Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
Hi. Everyone, it's Magna here with
0:02
a special from the archives podcast
0:04
Drop From On Point. Daniel Kahneman,
0:06
one of the world's most celebrated
0:09
economists, died this week at the
0:11
age of ninety. The Nobel prize
0:13
winner was one of the pioneers
0:16
in a field that later became
0:18
known as behavioral Economics. His groundbreaking
0:20
work showed that human intuitive reasoning
0:23
is flawed. In. Predictable Ways
0:25
the predictability being the breakthrough
0:27
kind of. Was also author
0:29
of thinking Fast and Slow,
0:31
a highly influential book that
0:33
debunked a long cherished beliefs
0:35
in economics that humans are
0:37
rational actors. His work, along
0:39
with others showed both qualitatively
0:41
and quantitatively that no matter
0:44
how much economists want to
0:46
believe it, human decision making
0:48
is not a rationally driven
0:50
process. Caught. Him and
0:52
was also a holocaust survivor. His
0:54
family was forced to wear the
0:57
Yellow Star of David in Occupied
0:59
France before their escape. He later
1:01
frequently said that his experience of
1:03
the holocaust was one of the
1:05
things that drove his powerful interest
1:07
in understanding the human mind. In
1:10
June of Twenty Twenty One we
1:12
spoke with Con I'm in an
1:14
his coauthor Olivier see Bony about
1:16
their latest book. It's called noise
1:18
or flaw in human judgment and
1:20
it's about how even though we're
1:22
told to trust our judgment, That.
1:25
Judgment is way more variable
1:27
than we think it is.
1:29
It's also about how that
1:31
variability or noise influences almost
1:33
every part of our lives.
1:35
So today from the archives
1:37
we offer you our conversation
1:39
with Daniel Kahneman. I. Hope
1:41
you enjoy. This
1:48
is on point. I Magna Chakrabarti.
1:50
Back when I was nineteen
1:52
years old, I suddenly suffered
1:54
from an auto immune disease
1:57
known as idiopathic thrombosis, I
1:59
do panic. It's
2:01
a mysterious ailment where my immune
2:03
system was attacking my own blood
2:05
platelets, and it was pretty
2:07
serious. Well, the first doctor
2:10
I saw said, we don't really know
2:12
what causes this, so I recommend
2:14
waiting and watching, just don't do
2:16
any major physical activity and it
2:18
might resolve itself. That
2:21
seemed too passive. So
2:24
the second doctor I went to, he said, to
2:27
reduce the autoimmune response,
2:29
I recommend surgery to
2:31
remove your spleen. Well
2:34
that was very aggressive.
2:37
So I asked, what's the chance that that
2:39
procedure will work? And he said, 50-50. Okay,
2:44
so I went to a third doctor and that doctor
2:46
said, you can take a complex
2:48
series of steroids for several months
2:50
and see what happens. Will
2:53
it work? I asked. And he said, I
2:56
don't know. Now
2:58
I didn't mind the uncertainty because
3:00
that is a fact in complex
3:02
systems and the body is a
3:05
profoundly complex system. What
3:07
threw me was the wildly different
3:09
solutions proposed by the three different
3:12
physicians all for the same ailment
3:14
in the same person. And as
3:16
a patient, I did not
3:18
know how to cope with that variability.
3:22
Now you know, I'm not a
3:24
particular fan of n equals
3:26
one examples or using one
3:28
anecdote to describe an entire
3:30
system, but it turns
3:32
out that kind of
3:34
variability is rampant in
3:37
the very professions, systems and
3:39
organizations whose judgment we are
3:41
meant to trust the most.
3:45
It is a huge, costly
3:47
and often unnoticed problem.
3:50
And it's a problem that Nobel
3:53
prize winning psychologist Daniel Kahneman, Olivier
3:55
Sibonie and Cass Sunstein write at
3:58
length about in their new- book,
4:00
Noise, a Flaw in
4:02
Human Judgment. And today,
4:05
Daniel Kahneman joins us. Professor Kahneman,
4:07
welcome to On Point. Glad
4:10
to be here. And Professor
4:12
Ciboney, welcome to you as well. Glad
4:15
to be here as well. Okay. So first, let
4:17
me ask you, I did open
4:19
up with that personal anecdote about
4:22
the medical system. But
4:25
Daniel Kahneman, how common or
4:27
how much noise is in
4:30
medicine, in decision-making amongst
4:32
doctors? Well, the
4:35
long and short of it is there is a
4:37
lot of noise. Doctors don't
4:40
agree with each other in many
4:42
cases, and they don't even agree
4:44
with themselves when shown the same
4:47
set of tests on
4:49
different occasions. So, yeah,
4:51
there's a lot of noise, and there's a
4:54
lot of noise in all professional judgment, not
4:56
only in medicine, but wherever
4:58
people make judgments, you
5:01
can expect to find noise, and you can expect
5:03
to find a surprising amount of noise.
5:06
A surprising amount. Well, the
5:08
thing about, I started with
5:11
medicine because it's one of the
5:13
systems that almost everyone has interactions
5:15
with at some point, if not
5:17
multiple points, in their lives.
5:20
Can you tell me, and Professor Ciboney, I'll
5:22
turn to you on this, can you tell
5:24
me a little bit more about what Daniel
5:26
Kahneman was saying about how doctors even disagree
5:29
with themselves when looking at sort
5:31
of the same set of information about a
5:33
particular case? How
5:36
do we understand that? A typical example
5:38
would be radiologists, and I
5:40
suspect that radiologists are not better
5:42
or worse than other doctors.
5:46
It's just that it's easier to test radiologists
5:48
because you can show them an x-ray that
5:50
they've seen some weeks or some months ago
5:52
and ask them, what is this? And
5:54
if they don't answer, of course, they cannot
5:56
recognize the x-ray because they see a lot
5:58
of x-rays. And if they. The tell you
6:00
something different from what they had told you
6:02
some weeks or so month ago. When looking
6:05
at the same x rayed you know that
6:07
is noise. Now that is by the way,
6:09
a different type of noise from the one
6:11
that you were dealing with in your example.
6:13
Make now because this would be noise in
6:15
the diagnosis which is a matter of sucks.
6:17
You either have this bizarre disease that you
6:20
were talking about the name of wish I
6:22
could not a member or you do at
6:24
least two. Three doctors seems to agree on
6:26
the diagnosis. They disagreed on the treatments which
6:28
is already something you you. Might find excuses
6:30
for the be there isn't an obvious treatment.
6:32
Maybe it's a very rare disease we don't
6:34
know, but in the examples that we document
6:36
in the book, the actually disagree on the
6:39
reality of the diagnosis of the disease that
6:41
is present. Their. Eyes. Which
6:43
is the bigger issue? Presumably okay.
6:45
Okay so so let's than step.
6:47
Step back here for a moment.
6:49
I suppose we should. We should
6:51
actually begin with basic definitions here,
6:54
Sir Daniel Home And when we're
6:56
talking about noise in assistant, right,
6:58
we're not homeless individuals that we're
7:00
talking about the organizational level. Hear?
7:02
how do we define what noises.
7:05
Well. Would apply noise as
7:07
unwanted variability in judgments
7:09
that should be identical
7:11
and thus the broad definition.
7:14
So you're you're you're three.
7:16
Physicians make judgments about the
7:19
same case and we
7:21
would expect them to give
7:23
identical answers. The
7:26
fact that they're variable is is
7:28
an indication that something is wrong
7:30
with the system. And
7:33
if I may, Ah! You're one
7:35
of. Not. One of I'd
7:37
say you probably the best known a
7:39
psychologist in in the world right now.
7:42
Okay, or at least one of them
7:44
are. And your previous work, ah, a
7:46
Sinking Fast and Slow is an incredibly.
7:49
Influential book Here does this interest
7:51
in the how a human judgments
7:54
across a systemic or organizational scale
7:56
it seems it must have in
7:58
naturally. Flows from your. Previous where it
8:01
doesn't not no, actually did not.
8:03
My previous work all my life
8:05
was studied. individuals have studied biases
8:08
and not noise and have been.
8:10
I knew that noise exist and
8:12
everyone knows that so that when
8:14
when it anything as a matter
8:17
of judgment people are not supposed
8:19
to agree exactly so there is
8:21
some noise. A with turned out
8:24
to be surprising was that as
8:26
some seven years ago and while
8:28
in on a consulting. Engagement
8:31
with an insurance company. I discovered
8:33
that there was much more disagreement
8:35
than anybody expected more than the
8:38
executives expected more than the underwriters
8:40
whom we looked at expected about
8:43
by factor of five, by the
8:45
way. so it's not a small
8:47
effect and that's that sets me
8:50
on the schools. Than Olivia join
8:52
me than cast joined us. and
8:55
and the book of came out
8:57
about seven years after as. And
8:59
Disagreement: A monks are using
9:02
underwriters in particular an insurance
9:04
industry. Vs. So you the way
9:06
that we conducted the experiment and
9:08
we call that the noise or
9:10
that because it's it's quite gentle.
9:12
You can conduct experiments like this.
9:14
In many cases they constructed cases
9:17
of that were realistic but fictitious.
9:19
You don't need to know the
9:21
correct answer in order to measure
9:23
noise. and then they present of
9:25
the same cases to above sixty
9:27
underwriters and each of them had
9:29
to give a dollar value. And
9:32
a question that we asked ourselves
9:34
and that we asked executives was
9:36
how much do they differ And
9:38
to get the sense of the
9:40
magnitude of the difference, think that
9:42
you pick to underwriters at random
9:44
from those who looked at the
9:46
same case and by how much
9:49
do they differ in percentages that
9:51
is it takes the average of
9:53
the of judgement difference. You divide
9:55
the difference by the average. Hello
9:58
to the difference. And
10:00
I asked the executives that question, not
10:03
all executives, but a few. And
10:06
since then, we especially Olivier
10:08
have collected a lot of
10:11
information on what people expect.
10:13
People expect about 10% variation
10:15
on quantitative judgment. That looks
10:17
sort of tolerable and reasonable.
10:19
You don't expect perfect agreement.
10:23
10% is sort of tolerable. The
10:25
answer for those underwriters
10:27
was 55%. So
10:31
that is not
10:33
an order of magnitude, but that
10:35
is qualitatively different from what anybody
10:37
had expected. It raises questions about
10:39
whether those underwriters were doing anything
10:42
useful for the company. I was
10:44
going to ask that because if
10:46
there's that much variability, what exactly are
10:48
they doing, right? It
10:51
is quite unclear. And I
10:53
think there is a movement
10:56
in insurance companies actually to take away
10:58
that role of judging, of evaluating risk
11:00
to take it away from underwriters and
11:02
to have them mainly as negotiators and
11:05
to have the judgments sort
11:08
of automated or made centrally. But
11:11
at the time, that was the
11:13
practice in that insurance company, underwriters
11:16
were actually setting dollar premiums. And
11:20
the striking thing that really set
11:22
this book in motion was not
11:25
only that there was a huge amount
11:27
of variability, but that the
11:29
executives in the company had not been aware
11:31
of it and that in fact the organization
11:33
did not know that it had the newest
11:35
problem. So when you can have
11:37
a problem of that magnitude that people are not aware
11:39
of, maybe there is something to
11:41
be studied. That's
11:44
what we did. So then I think we
11:46
need to understand more clearly, and Professor
11:49
Ciboni, I'll turn to you for this, how
11:51
then does noise
11:53
in your description
11:55
of it differ from another
11:57
word that Danny Kahneman used?
12:00
just a moment ago from bias. There
12:03
is actually a very easy way to think
12:05
about it and it's to think of an
12:07
example of measurement as opposed to judgment. It's
12:09
easier to figure it. So suppose
12:12
you step on your bathroom scale every
12:14
morning and on average
12:16
your bathroom scale is kind. It tells
12:19
you that you're a pound lighter than you
12:22
actually are on average every day.
12:25
That's a bias. That's an error of
12:27
a predictable direction and on
12:29
average is the error that your scale is
12:31
making. Now suppose that you step on your
12:33
bathroom scale three times in quick succession and
12:35
you read a different number each time. That
12:38
is random variability of something that should be
12:41
identical. It is noise. Now apply
12:43
this to judgment to see the difference between
12:45
the bias which is the average error and
12:48
the noise which is the random variability in
12:50
the judgment. Suppose that we're making
12:52
a forecast of say what the GDP growth
12:54
is going to be next year or something
12:56
like that. If on average all
12:59
of us who are making this forecast
13:01
tend to be optimistic that's a bias.
13:03
We overestimate that's an average error but
13:06
each of us is going to make a slightly different
13:08
forecast. The variability between our forecasts is
13:10
noise. So it's really quite simple. Bias
13:12
is a predictable error in a given
13:15
direction. It's the average error of a
13:17
number of people or a number of
13:19
observations by the same person. Noise
13:21
is a variability in those observations. Well
13:25
this hour we are talking with
13:27
Olivier Siboni and Daniel Kahneman,
13:29
the Nobel Prize-winning psychologist about their
13:32
new book Noise, a flaw
13:34
in human judgment and
13:36
how that flaw in human
13:38
judgment is amplified across organizations
13:41
and systems that can touch us all.
13:43
We'll be right back. Support
13:58
for the EnPointe podcast. The podcast comes from
14:01
Indeed. We're driven by the search for
14:03
better, but when it comes
14:05
to hiring, the best way to search
14:07
for a candidate isn't to search at
14:09
all. Don't search, match with Indeed. Ditch the
14:11
busy work and use Indeed for scheduling,
14:13
screening, and messaging so you can connect
14:15
with candidates faster. And listeners will
14:17
get a $75 sponsored job credit to
14:20
get your jobs more visibility
14:22
at indeed.com/on point. That's
14:25
indeed.com/on point. Terms
14:27
and conditions apply. Need to hire?
14:29
You need Indeed. This
14:32
is on point. I'm Meghna Chakrabarty,
14:34
and today we're talking with Daniel
14:37
Kahneman. He's the Nobel Prize-winning psychologist,
14:39
perhaps the world's most influential
14:41
psychologist. He's the
14:43
author of the book, Sinking Fast and
14:46
Slow, and we're talking as well
14:48
with Professor Olivier Sibonie. They,
14:50
along with Cass Sunstein, have co-authored a new
14:52
book called Noise, A Flaw
14:54
in Human Judgment, about how
14:57
that flaw, or
15:00
human judgment's flaws, gets amplified through systems, creates
15:03
noise, and makes it harder to come
15:06
to the right decisions and
15:08
what to do about it as well. So
15:11
I'd like to focus with
15:14
the two of you on one particular
15:16
system that you write at length
15:18
about, and that is
15:20
the judicial system. So Professor
15:22
Sibonie, I wonder if you can
15:24
help us understand, what
15:27
is the evidence that
15:29
there is a great deal of noise, or
15:32
this unwanted variability, as you both called
15:34
it, in judgments in the
15:36
judicial system? So
15:38
there has been evidence for quite a while. One
15:41
of the studies that we cite in the book
15:43
goes back to the 1970s, and in that study,
15:45
a great many judges, 208 judges to
15:48
be precise, looked
15:53
at vignettes describing cases,
15:55
so very simplified
15:58
descriptions of cases. where
16:00
you would expect pretty good agreement
16:03
on how to sentence a particular
16:05
defendant because the judges aren't
16:07
distracted by the particulars of what happens
16:09
in the courtroom or by the
16:11
looks of the defendant or by any distracting
16:15
information. You would expect some
16:17
consistency, perhaps not perfect consistency, but
16:19
at least some consistency. And
16:22
it turns out that on some of those
16:24
cases, one judge would say 15
16:27
days and another one would say 15 years.
16:29
On average, for a seven-year prison
16:31
term that was the average given
16:33
by the 200 judges, there
16:36
was, if you were to pick two
16:38
different judges, a difference of almost four
16:40
years in what the sentence would be.
16:43
Which basically tells you that if you're
16:45
a defendant, the moment you walk into
16:47
the courtroom because you've been assigned to
16:49
a particular judge, that has
16:51
already added two years or subtracted two
16:54
years from what would be otherwise a
16:56
seven-year sentence. That is
16:58
truly shocking. You would
17:01
want, of course, the specifics of the
17:03
case and the specifics of the defendant
17:05
and all the particular circumstances of a
17:07
particular offense to be taken into account.
17:10
But the particular circumstances of the
17:13
judge should not make a
17:15
big difference to the sentence, and they
17:17
do. And there have been quite
17:19
a few other studies replicating and amplifying
17:21
this finding, which basically tell you that
17:24
who the judge happens to be has
17:26
a very, very large
17:29
influence on the sentence. Of course
17:31
we know that, but it's
17:33
much larger than we suspect it is. Right.
17:35
I mean, the legal profession has known this
17:37
for quite some time, to your point, that
17:40
they would, you know, lawyers always talk about
17:42
hoping to get assigned particular judges for their
17:45
clients. But just to be clear, so
17:47
these judges in the studies that you're
17:49
talking about were given sort of
17:52
stripped-down information about cases so
17:54
that ostensibly the factors that
17:56
would normally contribute to bias
17:59
from the... individual judge were removed and yet
18:01
we still saw this variability in sentencing.
18:04
Is that what you're saying? That
18:06
is right. You can only expect
18:08
that in reality the noise would be
18:11
much worse than what we measure here
18:13
because these are stripped down cases where
18:15
all the distracting information that could add
18:17
and amplify the biases of the judge
18:20
has been taken out. Okay. So
18:23
Daniel Kahneman, do we know why there
18:25
was so much variability even in
18:27
these controlled circumstances amongst these judges
18:29
who – I mean their
18:32
profession is called judges. We are told
18:34
we are supposed to trust their judgment.
18:38
Well actually there is more than one source
18:40
of noise. We
18:43
distinguish three. So
18:45
one source of noise are differences in
18:48
the severity and that's hanging
18:51
judges and others. That's
18:53
the mean sentence,
18:56
the mean of a lot of sentences that the
18:58
judge gives differ across
19:00
judges. We call that
19:02
level noise. Then there is the
19:05
noise within a judge that
19:07
is elements that are
19:11
like the weather it turns out that
19:13
the sentences are more severe on hot
19:15
days. It turns out that judges are
19:17
more lenient when their football team just
19:19
won a game. Those
19:23
are small effects but they are reliable effects.
19:26
So it turns out that there
19:28
is a lot of noise within the judge
19:31
just as we were talking earlier
19:33
about radiologists. And
19:36
probably the largest source
19:38
of noise is that the judges
19:40
differ in how they see crimes.
19:43
They have different tastes in
19:45
crimes and different tastes in
19:47
defendants. Some of them are
19:49
more shocked by one kind of crime, others
19:51
by another. And there
19:54
are stable but idiosyncratic
19:56
differences among
19:59
judges. In that mysterious
20:01
set of differences, we call that
20:04
the judgment personality, seems to
20:07
account for much of the
20:09
differences among judges in
20:11
this traditional system and probably in
20:14
other professional judgments as well. And
20:16
is that judgment personality formed,
20:19
it must be formed over
20:21
time by that own judges,
20:23
I don't know, both their
20:25
DNA and their personal experiences
20:28
as they developed as humans? Absolutely,
20:32
except we know very little
20:34
about them because we
20:36
know it's just like personalities, we don't
20:39
expect personalities to be the same, but
20:41
we actually expect to see the world
20:43
in the same way. That
20:45
is, I don't expect you to like the same
20:47
things as I do, I don't expect you to
20:50
behave the same way as I do, but
20:52
I do expect you when we are looking
20:55
at the same situation to see it as
20:57
I do because I see it correctly. And
21:01
if I respect you, since I see
21:03
the situation the way it is, I
21:05
expect you to see exactly the same
21:08
thing that I do, and that expectation
21:10
is incorrect. I am curious
21:12
though about the
21:14
second factor that you talked about, even if it's less
21:18
influential, but the susceptibility
21:21
of everyone, but in this
21:23
case, judges sitting on the
21:25
bench to almost
21:27
imperceptible things like the weather or
21:30
whether their team won the
21:32
game the previous night or not.
21:35
Because if the susceptibility to
21:37
all manner of environmental
21:40
inputs is part of the
21:42
problem here, it seems as if
21:44
it would be impossible to meaningfully
21:46
reduce noise because it
21:48
would require changing what makes us human,
21:51
Professor Kahneman. Well, it
21:54
really depends on, to
21:56
some, we must expect that noise
21:58
will remain. so long as
22:01
there is judgment, because that actually
22:03
defines judgment. A matter of judgment
22:05
is a matter on which you
22:07
expect some disagreement. So you're not
22:09
going to resolve it completely. But
22:12
there are procedures, we think,
22:14
that if followed by
22:17
judges, are going to make them less
22:19
susceptible to variation. A source of variation
22:21
I'd like to mention, by the way,
22:23
is time of day. Procedures
22:26
and physicians are
22:29
different in the morning and in the afternoon when
22:31
they are hungry and when they are not hungry.
22:33
So those are substantial variabilities.
22:36
Wow. Well, so I want to talk more about
22:39
some of those procedures in a moment. But
22:41
Professor Siboni, let me turn back to you
22:43
here for a second, because I understand the
22:47
intellectual utility of the
22:49
kinds of studies that we're discussing here regarding
22:52
noise in the judicial system. But
22:55
at the same time, we're talking about judges looking at
22:57
cases that have been stripped of a
22:59
lot of detail, right? And
23:02
isn't part of what we
23:04
are actually entrusting to judges
23:07
is their discernment to come up with
23:09
the right sentence, given the
23:12
individual details of the cases that
23:14
they are hearing, that in fact
23:16
those details matter and then the
23:19
judgments made by the people
23:22
wearing the black robes should
23:24
be trusted. So how much
23:27
can we take these stripped down
23:29
studies and say that they
23:31
are really pointing to something fundamentally flawed in
23:33
the judicial system? Well, we have
23:36
every reason to believe that if you add
23:38
the real details that you see in a
23:40
real courtroom, it would make the noise worse.
23:43
Now there is an easy way to test
23:45
that, which would be to actually take judges,
23:47
I'm saying it's
23:50
actually not easy to do, but it's easy in principle, which
23:52
would be to take a number of judges and
23:54
have them sit in separate boxes
23:57
looking at the same trial, and
23:59
at the end actual trial having seen
24:01
the real descendants and the real jurors
24:03
and the real witnesses and so on
24:05
set a sentence. We would
24:07
see there what the real, that would be
24:09
a real full scale noise audit
24:12
if you will, where you would
24:14
see what the real noise is with real cases.
24:16
To our knowledge this hasn't been done because you
24:18
can see it's a cumbersome experiment. But
24:21
we are pretty convinced. I
24:23
think there's good reason to believe that all
24:25
the details you would see in an
24:28
actual trial like this would
24:30
only make the divergence between the
24:32
judges worse than it is in
24:34
the stripped down cases. Okay.
24:37
So, then help me understand.
24:41
Daniel Kahneman, have you, or actually both
24:43
of you, but Professor Kahneman, I'll turn to you for
24:45
this. Have you spoken
24:47
with judges about
24:50
this and how do they respond
24:52
when presented with this evidence of
24:55
the sort of built in noise
24:58
in their decision making? Well, I
25:01
have spoken with judges, but not
25:03
enough to form an opinion. But
25:05
a lot is known about the
25:07
reaction of judges to discussions
25:10
of noise and to
25:13
guidelines that were introduced in an
25:15
attempt to control the amount of
25:18
noise by setting boundaries for
25:21
different crimes. And
25:24
apparently judges hated them. The
25:27
guidelines were eliminated at some
25:29
point for reasons that are
25:31
not pertinent to the case. But
25:34
it turns out that judges are much
25:36
happier about their job ever since. And
25:38
clearly there is now more variability than
25:41
there was. So the
25:44
situation with there is a lot of
25:46
noise is a situation that judges are
25:48
entirely comfortable with. They
25:52
are comfortable with the situation. They don't know
25:54
there is noise. And
25:56
maybe I may add something here based on
25:58
my own. anecdotal
26:00
conversations with judges, here's
26:03
how the conversation basically goes. You
26:05
say there is noise and you give
26:07
them this evidence and basically they shrug.
26:09
They say, well, yeah, that's the reality
26:11
of making judgments. Every case is different.
26:13
So we're going to make different
26:15
judgments every time. And then
26:18
you ask them, well, okay, so the
26:20
same defendant is going to get a
26:22
different sentence depending on whether he's assigned
26:24
to you or to the judge next
26:26
door. And they say, yeah, that's life.
26:29
And then you ask them, well, what
26:31
if the same defendant got a different
26:33
sentence because his skin is of a
26:35
different color? And they
26:37
say, no, that would be completely unacceptable. And
26:40
then they realize that we have
26:43
a very different level
26:45
of outrage when we
26:47
can explain the cause of the
26:49
discrepancy, when we can identify a
26:52
bias, and when it
26:54
is noise that we can not identify.
26:56
And there isn't any obvious
26:58
reason why we should feel it's
27:00
completely acceptable for these differences to
27:02
appear for reasons we do not
27:04
understand. Whereas it is totally
27:06
unacceptable. And I think we would all agree
27:09
on that for them to appear because of
27:11
reasons that we do understand. And
27:13
that's what we're trying to point out when
27:15
we raise the question of noise in the
27:17
judicial system. Why do we tolerate large
27:20
differences that are caused by noise when
27:22
we would not tolerate them if they
27:24
were caused by bias? Okay. So you
27:27
mentioned both of you mentioned guidelines. And
27:30
Professor Kahneman, can you just elaborate
27:32
a little bit more about within the criminal
27:35
justice context, what you meant by guidelines?
27:38
Well, there was
27:41
a commission set up, I think, in the 1970s or 1980s to
27:43
discuss, well, to assign to each
27:51
crime as defined in
27:53
the law, to assign a
27:55
range of sentences and judges
27:58
were strongly discouraged from going
28:01
outside that
28:04
range, they were allowed to
28:07
do it. So there was discretion,
28:09
but clearly the
28:11
guidelines have a great deal of effect,
28:14
and the variability of sentences for any
28:16
given crime indeed diminished. So
28:18
those were just
28:21
part of the definition of the crime, is
28:23
the range of sentences that are allowed to
28:25
go with it. That's the guideline. Okay, so
28:28
the reason why I want to ask you about
28:30
that is because you do talk about the importance
28:32
of creating, I would say, the right kind of
28:34
guidelines to reduce noise in
28:37
organizations and systems, because the one
28:41
that pops up in my mind right now, which I think
28:44
has been deemed to
28:46
be something of a failure, is exactly what
28:49
you're talking about, mandatory minimum
28:51
sentencing, for example, in
28:53
drug crimes. You're right,
28:55
the judge's discretion was
28:57
removed from them with
29:00
mandatory minimum sentences, and the
29:03
part of the logic
29:06
behind those mandatory minimums was
29:08
to reduce variability in sentencing.
29:12
However, what we
29:14
saw, one of the outcomes
29:16
of that was also many, many
29:19
people being sentenced to
29:21
extremely long periods of
29:23
incarceration for relatively
29:26
minor drug crimes. So
29:29
there is still some sort
29:31
of systemic judgment that emerged
29:34
with those guidelines of
29:36
the mandatory minimums, which actually
29:38
made the problem of achieving justice even
29:42
worse. So Professor Siboni,
29:44
how do you find what the
29:46
right guidelines are without introducing a
29:48
whole other set of problems, in
29:51
trying to reduce unwanted variability and
29:54
reaching more identical solutions?
29:58
You can get a bunch of identical solutions. solutions that
30:00
aren't the right one. Absolutely.
30:03
And that would be bias, right? So that would
30:05
be an average error. If you think that the
30:08
proper sentence for a given crime
30:10
is one year in prison
30:12
and you set a mandatory minimum that is
30:15
ten years, you
30:19
have reduced noise because everybody will get ten years, but
30:21
you have created a lot of bias because everyone has
30:23
a sentence that is ten times worse than it should
30:25
be. And so that elevates
30:27
the question of what the proper sentence
30:29
should be to a debate
30:32
that has to take place in, I guess,
30:34
in the US Congress as
30:36
opposed to being a decision that is
30:38
being made separately by hundreds of judges
30:40
every day. Now it's
30:43
interesting that when it becomes a
30:45
problem of bias, when it becomes
30:47
a problem of the overall decision
30:49
being made at the wrong level,
30:52
it is at least a debate we can have. And
30:54
we can say three strikes and
30:56
you're out is terrible, mandatory minimum
30:58
sentences are terrible. We can have
31:01
that conversation. When that
31:03
decision is being made randomly
31:06
by judges all around the country every
31:08
day, the noise is
31:10
very hard to control and it leads to
31:12
many, many bad decisions as well.
31:15
Not all the decisions are uniformly bad,
31:18
but the randomness is in itself very bad.
31:21
That's the difference between bias and noise.
31:23
One is much easier to see, it's
31:26
much easier to counteract, it's much easier
31:28
to discuss and to combat.
31:31
The other is all over the place. And if you
31:33
don't do a noise audit to measure how much noise
31:35
there is, you can't even see it. Yeah.
31:38
Well, we are talking today with Olivier
31:41
Sibonie and Daniel Kahneman. Their new
31:43
book, along with Cass Sunstein, is
31:45
called Noise, a flaw in
31:47
human judgment. When we
31:50
come back, we'll talk about their recommendations
31:52
on how to reduce noise. This
31:54
is On Point. I'm
32:10
Kathleen Goldhar and I'm the host of
32:12
a new podcast, Crime Story. Every
32:15
week we bring you a different crime told
32:17
by the storyteller who knows it best. You
32:20
got one witness who can't be found.
32:22
You got another witness who's murdered. We
32:24
couldn't sugarcoat this story. I was getting
32:26
calls from Cosby's attorney threatening to sue
32:28
every day. Every crime in one
32:30
way or another is a reflection of who we
32:33
are as a people, as a city, as a
32:35
country. Find us wherever you get
32:37
your podcasts. This
32:40
is On Point. I'm Meghna Chakrabarti and
32:43
today we are talking with Olivier Ciboney
32:45
and Daniel Kahneman. Kahneman
32:47
is the Nobel Prize winning psychologist
32:49
and they, along with Cass
32:52
Sunstein, are co-authors of a fascinating
32:54
new book called Noise, a
32:57
Flaw and Human Judgment and
32:59
how noise or unwanted variability
33:02
in systems and organizations make
33:04
it really hard for those
33:06
systems and organizations to which we
33:09
all belong to operate at
33:11
our best interest. So we're trying to figure
33:13
out how to reduce noise. And
33:16
Professor Kahneman, before the break
33:18
we were talking about guidelines
33:20
and mandatory minimums in the
33:22
judicial system as one
33:24
perhaps flawed way in trying to
33:26
deal with the noise problem in
33:28
sentencing. I just wanted to quickly
33:30
hear your thoughts about that. Well,
33:34
you know, we should not
33:37
say that guidelines are a bad
33:39
idea because some guidelines were poorly
33:41
designed. In this case, clearly
33:44
there was a great deal of bias
33:46
in the setting of the guidelines. For
33:49
example, they distinguished among different kinds
33:51
of drugs in a way that
33:53
penalized track cocaine relative to
33:56
other drugs. Those are
33:58
poorly designed guidelines. guidelines
34:00
which will perpetuate bias rather
34:03
than eliminate error. But
34:06
you can design good guidelines
34:08
and the point about guidelines, and
34:10
here I echo something that Olivier
34:12
was saying earlier, the point about
34:14
guidelines is that you can see
34:16
them, you can discuss them, they're
34:19
out there. Noise is
34:21
something that you cannot see and you
34:24
cannot respond to appropriately. So
34:26
what other types of guidelines, just sticking
34:29
with the judicial system for one more
34:31
minute here, what other types of guidelines
34:33
that you suggest in the book might
34:36
be applicable here? Well,
34:40
if the guidelines are defined as
34:42
guidelines on sentencing, that's the
34:45
kind and that's the type, the only type.
34:48
We have ideas
34:52
about procedures, about
34:54
ways of thinking about the crime
34:57
and the defendant and the particular
34:59
case that we think might reduce
35:02
noise. But in terms of guidelines,
35:04
sentencing guidelines, well-designed sentencing
35:07
guidelines is what is
35:10
available I think. Okay, so then tell me
35:12
more about what you just said about the
35:17
other solutions for the judicial system. Well,
35:20
the general concept that we propose is
35:22
a concept that we call decision hygiene
35:24
and the term is almost
35:27
deliberately off-putting. It's to remind you
35:29
of what happens when
35:31
you wash your hands. And when you wash
35:33
your hands, you kill germs, you don't know
35:36
which germs you're killing and if you're successful,
35:38
you will never know. And it's
35:40
a sort of homely procedure but
35:43
it's extremely effective. And we have been
35:45
scouring the literature and what we
35:49
know to construct
35:52
a list of decision
35:54
hygiene procedures. And
35:57
one of them, just to give you an example, Well,
36:00
the most obvious one
36:02
is to ask several individuals
36:05
to make judgments independently because
36:07
that will reduce noise mechanically.
36:10
When you take several forecasters
36:12
and you average their forecast,
36:15
the average forecast is less
36:17
noisy than the individual forecast
36:19
and we know exactly by
36:22
what mathematical amount we
36:24
have cut down on the noise.
36:26
So that is another procedure. And
36:29
there are several others that Olivier,
36:31
I'm sure, can talk
36:33
about at least as well
36:35
as I can. Well, Mignon, just
36:37
to come back to guidelines for a second.
36:39
There is one field in which guidelines have
36:41
made a great difference and
36:44
that's medicine. You were talking,
36:46
as we started this conversation, about the
36:48
disease for which clearly there weren't guidelines
36:50
or if they're aware your three physicians
36:53
were not aware of them, sadly for
36:55
you. But in many fields,
36:57
guidelines have made a big difference. One
37:00
example that many people will have
37:02
encountered is that when a baby
37:04
is born, to determine if
37:06
the baby is healthy or needs to
37:08
be sent to neonatal care,
37:11
you use something called the
37:13
ABGAR score where you apply
37:15
five criteria, abbreviated A, P,
37:17
G, A, and R, and
37:20
you give this little baby that
37:22
is one minute or five minutes old a score
37:25
between zero and two on each of
37:27
those five criteria. And if the total
37:29
is six or less, the
37:32
baby has a problem. If the total is seven
37:34
or more, the baby is healthy. And
37:36
that has reduced massively the noise
37:39
and therefore the errors in the
37:41
treatment of newborn
37:43
babies. It's a great example
37:46
of a guideline that actually works. It's
37:48
a fairly simple guideline, but it's not
37:50
something one-dimensional like
37:53
a minimum sentencing guideline for
37:55
a particular crime. It takes into account
37:57
multiple factors, but it makes sure that
37:59
different... people will take the same factors into
38:01
account and will take them into account in the
38:03
same way so it reduces noise. Those
38:06
kinds of guidelines, when
38:08
they're well thought through, can actually
38:10
make a big difference. Okay.
38:14
So you also mentioned a
38:16
noise audit briefly in the last
38:18
segment there. Professor Ciboney, how
38:20
would you define what a noise audit is?
38:23
So a noise audit is not a way
38:25
to reduce noise. It's what you need to
38:28
do first. It's a way to measure noise.
38:30
So when we gave the
38:32
example of the underwriters or
38:34
the example of the justice system, these
38:36
are noise audits where you get a
38:38
feel for how large noise is. And
38:40
the reason you need to do that is that, as
38:44
Denny was pointing out, we don't imagine
38:46
that people see the world differently from
38:48
how we see it and therefore we
38:50
can't imagine that there is as much
38:52
noise as there is because
38:54
if I'm a judge, I never
38:56
hear what another judge would have
38:58
sentenced this particular defendant to because
39:00
each defendant is unique. And
39:02
if I'm a doctor and I look at an
39:05
x-ray, I never imagine that another doctor looking at
39:07
the same x-ray would see
39:09
something different from what I see. So a noise
39:11
audit makes this visible and tells you exactly how
39:13
much noise there is in your system. So
39:17
those are just a small
39:20
taste of the quite extensive
39:23
writing that you have in the book
39:25
about ways to reduce, to
39:28
know about, assess and reduce
39:30
noise in various systems and
39:33
organizations. But I'd like to just push
39:35
to one potential solution
39:38
and get both
39:41
your opinions on it. And that is
39:43
if you want to reduce variability entirely
39:45
on wanted variability,
39:48
you take the human condition out of it.
39:51
I mean, so of course I'm talking about
39:53
technology. People are actively trying to create AI
39:57
systems that achieve exactly what
39:59
you're talking about. about, take various
40:01
inputs and come up with the
40:03
same solution every single time. Professor
40:06
Kahneman, is that a desirable
40:08
way to reduce noise? Well,
40:11
in some, you know, there
40:15
have been many studies that
40:17
compared rules and algorithms to
40:19
human judgment and in
40:21
many of these studies human judgment comes
40:23
short and one of the
40:25
main reasons and that we know for that
40:28
humans come short is
40:31
because of noise because humans are noisy
40:33
and algorithms are not. You present the
40:35
same problem to an algorithm twice and
40:37
you get the same answer which is
40:39
not the case when you
40:42
do it with humans. So
40:45
we can expect that
40:47
algorithms when the information
40:49
is codeable. So there are some
40:51
conditions for algorithm to work well.
40:54
You need codeable information, you need
40:56
a lot of data and you
40:58
need a choice
41:02
about the criteria that you're applying
41:05
so as to eliminate bias to the
41:07
extent possible and then you
41:09
can have a system that is likely
41:11
to do better than humans and in
41:13
the judicial system there is an example
41:17
and the example is the
41:19
granting of bail where a
41:21
recent study using AI techniques,
41:25
I forget the number of millions
41:27
of cases that they looked at
41:29
but it's a very, very large
41:31
number, they were able to establish
41:33
that an algorithm would actually perform
41:36
better than the judicial system in
41:38
the sense that it would both
41:40
reduce crime and reduce
41:42
unnecessary and unjustified incarceration. So at
41:44
least in that domain there is
41:47
clear evidence that an algorithm can
41:49
do better than people. If the
41:51
algorithm has been eliminated of the
41:54
potential biases in its creation,
41:56
right, because I mean sticking with the
41:58
judicial system, I was reading several... several
42:00
years ago about how, for example, in
42:02
Washington, D.C. and this is also being
42:04
used everywhere, you were talking about bail
42:06
and this is actually regarding the AI
42:08
use in parole,
42:11
that prosecutors were
42:14
using an AI assessment system
42:17
to decide whether or not
42:19
to put parole on
42:21
the table for a particular defendant.
42:26
And defense lawyers had discovered
42:28
that that AI system was
42:30
making risk assessments based on
42:32
factors that included whether
42:34
or not a person lived in government subsidized
42:36
housing or whether they had
42:39
ever expressed negative attitudes about
42:41
the police. And
42:43
it seemed that there was ample opportunity
42:45
for bias to actually be
42:48
built into that AI system, which
42:50
was a problem. Absolutely. Absolutely.
42:53
No question. No
42:55
question about that and that's something to
42:57
be absolutely worried about. But
43:00
just to be clear, one biased algorithm
43:02
or two or ten do not
43:05
mean that all algorithms must be biased.
43:09
And algorithms have a big advantage over
43:11
humans, which is that, again, we can
43:13
have that conversation. We can measure whether
43:15
an algorithm is biased. We can have
43:18
ProPublica audit the algorithm and tell you
43:20
that the algorithm has this particular
43:23
bias or does not have that particular bias
43:25
and then it can get fixed. Of
43:27
course, that must happen. It
43:29
doesn't happen by magic. It takes action
43:31
from people who worry about it
43:34
and who make sure that algorithms improve.
43:37
But at least we can have that conversation.
43:39
A biased judge, a biased
43:41
human being is very difficult to
43:44
spot because of the noise
43:46
in the judgment, in part because of the
43:48
noise in the judgments of that person. No
43:51
judge is so consistently biased
43:54
that you would be able to indict
43:58
that particular judge for being biased. Yeah,
44:01
looks like a little internet instability
44:03
there. But Daniel
44:05
Kahneman, there's something I've been wanting
44:07
to ask you all our here because we're
44:10
talking about how noise can be
44:13
proliferated and amplified through a
44:16
system. But of course
44:18
that system is made up of individual
44:21
human beings. And I
44:23
wanted to hear from
44:25
you about how this actually does
44:27
connect to your previous
44:29
pathbreaking research about individual
44:32
judgment. Are there things
44:35
that individuals can do regarding their
44:37
thinking to reduce their contribution to
44:40
the noise? Well,
44:45
our idea on this matter is
44:47
really quite straightforward. And
44:50
it applies even to individuals' decisions,
44:52
not only to systems. It
44:54
applies to singular decisions, to
44:57
strategic decisions that people make.
44:59
And our argument is quite
45:01
straightforward. If decision
45:03
hygiene procedures work in repeated
45:05
cases, there is every reason
45:07
to believe that they will
45:09
apply as well to unique
45:12
or singular cases. So
45:14
decision hygiene recommendations
45:18
are applicable to any judgments
45:21
that people make. On
45:24
the argument, which actually Olivier,
45:26
the one who had that phrase that
45:29
we're very grateful for, that the singular
45:32
event is a repeated event that
45:35
happens only once so that everything
45:37
that we say about noise as
45:39
repeated events is actually applicable to
45:42
individual judgments. But
45:45
more specifically, from your book,
45:47
Thinking Fast and Slow, where you describe the
45:49
different types of thinking, aren't
45:53
there certain types of thinking that
45:55
achieve exactly what you're saying
45:57
that there are actually people who maybe
45:59
intuitively... are noise reducers?
46:04
We doubt that because and
46:06
the reason that we doubt that
46:08
and the connection with thinking fast
46:10
and slow is that
46:12
intuition is very rapid. Intuition
46:16
doesn't make efficient use of all
46:18
the information. So our
46:20
major recommendation in that context and
46:22
it's an important one is
46:25
that intuition should not be eliminated from
46:27
judgment but it should be delayed. That
46:30
is you want to not to
46:32
have a global intuition about a
46:34
case until you have made separate
46:37
judgments until you have all the
46:39
information whereas the human tendency is
46:41
to jump to conclusions. Jumping
46:44
to conclusions induces noise.
46:47
Well we only have a few minutes left. It
46:50
saddens my heart because I have so many more questions
46:52
for both of you but there's one more system I
46:55
wish to explore with you just briefly and
46:58
that is governance or political
47:00
systems and particularly let's just look
47:02
at the United States because
47:04
I feel like we are in a moment where
47:10
noise is the point. We
47:13
have very influential people in our
47:15
political system who have said, I'm
47:17
thinking of Steve Bannon for example
47:19
who said that his goal was
47:22
to flood the zone with
47:24
BS essentially. Is
47:29
there anything in your
47:31
book that we as citizens can
47:34
apply to reducing the noise and
47:37
improving the decision making in a
47:40
political system, in the American political system?
47:45
I don't think there is anything specific
47:48
that is to be applied. If
47:50
people thought better in
47:52
general and make better judgments
47:54
and better decisions we
47:56
might be better off but the
47:59
differences in the political system are
48:01
closer to issues of bias than to
48:03
issues of noise. And
48:05
bias and convictions and
48:08
convictions based on very little
48:10
evidence and on poor
48:12
evidence, those are political
48:15
problems. And to those,
48:17
we have no solution to offer
48:19
that I know of. Perhaps Olivier
48:21
can think of something, but I
48:23
have not. No, unfortunately
48:26
not. There is one
48:28
thing though, which is not a solution, but
48:30
which is part of the problem that we discuss
48:32
at some length in the book, which is that
48:35
groups and any forum,
48:37
including social media, in which people are
48:39
going to interact in a group, tend
48:42
to amplify the random noise that comes from
48:44
the opinion of a few people at the
48:46
beginning of the process. So any
48:50
system, and I'm thinking mostly of social
48:52
media, in which people are going to
48:54
be part of an eco
48:57
chamber is going to
48:59
add to the randomness in
49:01
the positions that people have eventually and is
49:03
going to add to the polarization of those
49:05
positions. Well if
49:07
Olivier Ciboney and Daniel Kahneman
49:09
especially don't necessarily have a
49:12
solution for the chaos
49:14
inducing noise in our political system,
49:17
I don't know who would, but
49:19
they along with Cass Sunstein are
49:21
authors of the new book Noise, a
49:23
flaw in human judgment, and we have
49:26
an excerpt of it at onpointradio.org. Professor
49:28
Kahneman, it's been a real pleasure to
49:31
speak with you. Thank you so much. My
49:33
pleasure. And Professor Ciboney likewise, thank you
49:36
so much for joining us. Thanks, Eumigna.
49:38
I'm Eumigna Chakrabarti. This is On Point. Thank
49:42
you.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More