would
appear that AI is going to be completely
harmless because in my industry we’re
continually using AI to make our games
more engaging in the game no man’s sky
algorithms created literally billions of
creatures and worlds so that every
single player could find a new world or
species awesome in Skyrim
computer-controlled characters each had
their own priorities their own goals
which were aside from the player’s quest
so when the player met them the world
felt more alive more exciting great fun
to explore because of AI but we also
have examples where the data driving our
so when Pokemon go was released in 2016
an unintended pattern was noticed in its
data if you don’t know Pokemon go it’s a
smartphone game where you go to physical
locations in the world and when you get
there you can collect it to animals in
your phone so if you live near to these
locations then you’re less likely to
have to pay money to progress in the
game and your real world travel costs to
get to those locations is going to be
less that’s okay right the problem was
the the data used to choose these
locations was based on the most popular
locations in a previous game and that
previous game was largely played by
affluent white city dwellers
and so Pokemon go players from
marginalized rural or poor neighborhoods
were financially disadvantaged if they
wanted to get the same experience as
people who lived in richer urban
neighborhoods now this affected many
different groups the one group which was
disproportionately affected was people
of color again completely
unintentionally the developers had used
data from the past that reinforced
systemic racism that chains privilege to
skin color now I don’t believe for a
second not for a moment there was any
intentional racism within that team
if those game developers are like any of
the ones that I know they would be
horrified by that idea but the data
drive in the game was not neutral it was
not objective in how it treated it all
people social divisions from the past
men are the players with the least
privilege in the present were being
penalized again and that’s going to
stepping slightly to the side of video
game examples do women program computers
I would hope that the answer is obvious
recruitment agencies looking for a top
go-getting programmer might have
hundreds of applications to filter and
these days at their AI is that’s too
quite likely to be doing the first
filter for them so if you look at the
historic data about the last couple of
decades of who are the top programmers
you might say well men are going to have
the best career in programming and so
that a I might just disregard
applications from women and people of
other genders without being told to do
this by humans an AI could be
reproducing gender bias
this isn’t AI predicting good future
employees this is AI shaping the future
workforce countering this is going to
need awareness and it’s going to need
deliberate action by ai’s developers so
that’s going to be our third step to
improve AR in the future we need to
recognize the data current contains
Myriad’s
sample biases we need to recognize that
it does not automatically become
socially morally ethically or
economically neutral if necessary put
data ethicists at the core of your team
to make sure you’re treating your data
and your algorithms in ways that means
the whole of society is treated equally
I think this is the way that we can help
ensure that we get the best out of AI
for everybody because artificial
intelligence it is a fantastic tool
I think we’re currently too accepting of
it neutrality we’re at the beginning of
our journey with it and I believe
there’s reactions I’ve discussed today
will help us make AI better for
everybody and use it more responsibly in
the future so those ideas again are
basically we must recognize that the
patterns found by AI and not the same as
unquestionable truth we must diversify
the tech industry to get the best
products and services in their AI driven
future and thirdly we need to
consciously engage with the data and –
rather than just assuming that it’s
neutral an even big data can miss the
most marginalized of groups and that
mistake that can multiply across all of
social strata to create imbalances in
society now as I said at the beginning
iOS decisions are constrained by their
creators and the data that we train them
with social divisions that are based on
race gender ethnicity wealth and other
personal factors so many personal
factors all the reality of our past and
they are the reality of present today
and I think these dangers could be
amplified by AI not overcome AI reflects
us good and bad a digital technologies
do not automatically fix the problematic
legacies of the past AI could help us
fix this but we will only do so if we
take conscious action to make it work
for everyone thank you all for listening [Applause]
Please follow and like us: