study after study has found overwhelming disparities and who is impacted by the
disparities and who is impacted by the
disparities and who is impacted by the American criminal justice system perhaps
American criminal justice system perhaps
American criminal justice system perhaps you’ve heard some of these black men are
you’ve heard some of these black men are
you’ve heard some of these black men are six times more likely to be incarcerated
six times more likely to be incarcerated
six times more likely to be incarcerated than white men on the road minorities
than white men on the road minorities
than white men on the road minorities are more likely to be stopped and
are more likely to be stopped and
are more likely to be stopped and searched than whites but are these
searched than whites but are these
searched than whites but are these disparities evidence of racial bias we
disparities evidence of racial bias we
disparities evidence of racial bias we often turn to data for a semblance of
often turn to data for a semblance of
often turn to data for a semblance of truth for rigor objectivity I love data
truth for rigor objectivity I love data
truth for rigor objectivity I love data it can reveal patterns tell stories and
it can reveal patterns tell stories and
it can reveal patterns tell stories and at their very core data can confirm or
at their very core data can confirm or
at their very core data can confirm or challenger theories but when it comes to
challenger theories but when it comes to
challenger theories but when it comes to racial bias definitive answers have been
racial bias definitive answers have been
racial bias definitive answers have been hard to come by
hard to come by
hard to come by although data initiatives have been
although data initiatives have been
although data initiatives have been sweeping the country drawbacks of
sweeping the country drawbacks of
sweeping the country drawbacks of current approaches have limited our
current approaches have limited our
current approaches have limited our ability to have an informed discussion
ability to have an informed discussion
ability to have an informed discussion sometimes contributing more confusion
sometimes contributing more confusion
sometimes contributing more confusion rather than clarity here’s one example
rather than clarity here’s one example
rather than clarity here’s one example one might argue that the higher search
one might argue that the higher search
one might argue that the higher search rates for minorities are due to racial
rates for minorities are due to racial
rates for minorities are due to racial bias the reality of policing is that
bias the reality of policing is that
bias the reality of policing is that officers tend to be deployed in high
officers tend to be deployed in high
officers tend to be deployed in high crime areas areas that tend to be
crime areas areas that tend to be
crime areas areas that tend to be predominantly black and Hispanic so a
predominantly black and Hispanic so a
predominantly black and Hispanic so a common counter-argument is that
common counter-argument is that
common counter-argument is that the elevated search rates we see are
the elevated search rates we see are
the elevated search rates we see are simply a result of effective policing
simply a result of effective policing
simply a result of effective policing how can we move forward and propose
how can we move forward and propose
how can we move forward and propose solutions if we can’t accurately
solutions if we can’t accurately
solutions if we can’t accurately identify or measure the very thing we’re
identify or measure the very thing we’re
identify or measure the very thing we’re trying to address racial bias our team
trying to address racial bias our team
trying to address racial bias our team at Stanford has been working on methods
at Stanford has been working on methods
at Stanford has been working on methods to overcome some of the current
to overcome some of the current
to overcome some of the current limitations but I first want to show you
limitations but I first want to show you
limitations but I first want to show you why it’s so difficult to measure racial
why it’s so difficult to measure racial
why it’s so difficult to measure racial bias as think of traffic stops now in
bias as think of traffic stops now in
bias as think of traffic stops now in order to identify bias we would need to
order to identify bias we would need to
order to identify bias we would need to know whether an office
know whether an office
know whether an office would given to drivers in otherwise
would given to drivers in otherwise
would given to drivers in otherwise identical contexts react differently
identical contexts react differently
identical contexts react differently when one is white and the other is not
when one is white and the other is not
when one is white and the other is not now obviously this is impossible so what
now obviously this is impossible so what
now obviously this is impossible so what can we do well the most commonly used
can we do well the most commonly used
can we do well the most commonly used tests to measure racial bias is known as
tests to measure racial bias is known as
tests to measure racial bias is known as the outcome test it compares the search
the outcome test it compares the search
the outcome test it compares the search success rate or hit rate among race
success rate or hit rate among race
success rate or hit rate among race groups what’s a hit well a hit is when
groups what’s a hit well a hit is when
groups what’s a hit well a hit is when an officer searches a car and finds
an officer searches a car and finds
an officer searches a car and finds contraband that is something illegal so
contraband that is something illegal so
contraband that is something illegal so it’s kind of like a homerun for the
it’s kind of like a homerun for the
it’s kind of like a homerun for the officer the idea is if searches of
officer the idea is if searches of
officer the idea is if searches of minorities turn up contraband 10% of the
minorities turn up contraband 10% of the
minorities turn up contraband 10% of the time where are the searches of whites
time where are the searches of whites
time where are the searches of whites turn up contraband 90% of the time then
turn up contraband 90% of the time then
turn up contraband 90% of the time then this adjusts officers are searching
this adjusts officers are searching
this adjusts officers are searching minorities on the basis of less evidence
minorities on the basis of less evidence
minorities on the basis of less evidence than whites now this test sounds very
than whites now this test sounds very
than whites now this test sounds very reasonable right but it turns out it has
reasonable right but it turns out it has
reasonable right but it turns out it has a well-known limitation and to see this
a well-known limitation and to see this
a well-known limitation and to see this let’s think about what happens during a
let’s think about what happens during a
let’s think about what happens during a traffic stop so here’s your typical
traffic stop so here’s your typical
traffic stop so here’s your typical driver during the course of the stop the
driver during the course of the stop the
driver during the course of the stop the officer has discretion to search if they
officer has discretion to search if they
officer has discretion to search if they have discretion to search if they have
have discretion to search if they have
have discretion to search if they have sufficient reason to believe the drivers
sufficient reason to believe the drivers
sufficient reason to believe the drivers can contraband
can contraband
can contraband so they observe the demographics of the
so they observe the demographics of the
so they observe the demographics of the driver the time of day whether it’s a
driver the time of day whether it’s a
driver the time of day whether it’s a dangerous neighborhood and they look for
dangerous neighborhood and they look for
dangerous neighborhood and they look for clues whether the driver is acting
clues whether the driver is acting
clues whether the driver is acting nervous or evasive there’s any odors in
nervous or evasive there’s any odors in
nervous or evasive there’s any odors in the car more factors than we can ever
the car more factors than we can ever
the car more factors than we can ever hope to measure and in order to make
hope to measure and in order to make
hope to measure and in order to make their decision we imagine they distill
their decision we imagine they distill
their decision we imagine they distill all this information down into a single
all this information down into a single
all this information down into a single number which represents their subjective
number which represents their subjective
number which represents their subjective estimate that the driver is going to
estimate that the driver is going to
estimate that the driver is going to possess contraband now if their estimate
possess contraband now if their estimate
possess contraband now if their estimate is sufficiently high they’re going to
is sufficiently high they’re going to
is sufficiently high they’re going to conduct a search that’s a pretty
conduct a search that’s a pretty
conduct a search that’s a pretty suspicious-looking driver right so now
suspicious-looking driver right so now
suspicious-looking driver right so now imagine we had this information for all
imagine we had this information for all
imagine we had this information for all drivers then we could order the drivers
drivers then we could order the drivers
drivers then we could order the drivers an increasing order of suspicion we’d
an increasing order of suspicion we’d
an increasing order of suspicion we’d expect that above a certain threshold of
expect that above a certain threshold of
expect that above a certain threshold of suspicion all drivers would be searched
suspicion all drivers would be searched
suspicion all drivers would be searched now say two out of the four searches
now say two out of the four searches
now say two out of the four searches were successful the hit rate would be
were successful the hit rate would be
were successful the hit rate would be one half okay so everyone with me so far
one half okay so everyone with me so far
one half okay so everyone with me so far now let’s see how the outcome task can
now let’s see how the outcome task can
now let’s see how the outcome task can fail so here we have two race groups
fail so here we have two race groups
fail so here we have two race groups green and blue the hit rate are equal
green and blue the hit rate are equal
green and blue the hit rate are equal indicating that the outcome test does
indicating that the outcome test does
indicating that the outcome test does not find evidence of bias but notice
not find evidence of bias but notice
not find evidence of bias but notice that the least suspicious green driver
that the least suspicious green driver
that the least suspicious green driver that was searched had an estimated 25
that was searched had an estimated 25
that was searched had an estimated 25 percent chance of having contraband
percent chance of having contraband
percent chance of having contraband whereas the lis suspicious blue driver
whereas the lis suspicious blue driver
whereas the lis suspicious blue driver that was searched had an estimated 51%
that was searched had an estimated 51%
that was searched had an estimated 51% chance so these green drivers would not
chance so these green drivers would not
chance so these green drivers would not have been searched had they been blue
have been searched had they been blue
have been searched had they been blue the lower threshold for green suggests
the lower threshold for green suggests
the lower threshold for green suggests that police require less evidence when
that police require less evidence when
that police require less evidence when searching green drivers so what we
searching green drivers so what we
searching green drivers so what we actually want to measure are the search
actually want to measure are the search
actually want to measure are the search thresholds themselves the only problem
thresholds themselves the only problem
thresholds themselves the only problem is we can’t observe them we can only
is we can’t observe them we can only
is we can’t observe them we can only observe the hit rates so we’re left with
observe the hit rates so we’re left with
observe the hit rates so we’re left with a flawed metric to measure one of the
a flawed metric to measure one of the
a flawed metric to measure one of the most invasive actions police can take to
most invasive actions police can take to
most invasive actions police can take to search someone it’s like we’re trying to
search someone it’s like we’re trying to
search someone it’s like we’re trying to diagnose a patient with unreliable tests
diagnose a patient with unreliable tests
diagnose a patient with unreliable tests using miscalibrated equipment so in
using miscalibrated equipment so in
using miscalibrated equipment so in order to address this problem our team
order to address this problem our team
order to address this problem our team at Stanford developed a mathematical
at Stanford developed a mathematical
at Stanford developed a mathematical model to determine whether searches are
model to determine whether searches are
model to determine whether searches are consistently effective given the number
consistently effective given the number
consistently effective given the number of stops searches and hits we can infer
of stops searches and hits we can infer
of stops searches and hits we can infer the search thresholds being used by a
the search thresholds being used by a
the search thresholds being used by a department for each race group we apply
department for each race group we apply
department for each race group we apply the test to 4.5 million stops in North
the test to 4.5 million stops in North
the test to 4.5 million stops in North Carolina each point here is a department
Carolina each point here is a department
Carolina each point here is a department the x-axis is the white threshold and
the x-axis is the white threshold and
the x-axis is the white threshold and the y-axis is the minority threshold so
the y-axis is the minority threshold so
the y-axis is the minority threshold so points below the line indicates that
points below the line indicates that
points below the line indicates that thresholds for minorities are
thresholds for minorities are
thresholds for minorities are consistently lower across the largest
consistently lower across the largest
consistently lower across the largest 100 departments in the state now we’ve
100 departments in the state now we’ve
100 departments in the state now we’ve tested and ruled out legitimate reasons
tested and ruled out legitimate reasons
tested and ruled out legitimate reasons why thresholds might vary such as time
why thresholds might vary such as time
why thresholds might vary such as time of day year or policy changes and
of day year or policy changes and
of day year or policy changes and crucially we cannot determine whether
crucially we cannot determine whether
crucially we cannot determine whether the lower thresholds are due to
the lower thresholds are due to
the lower thresholds are due to discrimination or implicit bias
discrimination or implicit bias
discrimination or implicit bias or some other unknown factor but the
or some other unknown factor but the
or some other unknown factor but the results do suggest there are some
results do suggest there are some
results do suggest there are some striking irregularities in how search
striking irregularities in how search
striking irregularities in how search decisions are made we must also be
decisions are made we must also be
decisions are made we must also be cognizant that the data we choose to
cognizant that the data we choose to
cognizant that the data we choose to collect and the models we build are
collect and the models we build are
collect and the models we build are laden with values the way we choose to
laden with values the way we choose to
laden with values the way we choose to discretize and record the world around
discretize and record the world around
discretize and record the world around us
us
us will determine the stories our data will
will determine the stories our data will
will determine the stories our data will be able to tell and models are built on
be able to tell and models are built on
be able to tell and models are built on assumptions ones that we must
assumptions ones that we must
assumptions ones that we must continuously test and perfect but the
continuously test and perfect but the
continuously test and perfect but the very fact that we lack perfect tools
very fact that we lack perfect tools
very fact that we lack perfect tools today is all the more reason to strive
today is all the more reason to strive
today is all the more reason to strive for rigor if we are to ever reach a
for rigor if we are to ever reach a
for rigor if we are to ever reach a common ground we are we must have the
common ground we are we must have the
common ground we are we must have the courage to ask some uncomfortable
courage to ask some uncomfortable
courage to ask some uncomfortable questions perhaps find some
questions perhaps find some
questions perhaps find some uncomfortable answers and start all over
uncomfortable answers and start all over
uncomfortable answers and start all over again when our solutions don’t work
again when our solutions don’t work
again when our solutions don’t work police undoubtedly have a tough job
police undoubtedly have a tough job
police undoubtedly have a tough job there’s a high cost to search in someone
there’s a high cost to search in someone
there’s a high cost to search in someone innocence the lack of trust and there’s
innocence the lack of trust and there’s
innocence the lack of trust and there’s a high cost to not search in someone
a high cost to not search in someone
a high cost to not search in someone who’s potentially dangerous and Safe
who’s potentially dangerous and Safe
who’s potentially dangerous and Safe Communities and higher crime Sir Robert
Communities and higher crime Sir Robert
Communities and higher crime Sir Robert Peel who’s commonly regarded as the
Peel who’s commonly regarded as the
Peel who’s commonly regarded as the father of modern policing envisioned a
father of modern policing envisioned a
father of modern policing envisioned a society where the bond between police
society where the bond between police
society where the bond between police and the communities they served was so
and the communities they served was so
and the communities they served was so strong that it would be like they were
strong that it would be like they were
strong that it would be like they were one in the same and until the day our
one in the same and until the day our
one in the same and until the day our data and our narratives will speak of
data and our narratives will speak of
data and our narratives will speak of Kiehl’s vision we mustn’t stop working
Kiehl’s vision we mustn’t stop working
Kiehl’s vision we mustn’t stop working [Applause]
”
Be First to Comment