so when I was a teenager growing up in Nigeria
Nigeria
Nigeria I was not directing music videos I did
I was not directing music videos I did
I was not directing music videos I did the style however to build an artificial
the style however to build an artificial
the style however to build an artificial intelligence I know I know
intelligence I know I know
intelligence I know I know I was taking computer science and logic
I was taking computer science and logic
I was taking computer science and logic courses at local university and I
courses at local university and I
courses at local university and I figured it would be easy being a 15 year
figured it would be easy being a 15 year
figured it would be easy being a 15 year old to just create an AI that could
old to just create an AI that could
old to just create an AI that could reason and argue logically I was
reason and argue logically I was
reason and argue logically I was inspired by the same dreams the early
inspired by the same dreams the early
inspired by the same dreams the early pioneers of artificial intelligence had
pioneers of artificial intelligence had
pioneers of artificial intelligence had that we could we could create this super
that we could we could create this super
that we could we could create this super mind and intelligence of pure logic and
mind and intelligence of pure logic and
mind and intelligence of pure logic and objectivity and that this super mind
objectivity and that this super mind
objectivity and that this super mind would be free of human traits traits
would be free of human traits traits
would be free of human traits traits like subjectivity and bias I was working
like subjectivity and bias I was working
like subjectivity and bias I was working with our guidance with really really old
with our guidance with really really old
with our guidance with really really old computers and with irregular electricity
computers and with irregular electricity
computers and with irregular electricity if you’ve been to Nigeria you might know
if you’ve been to Nigeria you might know
if you’ve been to Nigeria you might know what that’s about
what that’s about
what that’s about so let’s just say I was less than
so let’s just say I was less than
so let’s just say I was less than successful I did however go on to study
successful I did however go on to study
successful I did however go on to study in the United States and do a lot more
in the United States and do a lot more
in the United States and do a lot more research as an engineer on artificial
research as an engineer on artificial
research as an engineer on artificial intelligence around 2013 I was reminded
intelligence around 2013 I was reminded
intelligence around 2013 I was reminded of my dreams of objective AI when I
of my dreams of objective AI when I
of my dreams of objective AI when I heard of a Harvard professors
heard of a Harvard professors
heard of a Harvard professors experiments
experiments
experiments Latanya Sweeney at the University of
Latanya Sweeney at the University of
Latanya Sweeney at the University of Harvard University was exploring how
Harvard University was exploring how
Harvard University was exploring how search engine had change when you look
search engine had change when you look
search engine had change when you look up different types of names when she
up different types of names when she
up different types of names when she looked up names more predictive of being
looked up names more predictive of being
looked up names more predictive of being black she got ads for criminal justice
black she got ads for criminal justice
black she got ads for criminal justice services
so she’d look up names like DeSean or
so she’d look up names like DeSean or even Latonya and she get ads like has
even Latonya and she get ads like has
even Latonya and she get ads like has DeSean been arrested it’s not funny but
DeSean been arrested it’s not funny but
DeSean been arrested it’s not funny but he’s kind of funny do you need a bail
he’s kind of funny do you need a bail
he’s kind of funny do you need a bail bondsman when she looked up names less
bondsman when she looked up names less
bondsman when she looked up names less predictive of being black so names like
predictive of being black so names like
predictive of being black so names like Jeffrey or Emma not so much eventually
Jeffrey or Emma not so much eventually
Jeffrey or Emma not so much eventually she calculated that black identify names
she calculated that black identify names
she calculated that black identify names worth 25% more likely to generate ad
worth 25% more likely to generate ad
worth 25% more likely to generate ad suggestion that the person had an arrest
suggestion that the person had an arrest
suggestion that the person had an arrest record even when they did not but these
record even when they did not but these
record even when they did not but these are algorithms serving up ads can
are algorithms serving up ads can
are algorithms serving up ads can algorithms be prejudiced can a piece of
algorithms be prejudiced can a piece of
algorithms be prejudiced can a piece of software be racist can it be sexist
learning algorithms are recipes for
learning algorithms are recipes for teaching AI how to learn modern AI
teaching AI how to learn modern AI
teaching AI how to learn modern AI learns by consuming vast amounts of data
learns by consuming vast amounts of data
learns by consuming vast amounts of data they learn by spotting hidden
they learn by spotting hidden
they learn by spotting hidden correlations and patterns that humans
correlations and patterns that humans
correlations and patterns that humans cannot see they learn by applying
cannot see they learn by applying
cannot see they learn by applying neutral math equations but the more
neutral math equations but the more
neutral math equations but the more researchers have looked the more they
researchers have looked the more they
researchers have looked the more they found examples of artificial
found examples of artificial
found examples of artificial intelligence systems producing
intelligence systems producing
intelligence systems producing significantly biased outcomes and by
significantly biased outcomes and by
significantly biased outcomes and by bias we mean generally violations of
bias we mean generally violations of
bias we mean generally violations of ethical or social norms for example a
ethical or social norms for example a
ethical or social norms for example a recent study found that search engines
recent study found that search engines
recent study found that search engines show more ads for higher paying jobs to
show more ads for higher paying jobs to
show more ads for higher paying jobs to men than to women they’ve also been
men than to women they’ve also been
men than to women they’ve also been there was a recent study last year
there was a recent study last year
there was a recent study last year showing extensive bias in algorithms
showing extensive bias in algorithms
showing extensive bias in algorithms used in the criminal justice system so
used in the criminal justice system so
used in the criminal justice system so just using AI does not eliminate bias in
just using AI does not eliminate bias in
just using AI does not eliminate bias in decision-making but how exactly do these
decision-making but how exactly do these
decision-making but how exactly do these biases get into our algorithms in the
biases get into our algorithms in the
biases get into our algorithms in the first place modern AI is only as good as
first place modern AI is only as good as
first place modern AI is only as good as the data on which is trained AI eat
the data on which is trained AI eat
the data on which is trained AI eat and processes data and use the patterns
and processes data and use the patterns
and processes data and use the patterns it finds in the data to make future
it finds in the data to make future
it finds in the data to make future decisions now these decisions could be
decisions now these decisions could be
decisions now these decisions could be as trivial as distinguishing between
as trivial as distinguishing between
as trivial as distinguishing between cats versus dogs or they could be
cats versus dogs or they could be
cats versus dogs or they could be extremely important an insurance company
extremely important an insurance company
extremely important an insurance company might use artificial intelligence to
might use artificial intelligence to
might use artificial intelligence to help you decide who is insurable versus
help you decide who is insurable versus
help you decide who is insurable versus who is not insurable consider the case
who is not insurable consider the case
who is not insurable consider the case of an artificial intelligence trying to
of an artificial intelligence trying to
of an artificial intelligence trying to find the best nurses if we decide to
find the best nurses if we decide to
find the best nurses if we decide to feed this system stereotypical training
feed this system stereotypical training
feed this system stereotypical training data so data consisting mostly of female
data so data consisting mostly of female
data so data consisting mostly of female nursing profiles then chances are the
nursing profiles then chances are the
nursing profiles then chances are the system would more competently judge
system would more competently judge
system would more competently judge future female candidates as better fits
future female candidates as better fits
future female candidates as better fits unless it’s carefully designed not to do
unless it’s carefully designed not to do
unless it’s carefully designed not to do so the system isn’t being inherently
so the system isn’t being inherently
so the system isn’t being inherently sexist it’s simply learning the biases
sexist it’s simply learning the biases
sexist it’s simply learning the biases present in our data and applying them
present in our data and applying them
present in our data and applying them more consistently into the future now
more consistently into the future now
more consistently into the future now occasionally AI develops a case of what
occasionally AI develops a case of what
occasionally AI develops a case of what we might call our food poisoning back
we might call our food poisoning back
we might call our food poisoning back when IBM was trading up the future AI
when IBM was trading up the future AI
when IBM was trading up the future AI jeopardy champion that’s Watson they had
jeopardy champion that’s Watson they had
jeopardy champion that’s Watson they had to feed Watson vast amounts of data and
to feed Watson vast amounts of data and
to feed Watson vast amounts of data and then somebody decided it might be a good
then somebody decided it might be a good
then somebody decided it might be a good idea to feed Watson the urban dictionary
so after ingesting all that it’s just a
so after ingesting all that it’s just a colorful data Watson developed a little
colorful data Watson developed a little
colorful data Watson developed a little bit of a swearing habit
he began in inserting our weighted four
he began in inserting our weighted four letter words in his responses now
letter words in his responses now
letter words in his responses now personally I would have loved to watch
personally I would have loved to watch
personally I would have loved to watch that version of Watson on Jeopardy but
that version of Watson on Jeopardy but
that version of Watson on Jeopardy but the moral here is we need to be careful
the moral here is we need to be careful
the moral here is we need to be careful what we feed our algorithms on our AI
what we feed our algorithms on our AI
what we feed our algorithms on our AI and if n is AI is not going away
and if n is AI is not going away
and if n is AI is not going away AI already determined is what you see on
AI already determined is what you see on
AI already determined is what you see on your Twitter feed
your Twitter feed
your Twitter feed what you might want to watch on your own
what you might want to watch on your own
what you might want to watch on your own on Netflix who you might want to date
on Netflix who you might want to date
on Netflix who you might want to date and it’s going to keep making more and
and it’s going to keep making more and
and it’s going to keep making more and more of these decisions in the future
more of these decisions in the future
more of these decisions in the future because the amount of data we have and
because the amount of data we have and
because the amount of data we have and we create is vast and it just keeps
we create is vast and it just keeps
we create is vast and it just keeps growing artificial intelligence and
growing artificial intelligence and
growing artificial intelligence and algorithms present the only viable way
algorithms present the only viable way
algorithms present the only viable way of making sense of this much data so if
of making sense of this much data so if
of making sense of this much data so if you want better medical diagnosis if you
you want better medical diagnosis if you
you want better medical diagnosis if you want fewer car crashes or if you want to
want fewer car crashes or if you want to
want fewer car crashes or if you want to improve the quality of human lives and
improve the quality of human lives and
improve the quality of human lives and prevent needless suffering more
prevent needless suffering more
prevent needless suffering more generally you’re going to need AI
generally you’re going to need AI
generally you’re going to need AI systems that can learn quickly tackle
systems that can learn quickly tackle
systems that can learn quickly tackle complicated problems and make or inform
complicated problems and make or inform
complicated problems and make or inform hard decisions but we also want these
hard decisions but we also want these
hard decisions but we also want these systems to play fair so it’s great they
systems to play fair so it’s great they
systems to play fair so it’s great they begin to get behind or at least discard
begin to get behind or at least discard
begin to get behind or at least discard this illusion that AI modern-era is
this illusion that AI modern-era is
this illusion that AI modern-era is objective or infallible especially now
objective or infallible especially now
objective or infallible especially now as AI is ascendant and we can all get
as AI is ascendant and we can all get
as AI is ascendant and we can all get behind this idea that we do not want AI
behind this idea that we do not want AI
behind this idea that we do not want AI to make decisions that run contrary to
to make decisions that run contrary to
to make decisions that run contrary to our values but how exactly do we teach
our values but how exactly do we teach
our values but how exactly do we teach artificial intelligence to abide by
artificial intelligence to abide by
artificial intelligence to abide by social or ethical norms and who gets to
social or ethical norms and who gets to
social or ethical norms and who gets to decide what norms who gets to decide
decide what norms who gets to decide
decide what norms who gets to decide what is fair to me these are fascinating
what is fair to me these are fascinating
what is fair to me these are fascinating questions and researchers have started
questions and researchers have started
questions and researchers have started looking at a few approaches to tackling
looking at a few approaches to tackling
looking at a few approaches to tackling these questions so first we could
these questions so first we could
these questions so first we could disclose to consumers when AI is making
disclose to consumers when AI is making
disclose to consumers when AI is making decisions that affect them
decisions that affect them
decisions that affect them top-down government regulation may not
top-down government regulation may not
top-down government regulation may not be desirable or even feasible or
be desirable or even feasible or
be desirable or even feasible or effective
effective
effective as a way to tackle in biasing algorithms
as a way to tackle in biasing algorithms
as a way to tackle in biasing algorithms making consumers aware is an easy first
making consumers aware is an easy first
making consumers aware is an easy first step this also makes us more aware of a
step this also makes us more aware of a
step this also makes us more aware of a thousand little ways in which artificial
thousand little ways in which artificial
thousand little ways in which artificial intelligence affects our lives and more
intelligence affects our lives and more
intelligence affects our lives and more aware of our choices as consumers of
aware of our choices as consumers of
aware of our choices as consumers of services that rely on automated
services that rely on automated
services that rely on automated decisions second we could create
decisions second we could create
decisions second we could create processes for consumers to appeal
processes for consumers to appeal
processes for consumers to appeal automated decisions they’d be appealing
automated decisions they’d be appealing
automated decisions they’d be appealing to humans of course third we could have
to humans of course third we could have
to humans of course third we could have more human oversight when artificial
more human oversight when artificial
more human oversight when artificial intelligence is making decisions in
intelligence is making decisions in
intelligence is making decisions in high-risk domains this includes domains
high-risk domains this includes domains
high-risk domains this includes domains like defense or the criminal justice
like defense or the criminal justice
like defense or the criminal justice system these are being policy solutions
system these are being policy solutions
system these are being policy solutions so far there are a couple of technical
so far there are a couple of technical
so far there are a couple of technical fixes in the works for example we’ve
fixes in the works for example we’ve
fixes in the works for example we’ve been exploring ways of making AI models
been exploring ways of making AI models
been exploring ways of making AI models more transparent more observable other
more transparent more observable other
more transparent more observable other researchers have been exploring
researchers have been exploring
researchers have been exploring exploring ways of making AI models
exploring ways of making AI models
exploring ways of making AI models explain yet decisions the idea behind
explain yet decisions the idea behind
explain yet decisions the idea behind these approaches these technical
these approaches these technical
these approaches these technical approaches is to create a trail of
approaches is to create a trail of
approaches is to create a trail of breadcrumbs so that we can follow the
breadcrumbs so that we can follow the
breadcrumbs so that we can follow the trail of omi called intermediate
trail of omi called intermediate
trail of omi called intermediate inferences forward from the inputs all
inferences forward from the inputs all
inferences forward from the inputs all the way through to the final decisions
the way through to the final decisions
the way through to the final decisions basically show your work or we can
basically show your work or we can
basically show your work or we can follow the decisions the trail backward
follow the decisions the trail backward
follow the decisions the trail backward from the decisions all the way down to
from the decisions all the way down to
from the decisions all the way down to underlying causes or reasons these
underlying causes or reasons these
underlying causes or reasons these solutions they help us see where AI bias
solutions they help us see where AI bias
solutions they help us see where AI bias exists but they don’t always fix them
exists but they don’t always fix them
exists but they don’t always fix them and more fundamentally don’t get the
and more fundamentally don’t get the
and more fundamentally don’t get the more fundamental question of what
more fundamental question of what
more fundamental question of what qualifies as bias in specific domains as
qualifies as bias in specific domains as
qualifies as bias in specific domains as an engineer I fully own my tendency to
an engineer I fully own my tendency to
an engineer I fully own my tendency to think of or focus on technical
think of or focus on technical
think of or focus on technical virtuosity as the way to tackle and
virtuosity as the way to tackle and
virtuosity as the way to tackle and solve problems just as I try to do as a
solve problems just as I try to do as a
solve problems just as I try to do as a teenager all the way back then
teenager all the way back then
teenager all the way back then but I’m not convinced that we can just
but I’m not convinced that we can just
but I’m not convinced that we can just engineer our way out of this particular
engineer our way out of this particular
engineer our way out of this particular problem because the problem is less
problem because the problem is less
problem because the problem is less about the technology AI it’s more about
about the technology AI it’s more about
about the technology AI it’s more about all the social and cultural context in
all the social and cultural context in
all the social and cultural context in which we apply the technology at the
which we apply the technology at the
which we apply the technology at the moment we have AI been applied in fields
moment we have AI been applied in fields
moment we have AI been applied in fields as diverse as medicine law criminology
as diverse as medicine law criminology
as diverse as medicine law criminology and even education each of these fields
and even education each of these fields
and even education each of these fields have different norms different ethics an
have different norms different ethics an
have different norms different ethics an AI is like a very smart child it’s not
AI is like a very smart child it’s not
AI is like a very smart child it’s not always clear on the context of his
always clear on the context of his
always clear on the context of his decisions for example in many state it
decisions for example in many state it
decisions for example in many state it is perfectly legal to discriminate on
is perfectly legal to discriminate on
is perfectly legal to discriminate on the basis of demographic characteristics
the basis of demographic characteristics
the basis of demographic characteristics when you’re setting auto insurance rates
when you’re setting auto insurance rates
when you’re setting auto insurance rates young and male drivers tend to be
young and male drivers tend to be
young and male drivers tend to be riskier to insure at least so I’m told
riskier to insure at least so I’m told
riskier to insure at least so I’m told but the ability to calibrate this kind
but the ability to calibrate this kind
but the ability to calibrate this kind of risk is crucial for insurance on the
of risk is crucial for insurance on the
of risk is crucial for insurance on the other hand it’s illegal to discriminate
other hand it’s illegal to discriminate
other hand it’s illegal to discriminate on the basis of demographic
on the basis of demographic
on the basis of demographic characteristics when you’re issuing
characteristics when you’re issuing
characteristics when you’re issuing mortgages context determines the
mortgages context determines the
mortgages context determines the relevant norms and engineers cannot
relevant norms and engineers cannot
relevant norms and engineers cannot always teach AI how it ought to behave
always teach AI how it ought to behave
always teach AI how it ought to behave in every single application we need to
in every single application we need to
in every single application we need to work closer with social scientists so
work closer with social scientists so
work closer with social scientists so lawyers economists anthropologist
lawyers economists anthropologist
lawyers economists anthropologist linguist and I can tell you from
linguist and I can tell you from
linguist and I can tell you from personal experience working across these
personal experience working across these
personal experience working across these different fields is difficult we train
different fields is difficult we train
different fields is difficult we train differently and this is not a joke we
differently and this is not a joke we
differently and this is not a joke we really do speak different languages at
really do speak different languages at
really do speak different languages at my day job at the rank corporation we
my day job at the rank corporation we
my day job at the rank corporation we try to use as an example we try to use
try to use as an example we try to use
try to use as an example we try to use artificial intelligence models to study
artificial intelligence models to study
artificial intelligence models to study behaviors on social networks this
behaviors on social networks this
behaviors on social networks this includes behaviors like radicalization
includes behaviors like radicalization
includes behaviors like radicalization political polarization and even
political polarization and even
political polarization and even cyberbullying what you might consider
cyberbullying what you might consider
cyberbullying what you might consider fair or normative in any of those
fair or normative in any of those
fair or normative in any of those applications will depend on what you’re
applications will depend on what you’re
applications will depend on what you’re focused on if you focused on
focused on if you focused on
focused on if you focused on radicalization then you might not think
radicalization then you might not think
radicalization then you might not think it’s fair that we treat every
it’s fair that we treat every
it’s fair that we treat every opinion expressed on social media
opinion expressed on social media
opinion expressed on social media exactly alike here’s the crux of the
exactly alike here’s the crux of the
exactly alike here’s the crux of the problem we have created artificial
problem we have created artificial
problem we have created artificial intelligence in our own image and this
intelligence in our own image and this
intelligence in our own image and this is not a compliment
is not a compliment
is not a compliment AI to easily reflect human shortcomings
AI to easily reflect human shortcomings
AI to easily reflect human shortcomings it finds in our data streams it
it finds in our data streams it
it finds in our data streams it amplifies our flaws it’s going to take
amplifies our flaws it’s going to take
amplifies our flaws it’s going to take the full range of human intelligence to
the full range of human intelligence to
the full range of human intelligence to correct for these for the shortcomings
correct for these for the shortcomings
correct for these for the shortcomings but if you want to build better fairer
but if you want to build better fairer
but if you want to build better fairer societies we need AI systems that
societies we need AI systems that
societies we need AI systems that reflect and amplify the better parts of
reflect and amplify the better parts of
reflect and amplify the better parts of our nature thank you
”
Be First to Comment