1
00:00:00,881 --> 00:00:02,111
I have a question.

2
00:00:03,422 --> 00:00:05,365
Can a computer write poetry?

3
00:00:06,959 --> 00:00:09,036
This is a provocative question.

4
00:00:09,715 --> 00:00:11,433
You think about it for a minute,

5
00:00:11,457 --> 00:00:14,047
and you suddenly have a bunch
of other questions like:

6
00:00:14,769 --> 00:00:16,150
What is a computer?

7
00:00:16,710 --> 00:00:18,285
What is poetry?

8
00:00:18,707 --> 00:00:20,396
What is creativity?

9
00:00:21,650 --> 00:00:22,822
But these are questions

10
00:00:22,846 --> 00:00:25,916
that people spend their entire
lifetime trying to answer,

11
00:00:25,940 --> 00:00:28,164
not in a single TED Talk.

12
00:00:28,188 --> 00:00:30,633
So we're going to have to try
a different approach.

13
00:00:30,657 --> 00:00:32,800
So up here, we have two poems.

14
00:00:33,839 --> 00:00:36,115
One of them is written by a human,

15
00:00:36,139 --> 00:00:38,241
and the other one's written by a computer.

16
00:00:38,754 --> 00:00:41,164
I'm going to ask you to tell me
which one's which.

17
00:00:41,858 --> 00:00:43,014
Have a go:

18
00:00:43,038 --> 00:00:47,094
Poem 1: Little Fly / Thy summer's play, /
My thoughtless hand / Has brush'd away.

19
00:00:47,118 --> 00:00:50,512
Am I not / A fly like thee? /
Or art not thou / A man like me?

20
00:00:50,536 --> 00:00:53,835
Poem 2: We can feel / Activist
through your life's / morning /

21
00:00:53,859 --> 00:00:58,106
Pauses to see, pope I hate the / Non
all the night to start a / great otherwise (...)

22
00:00:58,130 --> 00:00:59,489
Alright, time's up.

23
00:00:59,513 --> 00:01:03,609
Hands up if you think Poem 1
was written by a human.

24
00:01:05,547 --> 00:01:07,037
OK, most of you.

25
00:01:07,061 --> 00:01:10,084
Hands up if you think Poem 2
was written by a human.

26
00:01:11,172 --> 00:01:12,362
Very brave of you,

27
00:01:12,855 --> 00:01:17,140
because the first one was written
by the human poet William Blake.

28
00:01:17,784 --> 00:01:20,733
The second one was written by an algorithm

29
00:01:20,757 --> 00:01:24,449
that took all the language
from my Facebook feed on one day

30
00:01:24,473 --> 00:01:27,236
and then regenerated it algorithmically,

31
00:01:27,260 --> 00:01:30,850
according to methods that I'll describe
a little bit later on.

32
00:01:31,218 --> 00:01:33,622
So let's try another test.

33
00:01:34,398 --> 00:01:36,491
Again, you haven't got ages to read this,

34
00:01:36,515 --> 00:01:38,127
so just trust your gut.

35
00:01:38,151 --> 00:01:42,196
Poem 1: A lion roars and a dog barks.
It is interesting / and fascinating

36
00:01:42,220 --> 00:01:46,523
that a bird will fly and not / roar
or bark. Enthralling stories about animals

37
00:01:46,547 --> 00:01:50,607
are in my dreams and I will sing them all
if I / am not exhausted or weary.

38
00:01:50,631 --> 00:01:54,616
Poem 2: Oh! kangaroos, sequins, chocolate
sodas! / You are really beautiful!

39
00:01:54,640 --> 00:01:58,998
Pearls, / harmonicas, jujubes, aspirins!
All / the stuff they've always talked about (...)

40
00:01:59,022 --> 00:02:00,180
Alright, time's up.

41
00:02:00,204 --> 00:02:03,341
So if you think the first poem
was written by a human,

42
00:02:03,365 --> 00:02:04,580
put your hand up.

43
00:02:05,687 --> 00:02:06,841
OK.

44
00:02:06,865 --> 00:02:09,540
And if you think the second poem
was written by a human,

45
00:02:09,564 --> 00:02:10,719
put your hand up.

46
00:02:11,779 --> 00:02:15,589
We have, more or less, a 50/50 split here.

47
00:02:16,157 --> 00:02:17,593
It was much harder.

48
00:02:17,617 --> 00:02:19,329
The answer is,

49
00:02:19,353 --> 00:02:22,836
the first poem was generated
by an algorithm called Racter,

50
00:02:22,860 --> 00:02:25,862
that was created back in the 1970s,

51
00:02:25,886 --> 00:02:29,075
and the second poem was written
by a guy called Frank O'Hara,

52
00:02:29,099 --> 00:02:31,767
who happens to be
one of my favorite human poets.

53
00:02:32,631 --> 00:02:35,689
(Laughter)

54
00:02:36,046 --> 00:02:39,274
So what we've just done now
is a Turing test for poetry.

55
00:02:40,018 --> 00:02:44,565
The Turing test was first proposed
by this guy, Alan Turing, in 1950,

56
00:02:44,589 --> 00:02:46,153
in order to answer the question,

57
00:02:46,177 --> 00:02:47,814
can computers think?

58
00:02:48,245 --> 00:02:51,015
Alan Turing believed that if
a computer was able

59
00:02:51,039 --> 00:02:54,117
to have a to have a text-based
conversation with a human,

60
00:02:54,141 --> 00:02:56,911
with such proficiency
such that the human couldn't tell

61
00:02:56,935 --> 00:02:59,901
whether they are talking
to a computer or a human,

62
00:02:59,925 --> 00:03:02,781
then the computer can be said
to have intelligence.

63
00:03:03,270 --> 00:03:06,565
So in 2013, my friend
Benjamin Laird and I,

64
00:03:06,589 --> 00:03:09,577
we created a Turing test
for poetry online.

65
00:03:09,601 --> 00:03:10,878
It's called bot or not,

66
00:03:10,902 --> 00:03:12,946
and you can go and play it for yourselves.

67
00:03:12,970 --> 00:03:15,221
But basically, it's the game
we just played.

68
00:03:15,245 --> 00:03:16,773
You're presented with a poem,

69
00:03:16,797 --> 00:03:19,825
you don't know whether it was written
by a human or a computer

70
00:03:19,849 --> 00:03:21,015
and you have to guess.

71
00:03:21,039 --> 00:03:24,230
So thousands and thousands
of people have taken this test online,

72
00:03:24,254 --> 00:03:25,703
so we have results.

73
00:03:25,727 --> 00:03:27,155
And what are the results?

74
00:03:27,704 --> 00:03:30,583
Well, Turing said that if a computer
could fool a human

75
00:03:30,607 --> 00:03:33,626
30 percent of the time
that it was a human,

76
00:03:33,650 --> 00:03:36,047
then it passes the Turing test
for intelligence.

77
00:03:36,625 --> 00:03:39,063
We have poems on the bot or not database

78
00:03:39,087 --> 00:03:42,066
that have fooled 65 percent
of human readers into thinking

79
00:03:42,090 --> 00:03:43,485
it was written by a human.

80
00:03:43,959 --> 00:03:46,776
So, I think we have an answer
to our question.

81
00:03:47,546 --> 00:03:49,894
According to the logic of the Turing test,

82
00:03:49,918 --> 00:03:51,846
can a computer write poetry?

83
00:03:51,870 --> 00:03:54,221
Well, yes, absolutely it can.

84
00:03:55,782 --> 00:03:58,128
But if you're feeling
a little bit uncomfortable

85
00:03:58,152 --> 00:04:00,079
with this answer, that's OK.

86
00:04:00,103 --> 00:04:02,419
If you're having a bunch
of gut reactions to it,

87
00:04:02,443 --> 00:04:05,648
that's also OK because
this isn't the end of the story.

88
00:04:06,594 --> 00:04:08,918
Let's play our third and final test.

89
00:04:10,000 --> 00:04:11,750
Again, you're going to have to read

90
00:04:11,774 --> 00:04:13,683
and tell me which you think is human.

91
00:04:13,707 --> 00:04:17,425
Poem 1: Red flags the reason
for pretty flags. / And ribbons.

92
00:04:17,449 --> 00:04:21,770
Ribbons of flags / And wearing material /
Reasons for wearing material. (...)

93
00:04:21,794 --> 00:04:25,712
Poem 2: A wounded deer leaps
highest, / I've heard the daffodil

94
00:04:25,736 --> 00:04:29,182
I've heard the flag to-day /
I've heard the hunter tell; /

95
00:04:29,206 --> 00:04:32,908
'Tis but the ecstasy of death, /
And then the brake is almost done (...)

96
00:04:32,932 --> 00:04:34,531
OK, time is up.

97
00:04:34,555 --> 00:04:38,392
So hands up if you think Poem 1
was written by a human.

98
00:04:39,973 --> 00:04:43,011
Hands up if you think Poem 2
was written by a human.

99
00:04:43,035 --> 00:04:45,366
Whoa, that's a lot more people.

100
00:04:46,327 --> 00:04:49,295
So you'd be surprised to find that Poem 1

101
00:04:49,319 --> 00:04:53,312
was written by the very
human poet Gertrude Stein.

102
00:04:54,100 --> 00:04:59,138
And Poem 2 was generated
by an algorithm called RKCP.

103
00:04:59,162 --> 00:05:02,481
Now before we go on, let me describe
very quickly and simply,

104
00:05:02,505 --> 00:05:04,286
how RKCP works.

105
00:05:04,873 --> 00:05:08,723
So RKCP is an algorithm
designed by Ray Kurzweil,

106
00:05:08,747 --> 00:05:10,969
who's a director of engineering at Google

107
00:05:10,993 --> 00:05:13,353
and a firm believer
in artificial intelligence.

108
00:05:13,822 --> 00:05:17,813
So, you give RKCP a source text,

109
00:05:17,837 --> 00:05:22,306
it analyzes the source text in order
to find out how it uses language,

110
00:05:22,330 --> 00:05:24,278
and then it regenerates language

111
00:05:24,302 --> 00:05:26,830
that emulates that first text.

112
00:05:26,854 --> 00:05:28,967
So in the poem we just saw before,

113
00:05:28,991 --> 00:05:31,616
Poem 2, the one that you all
thought was human,

114
00:05:31,640 --> 00:05:33,190
it was fed a bunch of poems

115
00:05:33,214 --> 00:05:35,249
by a poet called Emily Dickinson

116
00:05:35,273 --> 00:05:37,462
it looked at the way she used language,

117
00:05:37,486 --> 00:05:38,651
learned the model,

118
00:05:38,675 --> 00:05:42,933
and then it regenerated a model
according to that same structure.

119
00:05:44,732 --> 00:05:46,910
But the important thing to know about RKCP

120
00:05:46,934 --> 00:05:49,772
is that it doesn't know the meaning
of the words it's using.

121
00:05:50,359 --> 00:05:52,635
The language is just raw material,

122
00:05:52,659 --> 00:05:54,819
it could be Chinese,
it could be in Swedish,

123
00:05:54,843 --> 00:05:59,022
it could be the collected language
from your Facebook feed for one day.

124
00:05:59,046 --> 00:06:00,698
It's just raw material.

125
00:06:01,380 --> 00:06:04,077
And nevertheless, it's able
to create a poem

126
00:06:04,101 --> 00:06:07,428
that seems more human
than Gertrude Stein's poem,

127
00:06:07,452 --> 00:06:09,605
and Gertrude Stein is a human.

128
00:06:10,846 --> 00:06:14,918
So what we've done here is,
more or less, a reverse Turing test.

129
00:06:15,940 --> 00:06:21,119
So Gertrude Stein, who's a human,
is able to write a poem

130
00:06:21,143 --> 00:06:24,881
that fools a majority
of human judges into thinking

131
00:06:24,905 --> 00:06:26,731
that it was written by a computer.

132
00:06:27,176 --> 00:06:31,317
Therefore, according to the logic
of the reverse Turing test,

133
00:06:31,341 --> 00:06:33,257
Gertrude Stein is a computer.

134
00:06:33,281 --> 00:06:34,743
(Laughter)

135
00:06:35,358 --> 00:06:36,652
Feeling confused?

136
00:06:37,193 --> 00:06:38,708
I think that's fair enough.

137
00:06:39,546 --> 00:06:43,662
So far we've had humans
that write like humans,

138
00:06:43,686 --> 00:06:46,797
we have computers that write
like computers,

139
00:06:46,821 --> 00:06:49,876
we have computers that write like humans,

140
00:06:49,900 --> 00:06:53,532
but we also have,
perhaps most confusingly,

141
00:06:53,556 --> 00:06:55,931
humans that write like computers.

142
00:06:56,938 --> 00:06:58,704
So what do we take from all of this?

143
00:06:59,611 --> 00:07:02,768
Do we take that William Blake
is somehow more of a human

144
00:07:02,792 --> 00:07:04,041
than Gertrude Stein?

145
00:07:04,065 --> 00:07:07,111
Or that Gertrude Stein is more
of a computer than William Blake?

146
00:07:07,135 --> 00:07:08,687
(Laughter)

147
00:07:08,711 --> 00:07:11,034
These are questions
I've been asking myself

148
00:07:11,058 --> 00:07:12,523
for around two years now,

149
00:07:12,547 --> 00:07:14,856
and I don't have any answers.

150
00:07:14,880 --> 00:07:17,210
But what I do have are a bunch of insights

151
00:07:17,234 --> 00:07:19,768
about our relationship with technology.

152
00:07:20,999 --> 00:07:24,608
So my first insight is that,
for some reason,

153
00:07:24,632 --> 00:07:27,743
we associate poetry with being human.

154
00:07:28,197 --> 00:07:31,912
So that when we ask,
"Can a computer write poetry?"

155
00:07:31,936 --> 00:07:33,129
we're also asking,

156
00:07:33,153 --> 00:07:34,951
"What does it mean to be human

157
00:07:34,975 --> 00:07:38,147
and how do we put boundaries
around this category?

158
00:07:38,171 --> 00:07:41,829
How do we say who or what
can be part of this category?"

159
00:07:42,376 --> 00:07:45,727
This is an essentially
philosophical question, I believe,

160
00:07:45,751 --> 00:07:47,980
and it can't be answered
with a yes or no test,

161
00:07:48,004 --> 00:07:49,331
like the Turing test.

162
00:07:49,805 --> 00:07:52,850
I also believe that Alan Turing
understood this,

163
00:07:52,874 --> 00:07:56,179
and that when he devised
his test back in 1950,

164
00:07:56,203 --> 00:07:59,005
he was doing it
as a philosophical provocation.

165
00:08:01,124 --> 00:08:06,665
So my second insight is that,
when we take the Turing test for poetry,

166
00:08:06,689 --> 00:08:10,149
we're not really testing
the capacity of the computers

167
00:08:10,173 --> 00:08:13,066
because poetry-generating algorithms,

168
00:08:13,090 --> 00:08:17,653
they're pretty simple and have existed,
more or less, since the 1950s.

169
00:08:19,055 --> 00:08:22,173
What we are doing with the Turing
test for poetry, rather,

170
00:08:22,197 --> 00:08:26,812
is collecting opinions about what
constitutes humanness.

171
00:08:28,313 --> 00:08:31,042
So, what I've figured out,

172
00:08:31,066 --> 00:08:34,038
we've seen this when earlier today,

173
00:08:34,062 --> 00:08:36,540
we say that William Blake
is more of a human

174
00:08:36,564 --> 00:08:38,129
than Gertrude Stein.

175
00:08:38,153 --> 00:08:40,615
Of course, this doesn't mean
that William Blake

176
00:08:40,639 --> 00:08:42,467
was actually more human

177
00:08:42,491 --> 00:08:44,818
or that Gertrude Stein
was more of a computer.

178
00:08:45,533 --> 00:08:50,247
It simply means that the category
of the human is unstable.

179
00:08:51,450 --> 00:08:53,524
This has led me to understand

180
00:08:53,548 --> 00:08:56,311
that the human is not a cold, hard fact.

181
00:08:56,832 --> 00:08:59,964
Rather, it is something
that's constructed with our opinions

182
00:08:59,988 --> 00:09:02,843
and something that changes over time.

183
00:09:04,671 --> 00:09:09,150
So my final insight is that
the computer, more or less,

184
00:09:09,174 --> 00:09:13,180
works like a mirror
that reflects any idea of a human

185
00:09:13,204 --> 00:09:14,579
that we show it.

186
00:09:14,958 --> 00:09:16,842
We show it Emily Dickinson,

187
00:09:16,866 --> 00:09:19,187
it gives Emily Dickinson back to us.

188
00:09:19,768 --> 00:09:21,602
We show it William Blake,

189
00:09:21,626 --> 00:09:23,911
that's what it reflects back to us.

190
00:09:23,935 --> 00:09:25,774
We show it Gertrude Stein,

191
00:09:25,798 --> 00:09:28,268
what we get back is Gertrude Stein.

192
00:09:29,083 --> 00:09:31,451
More than any other bit of technology,

193
00:09:31,475 --> 00:09:36,640
the computer is a mirror that reflects
any idea of the human we teach it.

194
00:09:38,061 --> 00:09:40,348
So I'm sure a lot of you have been hearing

195
00:09:40,372 --> 00:09:43,234
a lot about artificial
intelligence recently.

196
00:09:44,694 --> 00:09:47,524
And much of the conversation is,

197
00:09:48,292 --> 00:09:49,481
can we build it?

198
00:09:50,383 --> 00:09:53,518
Can we build an intelligent computer?

199
00:09:53,542 --> 00:09:56,305
Can we build a creative computer?

200
00:09:56,329 --> 00:09:58,442
What we seem to be asking over and over

201
00:09:58,466 --> 00:10:01,190
is can we build a human-like computer?

202
00:10:01,961 --> 00:10:03,517
But what we've seen just now

203
00:10:03,541 --> 00:10:06,629
is that the human
is not a scientific fact,

204
00:10:06,653 --> 00:10:10,183
that it's an ever-shifting,
concatenating idea

205
00:10:10,207 --> 00:10:12,738
and one that changes over time.

206
00:10:12,762 --> 00:10:15,914
So that when we begin
to grapple with the ideas

207
00:10:15,938 --> 00:10:18,324
of artificial intelligence in the future,

208
00:10:18,348 --> 00:10:20,253
we shouldn't only be asking ourselves,

209
00:10:20,277 --> 00:10:21,645
"Can we build it?"

210
00:10:21,669 --> 00:10:23,563
But we should also be asking ourselves,

211
00:10:23,587 --> 00:10:27,300
"What idea of the human
do we want to have reflected back to us?"

212
00:10:27,820 --> 00:10:30,513
This is an essentially philosophical idea,

213
00:10:30,537 --> 00:10:33,534
and it's one that can't be answered
with software alone,

214
00:10:33,558 --> 00:10:38,535
but I think requires a moment
of species-wide, existential reflection.

215
00:10:39,040 --> 00:10:40,193
Thank you.

216
00:10:40,217 --> 00:10:42,912
(Applause)