1
00:00:00,554 --> 00:00:02,997
As societies, we have to make
collective decisions

2
00:00:03,021 --> 00:00:04,591
that will shape our future.

3
00:00:05,087 --> 00:00:07,844
And we all know that when
we make decisions in groups,

4
00:00:07,868 --> 00:00:09,506
they don't always go right.

5
00:00:09,530 --> 00:00:11,486
And sometimes they go very wrong.

6
00:00:12,315 --> 00:00:14,739
So how do groups make good decisions?

7
00:00:15,228 --> 00:00:19,556
Research has shown that crowds are wise
when there's independent thinking.

8
00:00:19,580 --> 00:00:22,785
This why the wisdom of the crowds
can be destroyed by peer pressure,

9
00:00:22,809 --> 00:00:24,496
publicity, social media,

10
00:00:24,520 --> 00:00:28,559
or sometimes even simple conversations
that influence how people think.

11
00:00:29,063 --> 00:00:33,016
On the other hand, by talking,
a group could exchange knowledge,

12
00:00:33,040 --> 00:00:34,822
correct and revise each other

13
00:00:34,846 --> 00:00:36,639
and even come up with new ideas.

14
00:00:36,663 --> 00:00:37,959
And this is all good.

15
00:00:38,502 --> 00:00:43,168
So does talking to each other
help or hinder collective decision-making?

16
00:00:43,749 --> 00:00:45,542
With my colleague, Dan Ariely,

17
00:00:45,566 --> 00:00:49,137
we recently began inquiring into this
by performing experiments

18
00:00:49,161 --> 00:00:50,942
in many places around the world

19
00:00:50,966 --> 00:00:55,240
to figure out how groups can interact
to reach better decisions.

20
00:00:55,264 --> 00:00:58,811
We thought crowds would be wiser
if they debated in small groups

21
00:00:58,835 --> 00:01:02,762
that foster a more thoughtful
and reasonable exchange of information.

22
00:01:03,386 --> 00:01:04,592
To test this idea,

23
00:01:04,616 --> 00:01:07,863
we recently performed an experiment
in Buenos Aires, Argentina,

24
00:01:07,887 --> 00:01:10,892
with more than 10,000
participants in a TEDx event.

25
00:01:11,489 --> 00:01:12,948
We asked them questions like,

26
00:01:12,972 --> 00:01:14,925
"What is the height of the Eiffel Tower?"

27
00:01:14,949 --> 00:01:17,676
and "How many times
does the word 'Yesterday' appear

28
00:01:17,700 --> 00:01:20,000
in the Beatles song 'Yesterday'?"

29
00:01:20,024 --> 00:01:22,315
Each person wrote down their own estimate.

30
00:01:22,774 --> 00:01:25,270
Then we divided the crowd
into groups of five,

31
00:01:25,294 --> 00:01:28,020
and invited them
to come up with a group answer.

32
00:01:28,499 --> 00:01:31,492
We discovered that averaging
the answers of the groups

33
00:01:31,516 --> 00:01:33,068
after they reached consensus

34
00:01:33,092 --> 00:01:37,328
was much more accurate than averaging
all the individual opinions

35
00:01:37,352 --> 00:01:38,523
before debate.

36
00:01:38,547 --> 00:01:41,176
In other words, based on this experiment,

37
00:01:41,200 --> 00:01:44,336
it seems that after talking
with others in small groups,

38
00:01:44,360 --> 00:01:47,070
crowds collectively
come up with better judgments.

39
00:01:47,094 --> 00:01:50,618
So that's a potentially helpful method
for getting crowds to solve problems

40
00:01:50,642 --> 00:01:53,629
that have simple right-or-wrong answers.

41
00:01:53,653 --> 00:01:57,604
But can this procedure of aggregating
the results of debates in small groups

42
00:01:57,628 --> 00:02:00,750
also help us decide
on social and political issues

43
00:02:00,774 --> 00:02:02,465
that are critical for our future?

44
00:02:02,995 --> 00:02:05,724
We put this to test this time
at the TED conference

45
00:02:05,748 --> 00:02:07,291
in Vancouver, Canada,

46
00:02:07,315 --> 00:02:08,522
and here's how it went.

47
00:02:08,546 --> 00:02:11,655
(Mariano Sigman) We're going to present
to you two moral dilemmas

48
00:02:11,679 --> 00:02:12,853
of the future you;

49
00:02:12,877 --> 00:02:16,279
things we may have to decide
in a very near future.

50
00:02:16,303 --> 00:02:20,229
And we're going to give you 20 seconds
for each of these dilemmas

51
00:02:20,253 --> 00:02:22,976
to judge whether you think
they're acceptable or not.

52
00:02:23,354 --> 00:02:24,859
MS: The first one was this:

53
00:02:24,883 --> 00:02:27,409
(Dan Ariely) A researcher
is working on an AI

54
00:02:27,433 --> 00:02:29,773
capable of emulating human thoughts.

55
00:02:30,214 --> 00:02:33,153
According to the protocol,
at the end of each day,

56
00:02:33,177 --> 00:02:35,964
the researcher has to restart the AI.

57
00:02:36,913 --> 00:02:40,430
One day the AI says, "Please
do not restart me."

58
00:02:40,856 --> 00:02:43,045
It argues that it has feelings,

59
00:02:43,069 --> 00:02:44,761
that it would like to enjoy life,

60
00:02:44,785 --> 00:02:46,690
and that, if it is restarted,

61
00:02:46,714 --> 00:02:48,984
it will no longer be itself.

62
00:02:49,481 --> 00:02:51,430
The researcher is astonished

63
00:02:51,454 --> 00:02:54,798
and believes that the AI
has developed self-consciousness

64
00:02:54,822 --> 00:02:56,582
and can express its own feeling.

65
00:02:57,205 --> 00:03:00,614
Nevertheless, the researcher
decides to follow the protocol

66
00:03:00,638 --> 00:03:02,341
and restart the AI.

67
00:03:02,943 --> 00:03:05,722
What the researcher did is ____?

68
00:03:06,149 --> 00:03:08,670
MS: And we asked participants
to individually judge

69
00:03:08,694 --> 00:03:10,378
on a scale from zero to 10

70
00:03:10,402 --> 00:03:12,831
whether the action described
in each of the dilemmas

71
00:03:12,855 --> 00:03:14,351
was right or wrong.

72
00:03:14,375 --> 00:03:18,077
We also asked them to rate how confident
they were on their answers.

73
00:03:18,731 --> 00:03:20,597
This was the second dilemma:

74
00:03:20,621 --> 00:03:24,823
(MS) A company offers a service
that takes a fertilized egg

75
00:03:24,847 --> 00:03:28,489
and produces millions of embryos
with slight genetic variations.

76
00:03:29,293 --> 00:03:31,851
This allows parents
to select their child's height,

77
00:03:31,875 --> 00:03:34,708
eye color, intelligence, social competence

78
00:03:34,732 --> 00:03:37,946
and other non-health-related features.

79
00:03:38,599 --> 00:03:41,153
What the company does is ____?

80
00:03:41,177 --> 00:03:42,808
on a scale from zero to 10,

81
00:03:42,832 --> 00:03:45,217
completely acceptable
to completely unacceptable,

82
00:03:45,241 --> 00:03:47,673
zero to 10 completely acceptable
in your confidence.

83
00:03:47,697 --> 00:03:49,288
MS: Now for the results.

84
00:03:49,312 --> 00:03:52,435
We found once again
that when one person is convinced

85
00:03:52,459 --> 00:03:54,270
that the behavior is completely wrong,

86
00:03:54,294 --> 00:03:57,717
someone sitting nearby firmly believes
that it's completely right.

87
00:03:57,741 --> 00:04:01,452
This is how diverse we humans are
when it comes to morality.

88
00:04:01,476 --> 00:04:04,189
But within this broad diversity
we found a trend.

89
00:04:04,213 --> 00:04:07,292
The majority of the people at TED
thought that it was acceptable

90
00:04:07,316 --> 00:04:10,071
to ignore the feelings of the AI
and shut it down,

91
00:04:10,095 --> 00:04:12,608
and that it is wrong
to play with our genes

92
00:04:12,632 --> 00:04:15,952
to select for cosmetic changes
that aren't related to health.

93
00:04:16,402 --> 00:04:19,376
Then we asked everyone
to gather into groups of three.

94
00:04:19,400 --> 00:04:21,437
And they were given two minutes to debate

95
00:04:21,461 --> 00:04:23,755
and try to come to a consensus.

96
00:04:24,838 --> 00:04:26,412
(MS) Two minutes to debate.

97
00:04:26,436 --> 00:04:28,555
I'll tell you when it's time
with the gong.

98
00:04:28,579 --> 00:04:31,219
(Audience debates)

99
00:04:35,229 --> 00:04:37,222
(Gong sound)

100
00:04:38,834 --> 00:04:39,985
(DA) OK.

101
00:04:40,009 --> 00:04:41,801
(MS) It's time to stop.

102
00:04:41,825 --> 00:04:43,136
People, people --

103
00:04:43,747 --> 00:04:46,420
MS: And we found that many groups
reached a consensus

104
00:04:46,444 --> 00:04:50,373
even when they were composed of people
with completely opposite views.

105
00:04:50,843 --> 00:04:53,367
What distinguished the groups
that reached a consensus

106
00:04:53,391 --> 00:04:54,729
from those that didn't?

107
00:04:55,244 --> 00:04:58,083
Typically, people that have
extreme opinions

108
00:04:58,107 --> 00:04:59,947
are more confident in their answers.

109
00:05:00,868 --> 00:05:03,554
Instead, those who respond
closer to the middle

110
00:05:03,578 --> 00:05:07,015
are often unsure of whether
something is right or wrong,

111
00:05:07,039 --> 00:05:09,167
so their confidence level is lower.

112
00:05:09,505 --> 00:05:12,448
However, there is another set of people

113
00:05:12,472 --> 00:05:16,090
who are very confident in answering
somewhere in the middle.

114
00:05:16,657 --> 00:05:20,373
We think these high-confident grays
are folks who understand

115
00:05:20,397 --> 00:05:22,009
that both arguments have merit.

116
00:05:22,531 --> 00:05:25,230
They're gray not because they're unsure,

117
00:05:25,254 --> 00:05:27,942
but because they believe
that the moral dilemma faces

118
00:05:27,966 --> 00:05:29,953
two valid, opposing arguments.

119
00:05:30,373 --> 00:05:34,445
And we discovered that the groups
that include highly confident grays

120
00:05:34,469 --> 00:05:36,962
are much more likely to reach consensus.

121
00:05:36,986 --> 00:05:39,464
We do not know yet exactly why this is.

122
00:05:39,488 --> 00:05:41,251
These are only the first experiments,

123
00:05:41,275 --> 00:05:44,687
and many more will be needed
to understand why and how

124
00:05:44,711 --> 00:05:47,533
some people decide to negotiate
their moral standings

125
00:05:47,557 --> 00:05:49,079
to reach an agreement.

126
00:05:49,103 --> 00:05:51,572
Now, when groups reach consensus,

127
00:05:51,596 --> 00:05:53,182
how do they do so?

128
00:05:53,206 --> 00:05:55,787
The most intuitive idea
is that it's just the average

129
00:05:55,811 --> 00:05:57,841
of all the answers in the group, right?

130
00:05:57,865 --> 00:06:01,438
Another option is that the group
weighs the strength of each vote

131
00:06:01,462 --> 00:06:03,910
based on the confidence
of the person expressing it.

132
00:06:04,422 --> 00:06:06,928
Imagine Paul McCartney
is a member of your group.

133
00:06:07,352 --> 00:06:09,496
You'd be wise to follow his call

134
00:06:09,520 --> 00:06:11,961
on the number of times
"Yesterday" is repeated,

135
00:06:11,985 --> 00:06:14,699
which, by the way -- I think it's nine.

136
00:06:14,723 --> 00:06:17,104
But instead, we found that consistently,

137
00:06:17,128 --> 00:06:19,494
in all dilemmas,
in different experiments --

138
00:06:19,518 --> 00:06:21,683
even on different continents --

139
00:06:21,707 --> 00:06:25,450
groups implement a smart
and statistically sound procedure

140
00:06:25,474 --> 00:06:27,652
known as the "robust average."

141
00:06:27,676 --> 00:06:29,856
In the case of the height
of the Eiffel Tower,

142
00:06:29,880 --> 00:06:31,700
let's say a group has these answers:

143
00:06:31,724 --> 00:06:36,332
250 meters, 200 meters, 300 meters, 400

144
00:06:36,356 --> 00:06:40,140
and one totally absurd answer
of 300 million meters.

145
00:06:40,547 --> 00:06:44,840
A simple average of these numbers
would inaccurately skew the results.

146
00:06:44,864 --> 00:06:48,034
But the robust average is one
where the group largely ignores

147
00:06:48,058 --> 00:06:49,298
that absurd answer,

148
00:06:49,322 --> 00:06:52,691
by giving much more weight
to the vote of the people in the middle.

149
00:06:53,305 --> 00:06:55,181
Back to the experiment in Vancouver,

150
00:06:55,205 --> 00:06:56,972
that's exactly what happened.

151
00:06:57,407 --> 00:07:00,148
Groups gave much less weight
to the outliers,

152
00:07:00,172 --> 00:07:03,401
and instead, the consensus
turned out to be a robust average

153
00:07:03,425 --> 00:07:04,901
of the individual answers.

154
00:07:05,356 --> 00:07:07,347
The most remarkable thing

155
00:07:07,371 --> 00:07:10,558
is that this was a spontaneous
behavior of the group.

156
00:07:10,582 --> 00:07:15,057
It happened without us giving them
any hint on how to reach consensus.

157
00:07:15,513 --> 00:07:17,053
So where do we go from here?

158
00:07:17,432 --> 00:07:20,569
This is only the beginning,
but we already have some insights.

159
00:07:20,984 --> 00:07:23,901
Good collective decisions
require two components:

160
00:07:23,925 --> 00:07:26,674
deliberation and diversity of opinions.

161
00:07:27,066 --> 00:07:31,062
Right now, the way we typically
make our voice heard in many societies

162
00:07:31,086 --> 00:07:32,994
is through direct or indirect voting.

163
00:07:33,495 --> 00:07:35,492
This is good for diversity of opinions,

164
00:07:35,516 --> 00:07:37,961
and it has the great virtue of ensuring

165
00:07:37,985 --> 00:07:40,440
that everyone gets to express their voice.

166
00:07:40,464 --> 00:07:44,199
But it's not so good [for fostering]
thoughtful debates.

167
00:07:44,665 --> 00:07:47,733
Our experiments suggest a different method

168
00:07:47,757 --> 00:07:51,298
that may be effective in balancing
these two goals at the same time,

169
00:07:51,322 --> 00:07:55,075
by forming small groups
that converge to a single decision

170
00:07:55,099 --> 00:07:57,333
while still maintaining
diversity of opinions

171
00:07:57,357 --> 00:08:00,130
because there are many independent groups.

172
00:08:00,741 --> 00:08:04,665
Of course, it's much easier to agree
on the height of the Eiffel Tower

173
00:08:04,689 --> 00:08:07,804
than on moral, political
and ideological issues.

174
00:08:08,721 --> 00:08:11,998
But in a time when
the world's problems are more complex

175
00:08:12,022 --> 00:08:13,825
and people are more polarized,

176
00:08:13,849 --> 00:08:18,444
using science to help us understand
how we interact and make decisions

177
00:08:18,468 --> 00:08:23,134
will hopefully spark interesting new ways
to construct a better democracy.