1
00:00:00,607 --> 00:00:05,357
- Catch this video ad free
and uncensored only on Nebula.

2
00:00:06,478 --> 00:00:09,228
Is it going to destroy the world?

3
00:00:10,623 --> 00:00:13,812
That's the question on
a lot of people's lips.

4
00:00:13,812 --> 00:00:16,176
"AI is coming", they say.

5
00:00:16,176 --> 00:00:19,037
"It's inevitable", they say.

6
00:00:19,037 --> 00:00:23,370
We'd better teach it to be
good and hope it listens.

7
00:00:24,377 --> 00:00:26,294
Can we make ethical AI?

8
00:00:28,808 --> 00:00:31,141
The future may depend on it.

9
00:00:32,199 --> 00:00:43,555
[intense music like NEEOOWW NEEEEOOOW electric guitars and explosions and stuff let's go it's Philosophy Tube time baybee]

10
00:00:43,555 --> 00:00:46,790
The AI people are most worried about

11
00:00:46,790 --> 00:00:50,290
is artificial general intelligence or AGI:

12
00:00:52,440 --> 00:00:55,693
a digital consciousness
that surpasses humanity

13
00:00:55,693 --> 00:00:58,891
and might decide to do away with us.

14
00:00:58,891 --> 00:01:02,891
A lot of money, effort,
and anxiety is right now

15
00:01:04,111 --> 00:01:09,028
being poured into making sure
AGI aligns with human values,

16
00:01:11,890 --> 00:01:14,466
but I'd like to put the alignment problem

17
00:01:14,466 --> 00:01:16,862
to one side for the moment.

18
00:01:16,862 --> 00:01:20,115
We will come back to it later.

19
00:01:20,115 --> 00:01:22,532
AGI does not currently exist.

20
00:01:23,639 --> 00:01:26,313
What we do have are a lot
of very different tools

21
00:01:26,313 --> 00:01:28,230
that are all called AI.

22
00:01:29,127 --> 00:01:31,138
We have algorithms that judge people

23
00:01:31,138 --> 00:01:33,374
on their parole hearings,
job applications,

24
00:01:33,374 --> 00:01:34,809
and credit scores.

25
00:01:34,809 --> 00:01:37,168
We have security systems
that can recognise faces

26
00:01:37,168 --> 00:01:39,866
and voices, and we have
text and image generators

27
00:01:39,866 --> 00:01:42,353
like Dall-E and ChatGPT.

28
00:01:42,353 --> 00:01:45,013
We are gonna be talking about all of that

29
00:01:45,013 --> 00:01:48,441
because before we get
anywhere close to Skynet,

30
00:01:48,441 --> 00:01:52,217
there are ethical dilemmas
that arise for regular AI.

31
00:01:52,217 --> 00:01:56,257
We know that AI can
produce unethical outcomes.

32
00:01:56,257 --> 00:01:59,612
For example, in 2018 Amazon
experimented with using it

33
00:01:59,612 --> 00:02:03,896
to review job applications and
rate them from one to five.

34
00:02:03,896 --> 00:02:05,078
They had to scrap it

35
00:02:05,078 --> 00:02:08,283
when it started downgrading
applications from women.

36
00:02:08,283 --> 00:02:11,155
So we know that things can go wrong,

37
00:02:11,155 --> 00:02:13,342
is there anything we can do about it?

38
00:02:13,342 --> 00:02:16,090
There are two broad approaches.

39
00:02:16,090 --> 00:02:18,978
The first is 'build it better.'

40
00:02:18,978 --> 00:02:20,728
Just build better AI.

41
00:02:21,901 --> 00:02:24,769
For example, data scientist
Fabian Beigang says

42
00:02:24,769 --> 00:02:26,267
we need to recognise

43
00:02:26,267 --> 00:02:30,384
algorithmic decision making has two steps.

44
00:02:30,384 --> 00:02:34,891
In step one, the model makes a
prediction, like 'This resume

45
00:02:34,891 --> 00:02:36,889
is a five out of five.'

46
00:02:36,889 --> 00:02:40,546
In step two, it allocates
something on the basis of that,

47
00:02:40,546 --> 00:02:44,201
like 'Therefore you should
give this person a job.'

48
00:02:44,201 --> 00:02:46,990
These are really two
separate tasks and therefore

49
00:02:46,990 --> 00:02:50,297
there are two different
ways things can go wrong.

50
00:02:50,297 --> 00:02:54,604
In step one prediction, the
model might be inaccurate.

51
00:02:54,604 --> 00:02:56,408
Like Amazon's AI,

52
00:02:56,408 --> 00:02:59,166
it looked at all the resumes
submitted to the company

53
00:02:59,166 --> 00:03:02,307
in the previous 10 years
and learned the patterns,

54
00:03:02,307 --> 00:03:03,809
which might sound good,

55
00:03:03,809 --> 00:03:06,828
but tech is a very male dominated field.

56
00:03:06,828 --> 00:03:09,691
So the system concluded from
its limited training data

57
00:03:09,691 --> 00:03:11,787
that men must be better applicants

58
00:03:11,787 --> 00:03:15,889
because they get offered more
jobs, and that's not true.

59
00:03:15,889 --> 00:03:17,706
So if the problem is accuracy,

60
00:03:17,706 --> 00:03:21,979
then more data and more
diverse data could help.

61
00:03:21,979 --> 00:03:25,758
In step two, allocation,
the model might be unfair.

62
00:03:25,758 --> 00:03:28,708
Like Amazon's AI offered jobs mainly to men.

63
00:03:28,708 --> 00:03:31,014
So we just constrain
what the model can do.

64
00:03:31,014 --> 00:03:33,543
For example, tell it that
if 30% of applications

65
00:03:33,543 --> 00:03:34,376
come from women,

66
00:03:34,376 --> 00:03:37,038
30% of jobs need to be
offered to women too.

67
00:03:37,038 --> 00:03:38,405
If the problem is fairness,

68
00:03:38,405 --> 00:03:43,279
just tell the computer to
be fairer and... we've done it!

69
00:03:43,279 --> 00:03:46,639
We've made ethical AI! And all
it took was one bit of code:

70
00:03:46,639 --> 00:03:48,306
'If evil, then don't.'

71
00:03:48,939 --> 00:03:55,896
[audience cheering, triumphant dubstep]

72
00:03:56,133 --> 00:03:58,550
But maybe it's not that easy.

73
00:03:59,748 --> 00:04:01,350
Here's an ethical dilemma.

74
00:04:01,350 --> 00:04:06,108
Suppose we are building an AI
to help a university decide

75
00:04:06,108 --> 00:04:09,289
which students should
be offered scholarships

76
00:04:09,289 --> 00:04:11,353
based on their resumes.

77
00:04:11,353 --> 00:04:15,484
The university wants to save
money by accurately predicting

78
00:04:15,484 --> 00:04:19,125
which students will achieve
a certain grade level

79
00:04:19,125 --> 00:04:21,252
that represents a good use of funds.

80
00:04:21,252 --> 00:04:23,319
And suppose at this particular university

81
00:04:23,319 --> 00:04:27,034
that students of colour drop
out more than white students

82
00:04:27,034 --> 00:04:28,914
and are therefore more likely to represent

83
00:04:28,914 --> 00:04:31,171
a loss of scholarship money.

84
00:04:31,171 --> 00:04:34,799
The AI we build will learn that pattern.

85
00:04:34,799 --> 00:04:38,555
However, suppose that students
of colour drop out more

86
00:04:38,555 --> 00:04:41,537
because there's problems
with faculty racism,

87
00:04:41,537 --> 00:04:42,846
and lack of affordable housing,

88
00:04:42,846 --> 00:04:44,977
and harassment by campus police.

89
00:04:44,977 --> 00:04:47,489
They do drop out at higher rates and race

90
00:04:47,489 --> 00:04:51,894
is a predicting factor,
but it's not their fault.

91
00:04:51,894 --> 00:04:55,561
So should we alter the
model to ignore race?

92
00:04:57,625 --> 00:05:02,292
That seems fair, but it would
make step one, prediction,

93
00:05:03,525 --> 00:05:06,160
less accurate and undermine the purpose

94
00:05:06,160 --> 00:05:09,726
for which the university
wants to use the AI.

95
00:05:09,726 --> 00:05:11,168
As for step two, allocation,

96
00:05:11,168 --> 00:05:13,709
we could tell the model
that if 30% of applications

97
00:05:13,709 --> 00:05:14,908
come from students of colour,

98
00:05:14,908 --> 00:05:17,137
then 30% of offers need to go there too.

99
00:05:17,137 --> 00:05:19,256
'If racism, then don't.'

100
00:05:19,256 --> 00:05:21,874
But that would mean some
scholarships get offered

101
00:05:21,874 --> 00:05:25,282
to students who really weren't
the most likely to succeed,

102
00:05:25,282 --> 00:05:28,974
which again undermines
the point of building it.

103
00:05:28,974 --> 00:05:30,176
And you might say, "Well
that doesn't matter.

104
00:05:30,176 --> 00:05:31,619
It's fairer that way."

105
00:05:31,619 --> 00:05:33,618
And personally I'd agree with you,

106
00:05:33,618 --> 00:05:36,437
but now we're not
talking about technology,

107
00:05:36,437 --> 00:05:38,634
we're talking about political philosophy.

108
00:05:38,634 --> 00:05:41,712
'Cause it turns out there's
a trade off between accuracy

109
00:05:41,712 --> 00:05:45,133
and the kind of society we should live in.

110
00:05:45,133 --> 00:05:47,674
And the trouble with that
is not everyone agrees

111
00:05:47,674 --> 00:05:50,511
what kind of society we should live in!

112
00:05:50,511 --> 00:05:52,859
This dilemma can come up a lot.

113
00:05:52,859 --> 00:05:55,513
Another example might be
if you search Google images

114
00:05:55,513 --> 00:05:58,919
for 'CEO,' should it show you pictures

115
00:05:58,919 --> 00:06:01,759
that accurately reflect most CEOs,

116
00:06:01,759 --> 00:06:03,932
which is to say pictures of men

117
00:06:03,932 --> 00:06:07,923
and therefore risk
reinforcing sexist biases?

118
00:06:07,923 --> 00:06:11,089
Or should it show you images
of the kind of world we want

119
00:06:11,089 --> 00:06:14,258
to live in where CEOs of all genders

120
00:06:14,258 --> 00:06:16,055
are guillotined for their crimes?

121
00:06:16,055 --> 00:06:18,700
So okay, how to make ethical AI?

122
00:06:18,700 --> 00:06:21,920
We've got 'build it better' and
there are some ideas there,

123
00:06:21,920 --> 00:06:24,066
but also some trade-offs.

124
00:06:24,066 --> 00:06:28,196
So maybe we also need 'police it better.'

125
00:06:28,196 --> 00:06:31,582
AI can never be perfect,
so when it goes wrong

126
00:06:31,582 --> 00:06:34,478
we need to make sure
that people have options.

127
00:06:34,478 --> 00:06:37,431
In the novel, "The Trial" by Franz Kafka,

128
00:06:37,431 --> 00:06:40,227
the protagonist Josef K is arrested

129
00:06:40,227 --> 00:06:42,922
and as he moves through the
criminal justice system

130
00:06:42,922 --> 00:06:46,131
he's never told what
crime he's accused of.

131
00:06:46,131 --> 00:06:48,321
The story is horrifying because Josef

132
00:06:48,321 --> 00:06:50,738
is rendered powerless by the bureaucracy.

133
00:06:50,738 --> 00:06:53,397
He can't take action and
he can't even find out

134
00:06:53,397 --> 00:06:57,864
what action he could take
until eventually he's executed

135
00:06:57,864 --> 00:06:59,975
and never told why.

136
00:06:59,975 --> 00:07:02,616
Philosopher Kate Vredenburg says that

137
00:07:02,616 --> 00:07:05,663
when AI goes wrong,
it's like a Kafka novel!

138
00:07:05,663 --> 00:07:08,129
You get denied something
and you don't know why

139
00:07:08,129 --> 00:07:10,553
and if you don't know, you can't improve.

140
00:07:10,553 --> 00:07:13,319
If your resume gets rejected
by the recruitment AI

141
00:07:13,319 --> 00:07:14,555
that doesn't give you any feedback,

142
00:07:14,555 --> 00:07:17,128
you can't write a better resume next time.

143
00:07:17,128 --> 00:07:19,052
Not only does that make you powerless,

144
00:07:19,052 --> 00:07:22,042
she says, it'll probably
undermine people's trust

145
00:07:22,042 --> 00:07:23,791
in institutions that use AI,

146
00:07:23,791 --> 00:07:25,184
which could be a big
problem if we're talking

147
00:07:25,184 --> 00:07:27,107
about say a court.

148
00:07:27,107 --> 00:07:30,774
So she proposes 'the
right to an explanation.'

149
00:07:32,715 --> 00:07:35,487
If an AI makes a decision
that affects your life,

150
00:07:35,487 --> 00:07:39,311
you should be entitled
to know why it did that.

151
00:07:39,311 --> 00:07:40,809
In some places like the EU,

152
00:07:40,809 --> 00:07:42,655
this is already a real legal right!

153
00:07:42,655 --> 00:07:44,905
But here's another dilemma.

154
00:07:45,906 --> 00:07:48,706
If I own the recruitment software

155
00:07:48,706 --> 00:07:53,048
that rejects your resume
and you say, "Explain this",

156
00:07:53,048 --> 00:07:57,048
what kind of explanation
do you want me to give?

157
00:07:58,235 --> 00:08:02,139
Do you want me to tell you how
the model works, generally?

158
00:08:02,139 --> 00:08:04,619
Like do you want me to show you the code?

159
00:08:04,619 --> 00:08:08,771
Or tell you what it did
specifically in your case?

160
00:08:08,771 --> 00:08:11,313
There might be some practical
trade-offs to be made here.

161
00:08:11,313 --> 00:08:14,190
If I'm a big company, then
I might not have the time

162
00:08:14,190 --> 00:08:16,194
to give you individual feedback.

163
00:08:16,194 --> 00:08:18,822
The most helpful kind of
explanation would probably be

164
00:08:18,822 --> 00:08:21,895
what's called a counterfactual one.

165
00:08:21,895 --> 00:08:25,131
That's where we say, "If you
had done this differently,

166
00:08:25,131 --> 00:08:27,143
the AI would've given you what you want."

167
00:08:27,143 --> 00:08:30,588
Like, "If you hadn't spelled
the company name wrong

168
00:08:30,588 --> 00:08:33,588
on your application, it
would've given you the job."

169
00:08:33,588 --> 00:08:35,370
That kind of explanation is useful

170
00:08:35,370 --> 00:08:37,331
because you can do better next time.

171
00:08:37,331 --> 00:08:38,383
Obviously though, if we're

172
00:08:38,383 --> 00:08:41,813
giving counterfactual explanations
of algorithmic decisions

173
00:08:41,813 --> 00:08:44,656
those counterfactuals need to be true!

174
00:08:44,656 --> 00:08:46,985
It needs to be true
that if you had spelled

175
00:08:46,985 --> 00:08:49,185
the company name right, the
AI would've given you the job.

176
00:08:49,185 --> 00:08:52,495
Otherwise that's not
actually a useful explanation

177
00:08:52,495 --> 00:08:55,505
and that can present a
problem because a lot of AIs

178
00:08:55,505 --> 00:08:58,486
are what's called black box models.

179
00:08:58,486 --> 00:09:01,144
They use multiple layers
of non-linear programming

180
00:09:01,144 --> 00:09:03,488
and they're so complicated
that their outputs

181
00:09:03,488 --> 00:09:06,452
are a mystery even to their creators.

182
00:09:06,452 --> 00:09:08,974
We could guess that if
you spelled the name right,

183
00:09:08,974 --> 00:09:13,085
it'd give you the job, but
how can we prove that to you?

184
00:09:13,085 --> 00:09:15,943
One solution to this is quite fun.

185
00:09:15,943 --> 00:09:17,427
Have you ever heard the nursery rhyme

186
00:09:17,427 --> 00:09:18,992
about the old lady who swallowed a fly

187
00:09:18,992 --> 00:09:21,410
so she swallowed a
spider to catch the fly?

188
00:09:21,410 --> 00:09:23,880
It turns out you can
kind of do that with AI!

189
00:09:23,880 --> 00:09:25,810
If we have a black box
model whose behaviour

190
00:09:25,810 --> 00:09:26,891
we want to explain

191
00:09:26,891 --> 00:09:31,196
we can build a second
"surrogate model" to approximate

192
00:09:31,196 --> 00:09:33,374
a simpler version of
what the first one did

193
00:09:33,374 --> 00:09:36,908
and get counterfactual
explanations from that.

194
00:09:36,908 --> 00:09:39,030
It's kindof like when police hire actors

195
00:09:39,030 --> 00:09:40,565
to do crime scene reenactments

196
00:09:40,565 --> 00:09:44,202
and show what probably happened,
like it's not bulletproof,

197
00:09:44,202 --> 00:09:46,069
but it is an option.

198
00:09:46,069 --> 00:09:49,355
However, here's another dilemma!

199
00:09:49,355 --> 00:09:53,731
Suppose we build an AI whose
outputs we know are unfair,

200
00:09:53,731 --> 00:09:56,692
like we deliberately
build the Racism Machine

201
00:09:56,692 --> 00:09:59,882
and we put it in charge of
our company's recruitment.

202
00:09:59,882 --> 00:10:01,052
It's a black box model,

203
00:10:01,052 --> 00:10:02,822
so when a candidate of colour gets rejected

204
00:10:02,822 --> 00:10:04,911
and requests their
right to an explanation,

205
00:10:04,911 --> 00:10:08,994
we turn to our surrogate
model, which tells lies.

206
00:10:11,460 --> 00:10:13,529
We build the Racism Machine

207
00:10:13,529 --> 00:10:17,060
and then we build the Uncle
Tom machine to tell everybody

208
00:10:17,060 --> 00:10:20,933
that the Racism Machine just
has some legitimate concerns

209
00:10:20,933 --> 00:10:22,389
and by calling it the Racism Machine

210
00:10:22,389 --> 00:10:24,187
you're actually silencing
artificial voices,

211
00:10:24,187 --> 00:10:27,182
which is not very tolerant of you.

212
00:10:27,182 --> 00:10:29,341
So you use your right to an explanation,

213
00:10:29,341 --> 00:10:33,110
but the explanation you get is
bullsh*t! And you probably

214
00:10:33,110 --> 00:10:37,289
don't have the time or
expertise to prove it.

215
00:10:37,289 --> 00:10:40,122
This process is called 'fairwashing'

216
00:10:41,037 --> 00:10:43,280
and it was first proposed
by a group of scientists

217
00:10:43,280 --> 00:10:46,058
in a paper published in 2019.

218
00:10:46,058 --> 00:10:50,753
Not content with establishing
the theoretical possibility,

219
00:10:50,753 --> 00:10:51,753
they did it!

220
00:10:52,946 --> 00:10:56,188
They built an AI they knew was biased,

221
00:10:56,188 --> 00:10:58,069
they deliberately built the Racism Machine,

222
00:10:58,069 --> 00:11:00,129
and then they built a
second surrogate model

223
00:11:00,129 --> 00:11:04,351
that they called Laundry
M.L. to launder the outputs

224
00:11:04,351 --> 00:11:05,305
of the first one.

225
00:11:05,305 --> 00:11:09,347
Turns out you can make
the results seem fair

226
00:11:09,347 --> 00:11:11,930
even when you know they're not.

227
00:11:13,510 --> 00:11:15,420
Weirdly, after that paper was published,

228
00:11:15,420 --> 00:11:17,953
Laundry M.L. got a job
writing for the telegraph.

229
00:11:17,953 --> 00:11:22,557
[fancy British music]

230
00:11:22,557 --> 00:11:24,913
Okay, how to make ethical AI?

231
00:11:24,913 --> 00:11:27,537
We've tried 'build it
better' and 'police it better'

232
00:11:27,537 --> 00:11:31,356
and there are options
there, but also limitations.

233
00:11:31,356 --> 00:11:33,932
Technical limitations, yes,

234
00:11:33,932 --> 00:11:37,599
but also there seems
to be a deeper problem...

235
00:11:41,584 --> 00:11:45,890
I'd like to talk about the
penis detection machine!

236
00:11:45,890 --> 00:11:50,344
[poppy girly music like whoa we just pivoted tone HARD here!]

237
00:11:50,344 --> 00:11:53,677
Fairwashing isn't just a technical snag.

238
00:11:54,663 --> 00:11:57,761
The reason it's an
issue is ordinary people

239
00:11:57,761 --> 00:12:01,021
wouldn't have the ability
to challenge institutions

240
00:12:01,021 --> 00:12:04,104
who misuse AI, so not a tech problem,

241
00:12:05,200 --> 00:12:07,865
a problem of unequal power.

242
00:12:07,865 --> 00:12:10,889
And indeed a whole suite of
problems present themselves

243
00:12:10,889 --> 00:12:14,808
when we remember that AI
(and all technology)

244
00:12:14,808 --> 00:12:18,141
is used in an already imperfect society.

245
00:12:19,276 --> 00:12:21,363
A fantastic illustration of this

246
00:12:21,363 --> 00:12:23,543
is the penis detection machine.

247
00:12:23,543 --> 00:12:25,461
If you've flown on a plane
in the last few years

248
00:12:25,461 --> 00:12:26,861
then you've probably seen one:

249
00:12:26,861 --> 00:12:29,751
it's the thing that you step
into at airport security

250
00:12:29,751 --> 00:12:33,130
and you put your hands up
and it goes VVVVVVVVVVFT.

251
00:12:33,130 --> 00:12:35,351
Its proper name is the full body scanner,

252
00:12:35,351 --> 00:12:37,885
but we in the transfem community know it

253
00:12:37,885 --> 00:12:41,653
as the dick detector, the
cock spotter, the wang RADAR,

254
00:12:41,653 --> 00:12:43,810
or the penis detection machine.

255
00:12:43,810 --> 00:12:47,644
If you are cisgender, which
is to say not transgender,

256
00:12:47,644 --> 00:12:49,072
then you might not know this,

257
00:12:49,072 --> 00:12:51,258
but just before you
step into that scanner

258
00:12:51,258 --> 00:12:53,827
someone on the other side looks at you

259
00:12:53,827 --> 00:12:57,994
and then presses one of
two buttons, blue or pink.

260
00:12:59,464 --> 00:13:02,036
The scanner looks for
lumps that might indicate

261
00:13:02,036 --> 00:13:04,279
something hidden under your clothing,

262
00:13:04,279 --> 00:13:07,545
which if you are a
transgender woman with a penis

263
00:13:07,545 --> 00:13:10,862
and breasts presents a dilemma.

264
00:13:10,862 --> 00:13:15,589
If they press the blue button for boys, 
then the machine thinks that your tiddies are the bomb.

265
00:13:15,589 --> 00:13:21,427
If they press the pink button for girls, then it's 
"unexpected item in shagging area."

266
00:13:21,427 --> 00:13:23,806
You might have to answer
some humiliating questions,

267
00:13:23,806 --> 00:13:26,811
possibly out yourself as
trans and maybe get groped

268
00:13:26,811 --> 00:13:28,849
by airport security.

269
00:13:28,849 --> 00:13:30,085
And we do laugh about it,

270
00:13:30,085 --> 00:13:32,868
but going through that
in front of my parents

271
00:13:32,868 --> 00:13:36,044
was one of the most humiliating
experiences of my life.

272
00:13:36,044 --> 00:13:38,706
The penis detection
machine sees the presence

273
00:13:38,706 --> 00:13:41,456
of a penis on a woman as strange.

274
00:13:42,756 --> 00:13:46,756
It does that because it
was built by cis people.

275
00:13:47,643 --> 00:13:49,325
If I'd built it, I'd make it so

276
00:13:49,325 --> 00:13:51,112
that it called every woman strange

277
00:13:51,112 --> 00:13:52,909
if she didn't have a penis,

278
00:13:52,909 --> 00:13:56,443
which would be extremely
funny and also makes the point

279
00:13:56,443 --> 00:13:59,693
that technology encodes ways of seeing.

280
00:14:01,959 --> 00:14:05,616
The real world is not
simple and black and white

281
00:14:05,616 --> 00:14:08,896
and doesn't always
correspond to binary gender,

282
00:14:08,896 --> 00:14:11,712
but the machine expects that.

283
00:14:11,712 --> 00:14:15,055
Moreover, by being part of
a system of surveillance

284
00:14:15,055 --> 00:14:17,709
(and occasionally humiliation and assault)

285
00:14:17,709 --> 00:14:20,292
it helps enforce binary gender.

286
00:14:21,261 --> 00:14:23,063
When all you have is a hammer,

287
00:14:23,063 --> 00:14:25,295
all your problems look like nails.

288
00:14:25,295 --> 00:14:27,453
We often assume that machines

289
00:14:27,453 --> 00:14:31,380
like that detect information
that's already there,

290
00:14:31,380 --> 00:14:33,852
but they also do the opposite:

291
00:14:33,852 --> 00:14:36,836
project information onto the body

292
00:14:36,836 --> 00:14:38,936
in a process that scholar Simone Browne

293
00:14:38,936 --> 00:14:41,519
calls digital epidermalization.

294
00:14:43,084 --> 00:14:46,070
The machine tells you
what certain features

295
00:14:46,070 --> 00:14:49,570
of your body signify, like "That's strange!"

296
00:14:50,901 --> 00:14:55,312
You, the person whose
body it is might be like,

297
00:14:55,312 --> 00:14:59,667
"This isn't strange? For me
this is normal and I like it."

298
00:14:59,667 --> 00:15:03,399
But people assume that machines
are neutral so they listen

299
00:15:03,399 --> 00:15:05,825
to the machine more
than they listen to you

300
00:15:05,825 --> 00:15:08,548
and that can land you in trouble.

301
00:15:08,548 --> 00:15:10,617
The term 'epidermalization' comes

302
00:15:10,617 --> 00:15:12,591
from philosopher Frantz Fanon

303
00:15:12,591 --> 00:15:15,042
who spent some time in France
in the forties and fifties

304
00:15:15,042 --> 00:15:17,579
and found that people called him strange,

305
00:15:17,579 --> 00:15:21,835
unwelcome, and dangerous
because he was black.

306
00:15:21,835 --> 00:15:25,018
They weren't just detecting
the colour of his skin,

307
00:15:25,018 --> 00:15:27,881
they were projecting meaning onto it,

308
00:15:27,881 --> 00:15:32,170
and this could happen
anywhere, on the bus, at work

309
00:15:32,170 --> 00:15:35,251
he was never secure 'cause
at any moment he could have

310
00:15:35,251 --> 00:15:39,668
the meaning of his being
assigned to him by a white person.

311
00:15:39,668 --> 00:15:41,256
And he said that's part of what it means

312
00:15:41,256 --> 00:15:42,886
to be in a racial minority.

313
00:15:42,886 --> 00:15:45,147
You don't get to stand
up and say, "This is me,

314
00:15:45,147 --> 00:15:46,369
this is my deal."

315
00:15:46,369 --> 00:15:49,821
White people tell you what your deal is.

316
00:15:49,821 --> 00:15:51,862
[Voice of Sophie from Mars]
"Digital epidermalization is the exercise

317
00:15:51,862 --> 00:15:54,052
of power cast by the disembodied gaze

318
00:15:54,052 --> 00:15:55,647
of certain surveillance technologies.

319
00:15:55,647 --> 00:15:56,852
For example, identity card

320
00:15:56,852 --> 00:15:59,589
and e-passport verification
machines that can be employed

321
00:15:59,589 --> 00:16:02,035
to do the work of alienating
the subject by producing

322
00:16:02,035 --> 00:16:05,407
a truth about the body and one's identity

323
00:16:05,407 --> 00:16:08,643
or identities despite
the subject's claims."

324
00:16:08,643 --> 00:16:10,645
- What does this have to do with AI?

325
00:16:10,645 --> 00:16:15,241
Well consider something
like facial recognition AI,

326
00:16:15,241 --> 00:16:19,680
and in particular automatic
gender recognition.

327
00:16:19,680 --> 00:16:21,844
In theory, that's a useful piece of tech:

328
00:16:21,844 --> 00:16:23,687
if we're looking for someone in a crowd

329
00:16:23,687 --> 00:16:24,872
using facial recognition

330
00:16:24,872 --> 00:16:27,049
and we know that we're looking for a man

331
00:16:27,049 --> 00:16:29,768
then by detecting gender we
could immediately rule out

332
00:16:29,768 --> 00:16:31,814
all the women and get a result faster.

333
00:16:31,814 --> 00:16:35,523
However, you can't always
tell someone's gender just

334
00:16:35,523 --> 00:16:37,895
by looking at their face.

335
00:16:37,895 --> 00:16:41,005
Automatic gender recognition
might work well enough

336
00:16:41,005 --> 00:16:42,702
most of the time,

337
00:16:42,702 --> 00:16:45,669
which is to say the meanings it assigns

338
00:16:45,669 --> 00:16:48,626
might be useful most of the time,

339
00:16:48,626 --> 00:16:50,386
but like the penis detection machine

340
00:16:50,386 --> 00:16:53,794
it encodes a particular way of seeing

341
00:16:53,794 --> 00:16:55,417
and it's always gonna make mistakes,

342
00:16:55,417 --> 00:16:57,179
particularly with trans people.

343
00:16:57,179 --> 00:16:59,742
Like the penis detection
machine, that could result

344
00:16:59,742 --> 00:17:02,515
in us being humiliated or even hurt.

345
00:17:02,515 --> 00:17:04,834
And you might say, "Well,
that's unfortunate,

346
00:17:04,834 --> 00:17:08,067
but trans people are a
very small minority."

347
00:17:08,067 --> 00:17:11,812
Two things. Firstly, a
mistake that keeps happening

348
00:17:11,812 --> 00:17:15,081
to a specific group of
people isn't just a mistake.

349
00:17:15,081 --> 00:17:17,269
That's systemic discrimination.

350
00:17:17,269 --> 00:17:21,186
Secondly, who told you
we're a small minority?

351
00:17:23,849 --> 00:17:27,352
Being left-handed used to be thought rare,

352
00:17:27,352 --> 00:17:29,661
but then we stopped punishing
children for being left-handed

353
00:17:29,661 --> 00:17:30,978
and we found out there's
actually a lot more

354
00:17:30,978 --> 00:17:33,444
of them than anyone thought!

355
00:17:33,444 --> 00:17:36,878
Aren't we right-handers
lucky that left-handed people

356
00:17:36,878 --> 00:17:38,561
never sought what would surely

357
00:17:38,561 --> 00:17:41,300
have been their justified revenge?!

358
00:17:41,300 --> 00:17:42,472
[adorkable snort laugh]

359
00:17:42,472 --> 00:17:47,103
So how to make ethical AI?
Not just facial recognition,

360
00:17:47,103 --> 00:17:49,715
but any AI that categorizes people,

361
00:17:49,715 --> 00:17:52,263
recruitment software,
university applications, anything?

362
00:17:52,263 --> 00:17:56,024
Wwe could try 'build it better'

363
00:17:56,024 --> 00:17:58,799
- more data, more diverse data,

364
00:17:58,799 --> 00:18:01,458
more buttons on the
penis detection machine -

365
00:18:01,458 --> 00:18:04,621
and we could try 'police it
better' - give people options

366
00:18:04,621 --> 00:18:09,454
if they are miscategorised -
but that still might not be enough.

367
00:18:10,711 --> 00:18:14,982
Tech researcher Os Keyes
says gender recognition AI,

368
00:18:14,982 --> 00:18:17,773
no matter how inclusive we try to make it,

369
00:18:17,773 --> 00:18:20,087
is always premised on the idea

370
00:18:20,087 --> 00:18:22,733
that someone else can look at you

371
00:18:22,733 --> 00:18:25,316
and tell you what your deal is.

372
00:18:26,647 --> 00:18:28,926
And that's always gonna
be fundamentally incorrect

373
00:18:28,926 --> 00:18:31,655
because one of the fun
and beautiful things

374
00:18:31,655 --> 00:18:36,155
about human gender is deciding
your deal for yourself.

375
00:18:37,656 --> 00:18:39,368
[Voice of Sophie from Mars]
"Automated gender recognition purports

376
00:18:39,368 --> 00:18:42,155
to recognize gender - but
my analysis of the way in

377
00:18:42,155 --> 00:18:43,026
which it operates shows

378
00:18:43,026 --> 00:18:44,926
that this is only true
if one denies the role

379
00:18:44,926 --> 00:18:47,135
that self-knowledge plays in gender

380
00:18:47,135 --> 00:18:50,674
and consequently denies the
existence of trans people.

381
00:18:50,674 --> 00:18:52,060
Using a more inclusive view of gender,

382
00:18:52,060 --> 00:18:54,726
one that recognizes the
primacy of the self,

383
00:18:54,726 --> 00:18:58,084
we can see that the systems
are not recognizing gender,

384
00:18:58,084 --> 00:19:01,374
which would be impossible
to reliably do by inference,

385
00:19:01,374 --> 00:19:04,446
but instead imposing
their own views of gender

386
00:19:04,446 --> 00:19:07,232
on unwitting users and research subjects.

387
00:19:07,232 --> 00:19:09,865
30 years of research has
not produced a single system

388
00:19:09,865 --> 00:19:12,664
that can genuinely recognise gender,

389
00:19:12,664 --> 00:19:16,441
it has only produced systems
that try to assign it."

390
00:19:16,441 --> 00:19:21,212
- To summarize, AI is used in
an already imperfect society

391
00:19:21,212 --> 00:19:24,201
and might exacerbate existing inequality.

392
00:19:24,201 --> 00:19:27,322
The solutions may not
always be technological.

393
00:19:27,322 --> 00:19:31,084
So when building it, we need
to pay careful attention

394
00:19:31,084 --> 00:19:34,329
and broaden the scope of our inquiries

395
00:19:34,329 --> 00:19:37,912
and critically interrogate
our assumptions.

396
00:19:39,861 --> 00:19:41,609
I've read a lot of papers about this issue

397
00:19:41,609 --> 00:19:43,991
and in the final paragraph
they always pull back

398
00:19:43,991 --> 00:19:46,324
and use a vague sentence like that

399
00:19:46,324 --> 00:19:47,635
because academic philosophers

400
00:19:47,635 --> 00:19:49,926
don't like telling people what to do.

401
00:19:49,926 --> 00:19:52,593
Their go-to solution is usually

402
00:19:52,805 --> 00:19:56,586
"Uhhhhh let's try doing some more philosophy?"

403
00:19:56,586 --> 00:20:00,145
Practical steps! Tech
companies can hire sociologists

404
00:20:00,145 --> 00:20:03,711
and philosophers to work on
every step of AI development

405
00:20:03,711 --> 00:20:05,412
and pay particular attention to the ways

406
00:20:05,412 --> 00:20:09,360
in which the AI might exacerbate
existing inequalities.

407
00:20:09,360 --> 00:20:12,675
When in doubt, hire someone
with a humanities degree!

408
00:20:12,675 --> 00:20:15,247
Woo, we have covered a lot of ground!

409
00:20:15,247 --> 00:20:18,647
We've got technical problems
and societal problems,

410
00:20:18,647 --> 00:20:21,085
but I'm sure that if we host
some diversity, equality,

411
00:20:21,085 --> 00:20:22,902
and inclusion seminars,
we can solve most of that.

412
00:20:22,902 --> 00:20:25,819
So we are ready to make ethical AI!

413
00:20:29,507 --> 00:20:34,007
Right? I mean, we must have
covered everything by now.

414
00:20:34,889 --> 00:20:38,597
It's not like there's an even bigger

415
00:20:38,597 --> 00:20:42,980
and even more fundamental probl -

416
00:20:42,980 --> 00:20:48,543
A while ago, someone used AI to make non-consensual pornography of me.

417
00:20:48,543 --> 00:20:52,054
[dramatic sting like OOOOOHHHH we are GOING THERE]

418
00:20:52,054 --> 00:20:53,810
- This person took screenshots

419
00:20:53,810 --> 00:20:56,095
from a previous episode
of Philosophy Tube.

420
00:20:56,095 --> 00:20:58,101
Do you remember the one we
did about effective altruism

421
00:20:58,101 --> 00:21:00,664
when I was dressed as the latex nun?

422
00:21:00,664 --> 00:21:04,146
And they fed those images
to an AI image generator,

423
00:21:04,146 --> 00:21:06,980
which produced a series of pictures of me,

424
00:21:06,980 --> 00:21:10,242
all of which were quite
bad, in similar costumes,

425
00:21:10,242 --> 00:21:13,454
and also some where I was partially nude.

426
00:21:13,454 --> 00:21:17,037
And then they posted
those images publicly.

427
00:21:19,138 --> 00:21:22,907
Okay, I'd like to tell
you a little bit about how

428
00:21:22,907 --> 00:21:25,433
that look came together.

429
00:21:25,433 --> 00:21:28,741
In the show notes I
called it the Techno Nun.

430
00:21:28,741 --> 00:21:32,022
I came up with the idea after
researching effective altruism

431
00:21:32,022 --> 00:21:36,933
for a long time, I wrote about
53,000 words worth of notes

432
00:21:36,933 --> 00:21:40,059
for that episode. And I
wanted a look that combined

433
00:21:40,059 --> 00:21:44,175
the idea of altruism, helping
other people, with science,

434
00:21:44,175 --> 00:21:46,732
hence 'techno' and 'nun.'

435
00:21:46,732 --> 00:21:48,664
And when I had a design I was happy with

436
00:21:48,664 --> 00:21:50,064
I worked with Brian,

437
00:21:50,064 --> 00:21:53,449
who is the Philosophy
Tube stylist, to refine it,

438
00:21:53,449 --> 00:21:56,115
and then we went to a
lovely lady named Nange

439
00:21:56,115 --> 00:21:58,943
who runs a latex company
called Dead Lotus Couture,

440
00:21:58,943 --> 00:22:01,987
and we worked a little bit
more with her and then her

441
00:22:01,987 --> 00:22:05,190
and her assistants made that outfit.

442
00:22:05,190 --> 00:22:09,614
Fun fact, Lady Gaga gets her latex from the same place as me!

443
00:22:09,614 --> 00:22:12,204
Stef, you're such a Philosophy Tube rip-off!

444
00:22:12,204 --> 00:22:16,658
[laughing]
I'm kidding, I'm sorry! I loved House of Gucci!

445
00:22:16,658 --> 00:22:18,327
[laughing]

446
00:22:18,327 --> 00:22:24,019
We filmed that episode in a studio in North London, which I hired.
Brian was my dresser on the day,

447
00:22:24,019 --> 00:22:26,616
and Nicki, the Philosophy
Tube makeup artist,

448
00:22:26,616 --> 00:22:28,426
did all of the hair and the makeup.

449
00:22:28,426 --> 00:22:30,823
Nicki works so hard on this show,

450
00:22:30,823 --> 00:22:33,416
like she often comes up
with these little ideas

451
00:22:33,416 --> 00:22:36,032
at the last minute, which
just like elevate things.

452
00:22:36,032 --> 00:22:36,970
Like on the previous episode we did

453
00:22:36,970 --> 00:22:39,819
when I was the demon
lawyer, using victory rolls

454
00:22:39,819 --> 00:22:43,335
in my hair to look like
horns, was Nicki's idea!

455
00:22:43,335 --> 00:22:46,588
And then Mr. X, who is the
Philosophy Tube camera operator

456
00:22:46,588 --> 00:22:48,754
who prefers to remain anonymous,

457
00:22:48,754 --> 00:22:52,011
set up the camera and the
lights and operated the camera

458
00:22:52,011 --> 00:22:54,173
and colour graded all of the footage

459
00:22:54,173 --> 00:22:55,943
and then I edited the footage.

460
00:22:55,943 --> 00:22:58,289
I do all the editing on
Philosophy Tube myself.

461
00:22:58,289 --> 00:23:01,229
It usually takes me about
a week of 12 hour days.

462
00:23:01,229 --> 00:23:03,543
And then I uploaded the
final video to YouTube

463
00:23:03,543 --> 00:23:05,180
and to Nebula.

464
00:23:05,180 --> 00:23:07,702
I did the thumbnails, the titles,

465
00:23:07,702 --> 00:23:09,896
the subtitles, the metadata,

466
00:23:09,896 --> 00:23:11,830
and all the stuff
that takes a video file

467
00:23:11,830 --> 00:23:15,065
and turns it into an
episode of Philosophy Tube.

468
00:23:15,065 --> 00:23:17,202
I could go on, I haven't
even mentioned the people

469
00:23:17,202 --> 00:23:18,583
who built the cameras

470
00:23:18,583 --> 00:23:21,653
and drove us to set and made the coffee.

471
00:23:21,653 --> 00:23:26,236
If you added up all the time
and money and work it took

472
00:23:27,102 --> 00:23:31,539
to produce those screenshots,
it would be staggering.

473
00:23:31,539 --> 00:23:34,561
There's a reason that Philosophy
Tube has a Patreon page

474
00:23:34,561 --> 00:23:37,466
where people who like the
show chip in what they can,

475
00:23:37,466 --> 00:23:42,466
and it's because the people
who make it work f*****g hard!

476
00:23:43,527 --> 00:23:46,979
And this person took those screenshots

477
00:23:46,979 --> 00:23:50,229
and flattened all of that work in order

478
00:23:51,532 --> 00:23:55,949
to produce some shoddy images
of me with my tits out.

479
00:23:58,392 --> 00:24:01,646
They evidently didn't consider
the ethics of doing that.

480
00:24:01,646 --> 00:24:04,472
They appropriated all that
labor from all those people

481
00:24:04,472 --> 00:24:06,992
without permission and
presented it as their own,

482
00:24:06,992 --> 00:24:11,492
and I don't think they
realised how insulting that is.

483
00:24:12,840 --> 00:24:17,478
They assumed that what we had
labored to make was just data,

484
00:24:17,478 --> 00:24:20,182
just something they could
take for themselves.

485
00:24:20,182 --> 00:24:23,432
And I personally don't see it that way.

486
00:24:24,898 --> 00:24:28,283
The fact that the end result
was non-consensual pornography

487
00:24:28,283 --> 00:24:31,533
was kind of just the icing on the turd.

488
00:24:32,686 --> 00:24:34,948
Interestingly, a few months prior to that,

489
00:24:34,948 --> 00:24:39,352
somebody made hand-drawn
non-consensual porn of me

490
00:24:39,352 --> 00:24:43,490
and that felt more violating,
weirdly, than the AI stuff!

491
00:24:43,490 --> 00:24:46,693
I think because the intent
was quite different.

492
00:24:46,693 --> 00:24:48,910
The person who made the AI porn of me,

493
00:24:48,910 --> 00:24:50,824
I think genuinely didn't know

494
00:24:50,824 --> 00:24:53,239
that the AI was gonna do that
and they didn't look closely

495
00:24:53,239 --> 00:24:55,454
and they posted it in
a spirit of excitement.

496
00:24:55,454 --> 00:24:57,316
And as soon as I said, "Oh my God,

497
00:24:57,316 --> 00:24:58,848
that's disgusting, please don't",

498
00:24:58,848 --> 00:25:00,932
they apologised and deleted it.

499
00:25:00,932 --> 00:25:05,710
Whereas the person who made the
hand-drawn version obviously

500
00:25:05,710 --> 00:25:08,455
had spent a long time thinking about me

501
00:25:08,455 --> 00:25:11,455
in a very obsessive and lustful way

502
00:25:12,377 --> 00:25:13,693
and by posting it publicly,

503
00:25:13,693 --> 00:25:14,646
they were deliberately trying

504
00:25:14,646 --> 00:25:16,410
to make me feel sexually violated.

505
00:25:18,910 --> 00:25:21,646
Which... yeah, I mean that sucks

506
00:25:21,646 --> 00:25:23,563
if I'm honest with you!

507
00:25:24,922 --> 00:25:27,478
It's hard to deal with that
first thing in the morning

508
00:25:27,478 --> 00:25:29,561
and then just go to work.

509
00:25:30,927 --> 00:25:31,823
I was insulted.

510
00:25:31,823 --> 00:25:34,142
My real naked body looks
way better than that!

511
00:25:34,142 --> 00:25:35,961
I tell you this story to introduce

512
00:25:35,961 --> 00:25:38,294
the data flattening problem.

513
00:25:39,298 --> 00:25:42,975
That process is how AI's like
ChatGPT and Dall-E Mini work:

514
00:25:42,975 --> 00:25:45,149
they scrape the internet for data.

515
00:25:45,149 --> 00:25:48,304
No consent is taken, no
care or thought is given,

516
00:25:48,304 --> 00:25:50,294
the people who make these models assume

517
00:25:50,294 --> 00:25:53,452
that they can do that,
because it's just data,

518
00:25:53,452 --> 00:25:58,168
just abstract and immaterial
and that's not true.

519
00:25:58,168 --> 00:26:01,585
Training data is made by and from people.

520
00:26:03,525 --> 00:26:05,343
You might be aware that
in the United States,

521
00:26:05,343 --> 00:26:07,411
the Writer's Guild and
the Screen Actor's Guild

522
00:26:07,411 --> 00:26:11,222
have been on strike? (
At time of recording the WGA might

523
00:26:11,222 --> 00:26:12,504
have reached a provisional agreement

524
00:26:12,504 --> 00:26:14,594
with the studios in the
last like three hours

525
00:26:14,594 --> 00:26:16,419
but we dunno the content of it yet.)

526
00:26:16,419 --> 00:26:20,039
Part of what they're striking about is AI

527
00:26:20,039 --> 00:26:22,289
because text and image
generators obviously

528
00:26:22,289 --> 00:26:24,498
pose a threat to artists' livelihoods.

529
00:26:24,498 --> 00:26:28,156
Studios love the idea of
AI because you can use it

530
00:26:28,156 --> 00:26:31,340
to create a lot of content cheap and fast.

531
00:26:31,340 --> 00:26:32,655
And yeah, most of it's gonna be crap

532
00:26:32,655 --> 00:26:35,161
but you only have to generate
House of the Dragon once

533
00:26:35,161 --> 00:26:37,477
and you've got the
biggest show in the world.

534
00:26:37,477 --> 00:26:41,293
Actors and writers don't
want our work to train AI

535
00:26:41,293 --> 00:26:44,011
and we don't wanna work
on stuff generated by AI,

536
00:26:44,011 --> 00:26:46,784
and part of the reason is that a bot

537
00:26:46,784 --> 00:26:51,012
would scrape a thousand
scripts or performances

538
00:26:51,012 --> 00:26:53,679
crafted by a person and spit out

539
00:26:55,924 --> 00:26:57,849
a soulless ripoff.

540
00:26:57,849 --> 00:27:00,865
It's an insulting misunderstanding

541
00:27:00,865 --> 00:27:03,448
of what we do when we make art.

542
00:27:04,428 --> 00:27:06,928
So how can we make ethical AI?

543
00:27:07,975 --> 00:27:12,435
Is there a way to use data
that prioritises the consent

544
00:27:12,435 --> 00:27:16,007
and the labour of the people
who work to produce it?

545
00:27:16,007 --> 00:27:19,638
Right now, the main suggestion
is to regulate tech companies

546
00:27:19,638 --> 00:27:21,517
through things like privacy laws,

547
00:27:21,517 --> 00:27:24,506
antitrust laws, copyright laws.

548
00:27:24,506 --> 00:27:27,155
Beyond that, there's a slightly
stronger set of proposals

549
00:27:27,155 --> 00:27:29,697
called Data-Owning Democracy.

550
00:27:29,697 --> 00:27:32,091
The idea is that all the data you produce,

551
00:27:32,091 --> 00:27:35,697
like your location data
and your purchase histories

552
00:27:35,697 --> 00:27:38,792
and your browser data, you'd
get a copy of all of that

553
00:27:38,792 --> 00:27:40,165
and then you could decide
what to do with it,

554
00:27:40,165 --> 00:27:42,888
like sell it or donate it maybe.

555
00:27:42,888 --> 00:27:45,257
Beyond that, there's an even
stronger set of proposals

556
00:27:45,257 --> 00:27:48,113
called Digital Socialism
where tech companies

557
00:27:48,113 --> 00:27:50,237
would be democratically
controlled by their workers

558
00:27:50,237 --> 00:27:51,807
and the infrastructure they use

559
00:27:51,807 --> 00:27:55,140
would be owned by society. So, socialism!

560
00:27:56,127 --> 00:27:58,037
Those last two sets of proposals,

561
00:27:58,037 --> 00:28:01,388
Data-Owning Democracy and
Digital Socialism, recommend

562
00:28:01,388 --> 00:28:03,777
redistributing property a little bit

563
00:28:03,777 --> 00:28:07,320
to increase the economic
power of citizens.

564
00:28:07,320 --> 00:28:10,820
And they do that because AI kind of messes

565
00:28:11,718 --> 00:28:15,438
with our concept of private property.

566
00:28:15,438 --> 00:28:17,995
The English philosopher John Locke said

567
00:28:17,995 --> 00:28:22,445
you own your body and
yourself, your person,

568
00:28:22,445 --> 00:28:26,577
and that includes your ability to work.

569
00:28:26,577 --> 00:28:29,809
And if you take the raw
materials that nature provides

570
00:28:29,809 --> 00:28:33,081
and apply your ability to
work, you can transform them.

571
00:28:33,081 --> 00:28:36,905
Like you take a tree and
you carve it into a table.

572
00:28:36,905 --> 00:28:40,293
The things that you make are
like an extension of you.

573
00:28:40,293 --> 00:28:43,586
You put yourself, your
labour power, into them,

574
00:28:43,586 --> 00:28:47,194
and so just as you own yourself,
you now own those things.

575
00:28:47,194 --> 00:28:50,129
They are your property. And if you want to

576
00:28:50,129 --> 00:28:51,661
you can take the things that you own

577
00:28:51,661 --> 00:28:53,751
and you can swap them for money

578
00:28:53,751 --> 00:28:55,498
and you can take the
money and you can swap it

579
00:28:55,498 --> 00:28:59,079
for... body stockings and bondage gear,

580
00:28:59,079 --> 00:29:01,132
and then you own those things too!

581
00:29:01,132 --> 00:29:03,469
I think a lot of people's
idea of private property

582
00:29:03,469 --> 00:29:06,935
is something like that, kindof?

583
00:29:06,935 --> 00:29:08,690
Like if we stopped to
really think about it?

584
00:29:08,690 --> 00:29:10,416
Which let's face it, we often don't

585
00:29:10,416 --> 00:29:12,713
because it's such a
fundamental part of our society

586
00:29:12,713 --> 00:29:15,012
that it often goes unquestioned.

587
00:29:15,012 --> 00:29:16,730
The economist Friedrich
Hayek actually said that

588
00:29:16,730 --> 00:29:19,761
private property is the
cornerstone of civilisation:

589
00:29:19,761 --> 00:29:21,657
once you start breaking those rules, wahey!

590
00:29:21,657 --> 00:29:23,433
The whole thing's gonna come down!

591
00:29:23,433 --> 00:29:27,989
Trouble is, AI does kindof break those rules.

592
00:29:27,989 --> 00:29:32,006
If a black box AI scrapes
10,000 movie scripts

593
00:29:32,006 --> 00:29:35,268
and spits out a new one,
how can we ever tell

594
00:29:35,268 --> 00:29:37,588
whose labour went into the final product

595
00:29:37,588 --> 00:29:38,679
and in what proportions,

596
00:29:38,679 --> 00:29:42,346
and therefore how much
they should be paid?

597
00:29:42,346 --> 00:29:43,757
Probably we can't.

598
00:29:43,757 --> 00:29:45,507
And if we try using a surrogate model,

599
00:29:45,507 --> 00:29:47,629
we're going to encounter
the fairwashing problem

600
00:29:47,629 --> 00:29:50,409
from earlier, which means
the people who produced

601
00:29:50,409 --> 00:29:52,701
that training data are almost certainly

602
00:29:52,701 --> 00:29:54,372
gonna get ripped off.

603
00:29:54,372 --> 00:29:57,535
In this scenario the AI basically steals

604
00:29:57,535 --> 00:29:59,160
the property from writers

605
00:29:59,160 --> 00:30:02,252
and slaps a new label on it,
so they can't do anything,

606
00:30:02,252 --> 00:30:06,391
which is another reason why
screenwriters don't like it!

607
00:30:06,391 --> 00:30:09,714
Annoyingly, it's so easy that just

608
00:30:09,714 --> 00:30:12,995
because it breaks the rules
doesn't mean corporations

609
00:30:12,995 --> 00:30:14,213
won't use it.

610
00:30:14,213 --> 00:30:18,296
And so some corporations
think we need new rules.

611
00:30:20,518 --> 00:30:25,351
Introducing Sam Altman, CEO
of Open AI, makers of ChatGPT.

612
00:30:27,080 --> 00:30:30,531
Sam thinks AI is about to
revolutionise the economy.

613
00:30:30,531 --> 00:30:33,477
It's gonna generate a lot of
content quickly and cheaply,

614
00:30:33,477 --> 00:30:36,066
which means it's gonna
generate a lot of wealth.

615
00:30:36,066 --> 00:30:39,023
But because it's also gonna put
a lot of people like writers

616
00:30:39,023 --> 00:30:40,708
out of a job, most of that wealth

617
00:30:40,708 --> 00:30:43,090
is gonna go to private corporations.

618
00:30:43,090 --> 00:30:45,327
Governments will struggle
to raise income taxes

619
00:30:45,327 --> 00:30:48,754
since most people won't
have income anymore,

620
00:30:48,754 --> 00:30:51,212
so if we're not careful, this could be

621
00:30:51,212 --> 00:30:53,045
the AI jobs-pocalypse.

622
00:30:53,909 --> 00:30:56,624
Altman compares himself
to Robert Oppenheimer

623
00:30:56,624 --> 00:30:58,728
and he takes himself very, very seriously.

624
00:30:58,728 --> 00:31:01,757
So something fun that I
like to do is every time

625
00:31:01,757 --> 00:31:04,294
he mentions computer models,

626
00:31:04,294 --> 00:31:07,051
pretend he's talking about fashion models!

627
00:31:07,051 --> 00:31:08,569
[Voice of Dan Olson]
"Regulatory intervention by governments

628
00:31:08,569 --> 00:31:10,125
will be critical to mitigate the risks

629
00:31:10,125 --> 00:31:12,988
of increasingly powerful models.

630
00:31:12,988 --> 00:31:14,506
I think we're at the end of the era

631
00:31:14,506 --> 00:31:18,582
where it's gonna be
these like giant models.

632
00:31:18,582 --> 00:31:20,246
The model will confidently state things

633
00:31:20,246 --> 00:31:23,852
as if they were facts that
are entirely made up."

634
00:31:23,852 --> 00:31:26,486
- Altman thinks he has a solution

635
00:31:26,486 --> 00:31:28,728
to this jobs-pocalypse problem.

636
00:31:28,728 --> 00:31:30,860
He's a Georgist, a philosophy

637
00:31:30,860 --> 00:31:34,594
named after 19th century
economist Henry George.

638
00:31:34,594 --> 00:31:38,214
Georgists believe that all
taxes should be eliminated

639
00:31:38,214 --> 00:31:41,066
and replaced with a tax on land

640
00:31:41,066 --> 00:31:44,471
because land only gets
its value from the work

641
00:31:44,471 --> 00:31:46,790
that other people do around it.

642
00:31:46,790 --> 00:31:51,192
For example, property in central
London is really expensive

643
00:31:51,192 --> 00:31:53,409
because you're close to all
the theatres and restaurants

644
00:31:53,409 --> 00:31:56,404
and tourist attractions
that people want to go to,

645
00:31:56,404 --> 00:32:00,947
but landlords didn't make
those things, society did.

646
00:32:00,947 --> 00:32:04,142
So the idea is they should
share the value of their land

647
00:32:04,142 --> 00:32:06,658
with society through a tax.

648
00:32:06,658 --> 00:32:10,197
Georgists believe that
capitalism would work great

649
00:32:10,197 --> 00:32:12,914
if landlords just got out of the way

650
00:32:12,914 --> 00:32:15,728
and let the free market be free.

651
00:32:15,728 --> 00:32:18,925
It used to be a fairly popular idea,

652
00:32:18,925 --> 00:32:21,453
partly because it sounds radical,

653
00:32:21,453 --> 00:32:23,478
but you don't actually
have to redistribute

654
00:32:23,478 --> 00:32:26,852
anybody's property - just tax
it a little bit differently.

655
00:32:26,852 --> 00:32:30,036
And partly I think it taps
into that relatable feeling you

656
00:32:30,036 --> 00:32:31,234
get when you pay your rent.

657
00:32:31,234 --> 00:32:35,062
Like, "Oh my God, I am
getting ripped off!"

658
00:32:35,062 --> 00:32:38,825
Altman says the USA should
pass a constitutional amendment

659
00:32:38,825 --> 00:32:43,825
to eliminate all taxes except
land tax and corporation tax.

660
00:32:43,958 --> 00:32:46,798
This tax money should
then be used to buy share

661
00:32:46,798 --> 00:32:49,251
in companies which get given to citizens

662
00:32:49,251 --> 00:32:51,775
as universal basic income.

663
00:32:51,775 --> 00:32:53,001
So they don't need jobs

664
00:32:53,001 --> 00:32:55,867
and they have a stake in
the economy doing well.

665
00:32:55,867 --> 00:32:57,596
So it's like capitalism,

666
00:32:57,596 --> 00:33:00,322
but we've made it work for everybody!

667
00:33:00,322 --> 00:33:03,405
Hang on a minute, did you catch that?

668
00:33:06,018 --> 00:33:09,627
The AI billionaire is
saying, "Don't regulate me,

669
00:33:09,627 --> 00:33:12,359
don't redistribute property
to empower citizens

670
00:33:12,359 --> 00:33:14,109
and don't tax me now,

671
00:33:15,647 --> 00:33:20,647
tax me later when my company
has made trillions of dollars,

672
00:33:20,705 --> 00:33:22,778
which it definitely will

673
00:33:22,778 --> 00:33:25,135
because AI is definitely not a bubble,

674
00:33:25,135 --> 00:33:27,405
even though it costs my
company a huge amount

675
00:33:27,405 --> 00:33:30,464
and our main product seems to be hype.

676
00:33:30,464 --> 00:33:34,127
Tax me later when I've got all that money

677
00:33:34,127 --> 00:33:37,555
and do it by passing
a piece of legislation

678
00:33:37,555 --> 00:33:40,283
that's very, very difficult to pass.

679
00:33:40,283 --> 00:33:44,354
And in the meantime, don't
change any of the data flattening

680
00:33:44,354 --> 00:33:46,271
that lets me get rich."

681
00:33:49,028 --> 00:33:53,792
Cool. Cool, man. Smart!

682
00:33:53,792 --> 00:33:55,375
Cards on the table,

683
00:33:55,375 --> 00:33:58,975
I don't think Altman and
people like him are facing

684
00:33:58,975 --> 00:34:02,702
the real challenges posed by AI at all.

685
00:34:02,702 --> 00:34:06,833
I think they might be
doing something else.

686
00:34:06,833 --> 00:34:10,090
Altman owns a company called Worldcoin.

687
00:34:10,090 --> 00:34:12,228
They scan people's irises and pay them

688
00:34:12,228 --> 00:34:14,262
for that data in cryptocurrency.

689
00:34:14,262 --> 00:34:16,905
I've seen it advertised on
TikTok! And that might sound

690
00:34:16,905 --> 00:34:18,173
like data owning democracy

691
00:34:18,173 --> 00:34:20,199
'cause like we're paying
people for that data, right?

692
00:34:20,199 --> 00:34:22,792
But then you think about it
for more than five minutes

693
00:34:22,792 --> 00:34:27,252
and you realize, okay,
so Altman thinks that AI

694
00:34:27,252 --> 00:34:28,825
is gonna wipe out people's jobs

695
00:34:28,825 --> 00:34:31,223
so people need universal basic income.

696
00:34:31,223 --> 00:34:34,973
And his plan for that
is to give them crypto?

697
00:34:37,197 --> 00:34:39,805
So can you actually spend

698
00:34:39,805 --> 00:34:44,331
that on anything like rent or groceries? No...

699
00:34:44,331 --> 00:34:50,818
In fact, I know from a lot of my friends that 
crypto is often used to pay OnlyFans girls

700
00:34:50,818 --> 00:34:54,772
so the men using it don't
get found out by their wives!

701
00:34:54,772 --> 00:34:58,118
If you're married to a crypto bro, check his browser history!

702
00:34:58,118 --> 00:35:00,117
But other than paying for porn,

703
00:35:00,117 --> 00:35:02,876
how is this actually helping?

704
00:35:02,876 --> 00:35:05,357
Turning Worldcoin into
global universal basic income

705
00:35:05,357 --> 00:35:07,938
would take a huge amount of
legislation and infrastructure

706
00:35:07,938 --> 00:35:10,790
that Altman has no plan, no authority,

707
00:35:10,790 --> 00:35:12,542
and no ability to deliver,

708
00:35:12,542 --> 00:35:15,434
which might explain why
within a month of its launch,

709
00:35:15,434 --> 00:35:18,661
Worldcoin fell by 44% and its offices

710
00:35:18,661 --> 00:35:20,494
were raided by police.

711
00:35:21,726 --> 00:35:23,530
Nice job, dude.

712
00:35:23,530 --> 00:35:27,396
It seems like Altman's
real plan is to tell

713
00:35:27,396 --> 00:35:29,953
a scary story about AI,

714
00:35:29,953 --> 00:35:33,819
dazzle some investors,
harvest people's data,

715
00:35:33,819 --> 00:35:36,819
and pump the value of a crypto scam.

716
00:35:38,171 --> 00:35:40,225
This isn't making ethical AI.

717
00:35:40,225 --> 00:35:42,784
This isn't engaging with
any of the real questions

718
00:35:42,784 --> 00:35:45,752
it throws up about private
property or anything else.

719
00:35:45,752 --> 00:35:49,170
And a lot of AI chat is like this.

720
00:35:49,170 --> 00:35:53,800
It's not ethics; it's marketing!

721
00:35:54,773 --> 00:35:59,327
Hi! I'm Kelly Slaughter,
founder and CEO of Allie,

722
00:35:59,327 --> 00:36:02,275
the world's first feminine AI company!

723
00:36:02,275 --> 00:36:04,787
AI is transforming our world,

724
00:36:04,787 --> 00:36:07,922
unlocking quadrillions
of dollars of value,

725
00:36:07,922 --> 00:36:10,929
but we need to share the
benefits with everybody.

726
00:36:10,929 --> 00:36:14,317
That's why I'm proud to introduce Allie,

727
00:36:14,317 --> 00:36:17,884
artificial intelligence by women for women!

728
00:36:17,884 --> 00:36:20,334
Our story begins a few
years ago when I was

729
00:36:20,334 --> 00:36:22,390
at a business conference in Switzerland

730
00:36:22,390 --> 00:36:24,898
about lethal autonomous weapons.

731
00:36:24,898 --> 00:36:28,520
Everybody was talking about the
military applications of AI,

732
00:36:28,520 --> 00:36:31,207
how it could enhance our
battlefield capabilities

733
00:36:31,207 --> 00:36:34,355
and increase our power
to inflict violence.

734
00:36:34,355 --> 00:36:37,383
And I looked around the
room and I realized,

735
00:36:37,383 --> 00:36:38,300
"Oh my god,

736
00:36:40,351 --> 00:36:43,192
I'm the only woman here."

737
00:36:43,192 --> 00:36:45,858
Women are so underrepresented in STEM.

738
00:36:45,858 --> 00:36:48,449
I had to work so much harder
than my male colleagues

739
00:36:48,449 --> 00:36:49,801
to become a founder.

740
00:36:49,801 --> 00:36:53,509
When I left Harvard, I
had nothing but a dream

741
00:36:53,509 --> 00:36:55,822
and my dad's software company.

742
00:36:55,822 --> 00:37:02,286
I sacrificed so much; family, health, sleep, friendships,

743
00:37:02,934 --> 00:37:05,681
... my marriage.

744
00:37:07,394 --> 00:37:09,236
And it's all been worth i!

745
00:37:09,236 --> 00:37:12,225
Because I followed that
dream and now I'm here!

746
00:37:12,225 --> 00:37:16,549
I believe every woman should have the opportunity to follow her dream

747
00:37:16,549 --> 00:37:18,058
by working for me.

748
00:37:18,058 --> 00:37:20,286
And that's when I have the idea for Allie!

749
00:37:20,286 --> 00:37:24,961
Just like the journey of
a thousand miles begins with a single step,

750
00:37:24,961 --> 00:37:28,206
the most sophisticated
artificial minds are assembled

751
00:37:28,206 --> 00:37:32,252
from millions of tiny
computer-based tasks!

752
00:37:32,252 --> 00:37:34,438
We're giving women in refugee camps

753
00:37:34,438 --> 00:37:37,386
and prisons the chance at a global family

754
00:37:37,386 --> 00:37:41,007
by empowering them to
become micro-entrepreneurs.

755
00:37:41,007 --> 00:37:45,392
Teaching our algorithms
to recognize, text, faces,

756
00:37:45,392 --> 00:37:47,809
and more one click at a time.

757
00:37:48,816 --> 00:37:52,762
The tasks are flexible, easy - and fun!

758
00:37:52,762 --> 00:37:54,807
Every one of our micro-entrepreneurs

759
00:37:54,807 --> 00:37:59,561
are guaranteed 100% female and
they get on the job childcare

760
00:37:59,561 --> 00:38:02,993
when they bring their daughters
to work alongside them.

761
00:38:02,993 --> 00:38:06,338
The data they provide go
toward building Allie,

762
00:38:06,338 --> 00:38:07,987
our artificially intelligent

763
00:38:07,987 --> 00:38:11,166
but genuinely feminine AI system.

764
00:38:11,166 --> 00:38:13,901
Maybe she'll be part of
a self-driving vehicle,

765
00:38:13,901 --> 00:38:17,677
facial recognition
software, a security system.

766
00:38:17,677 --> 00:38:19,961
The possibilities are endless.

767
00:38:19,961 --> 00:38:24,234
We need smart business to
win the AI war with China

768
00:38:24,234 --> 00:38:27,982
and I believe diversity
is a strategic priority!

769
00:38:27,982 --> 00:38:35,688
When we make the next generation of lethal autonomous weapons, 
half those weapons should be made by women!

770
00:38:35,688 --> 00:38:39,103
The technology we're building
at Allie has the potential

771
00:38:39,103 --> 00:38:41,507
to reshape the world.

772
00:38:41,507 --> 00:38:45,840
It's not intelligence, it's
artificial intelligence!

773
00:38:51,519 --> 00:39:01,431
[intense music like OOOOOHHHHH HERE WE GO BAYBEEEEE
DUM DUM DUM DUM DUM DUM DUM DUM]

774
00:39:01,431 --> 00:39:04,835
At the risk of sounding very obvious,

775
00:39:04,835 --> 00:39:09,835
if you wanna make AI, you
need electricity! Cables,

776
00:39:09,839 --> 00:39:12,187
power plants... lithium!

777
00:39:12,187 --> 00:39:14,569
Lithium is used in rechargeable batteries.

778
00:39:14,569 --> 00:39:18,370
It's in laptops and
cameras and electric cars.

779
00:39:18,370 --> 00:39:26,158
I own a small battery powered device, 
which almost never leaves my hand, 
which needs lithium to function.

780
00:39:26,158 --> 00:39:28,958
And as well as my vibrator,
I also own a smartphone.

781
00:39:28,958 --> 00:39:31,089
Lithium has to be mined out of the ground

782
00:39:31,089 --> 00:39:34,222
and then taken to a refinery
on trucks and ships.

783
00:39:34,222 --> 00:39:35,980
It has to be turned into components.

784
00:39:35,980 --> 00:39:38,004
Those also have to be transported.

785
00:39:38,004 --> 00:39:42,869
Every step uses up non-renewable
resources and emits CO2.

786
00:39:42,869 --> 00:39:47,452
This is all to say that AI
is not a free-floating mind.

787
00:39:48,644 --> 00:39:51,394
It is a physical, material thing.

788
00:39:52,261 --> 00:39:55,153
Not just material, but dirty.

789
00:39:55,153 --> 00:39:58,966
Mining in particular causes
a lot of damage to people

790
00:39:58,966 --> 00:40:00,090
and the environment.

791
00:40:00,090 --> 00:40:02,979
Damage that can be expressed
in terms of financial costs,

792
00:40:02,979 --> 00:40:05,737
like 'X billion dollars worth of damage.'

793
00:40:05,737 --> 00:40:08,063
The cost of digging
stuff out of the ground

794
00:40:08,063 --> 00:40:12,768
is almost always more than
the value of what you dig.

795
00:40:12,768 --> 00:40:16,320
The only way mining works as a business is

796
00:40:16,320 --> 00:40:20,047
because mining corporations
don't pay that cost.

797
00:40:20,047 --> 00:40:23,667
The cost to the miners'
health is borne by the miners.

798
00:40:23,667 --> 00:40:27,361
The cost to the environment
is borne by everyone.

799
00:40:27,361 --> 00:40:29,632
It's worth remembering
that because the idea

800
00:40:29,632 --> 00:40:31,221
that the tech sector including,

801
00:40:31,221 --> 00:40:38,300
but not limited to AI, is
"clean," is a very profitable lie.

802
00:40:38,897 --> 00:40:40,434
[Voice of Jordan Harrod]
"The term artificial intelligence

803
00:40:40,434 --> 00:40:43,251
may invoke ideas of algorithms, data,

804
00:40:43,251 --> 00:40:45,249
and computer architectures,

805
00:40:45,249 --> 00:40:48,615
but none of that can
function without the minerals

806
00:40:48,615 --> 00:40:52,425
and resources that build
computing's core components.

807
00:40:52,425 --> 00:40:55,372
Rechargeable lithium ion
batteries are essential

808
00:40:55,372 --> 00:40:58,558
for mobile devices and laptops
in-home digital assistance

809
00:40:58,558 --> 00:41:00,839
and data center backup power.

810
00:41:00,839 --> 00:41:03,144
They undergird the internet
and every commerce platform

811
00:41:03,144 --> 00:41:05,408
that runs on it, from banking to retail

812
00:41:05,408 --> 00:41:07,686
to stock market trades.

813
00:41:07,686 --> 00:41:09,685
The cloud is the backbone

814
00:41:09,685 --> 00:41:12,155
of the artificial intelligence industry

815
00:41:12,155 --> 00:41:16,894
and it's made of rocks and
lithium, brine and crude oil."

816
00:41:16,894 --> 00:41:20,064
- The other thing you
need to make AI is people.

817
00:41:20,064 --> 00:41:22,168
Somebody has to go down the
lithium mine and dig it up.

818
00:41:22,168 --> 00:41:23,456
Somebody has to work in the refinery,

819
00:41:23,456 --> 00:41:25,010
drive the trucks and the ships.

820
00:41:25,010 --> 00:41:27,497
Like we said before,
somebody has to produce all

821
00:41:27,497 --> 00:41:28,624
that training data.

822
00:41:28,624 --> 00:41:33,624
Someone also has to label that
data: to tell the algorithm

823
00:41:33,979 --> 00:41:36,215
'that's a human face; that's an object,'

824
00:41:36,215 --> 00:41:38,047
'that's hate speech; that's free speech,'

825
00:41:38,047 --> 00:41:39,945
'that's a car; that's a bollard.'

826
00:41:39,945 --> 00:41:42,829
It actually takes a lot
of people to do that job.

827
00:41:42,829 --> 00:41:44,744
If you added up everyone who does it

828
00:41:44,744 --> 00:41:48,852
they would make up one of the
largest workforces on Earth.

829
00:41:48,852 --> 00:41:50,044
In the global south,

830
00:41:50,044 --> 00:41:51,285
that job is done by people in countries

831
00:41:51,285 --> 00:41:53,444
like the Philippines, India, and Uganda.

832
00:41:53,444 --> 00:41:55,115
Sometimes people in refugee camps,

833
00:41:55,115 --> 00:41:57,006
like in Kenya and Lebanon.

834
00:41:57,006 --> 00:41:59,920
In the global north,
it's done by prisoners,

835
00:41:59,920 --> 00:42:02,539
old people who can't afford to retire,

836
00:42:02,539 --> 00:42:04,084
and people who can't get other jobs

837
00:42:04,084 --> 00:42:06,809
because for example,
they have a disability.

838
00:42:06,809 --> 00:42:08,104
[Voice of F.D. Signifier]
"Behind the search engines,

839
00:42:08,104 --> 00:42:10,688
apps and smart devices stand workers,

840
00:42:10,688 --> 00:42:14,072
often those banished to the
margins of our global system,

841
00:42:14,072 --> 00:42:16,066
who for lack of other options,

842
00:42:16,066 --> 00:42:19,228
are forced to clean data
and oversee algorithms

843
00:42:19,228 --> 00:42:21,380
for little more than a few cents.

844
00:42:21,380 --> 00:42:24,077
The feeds of Facebook and Twitter may seem

845
00:42:24,077 --> 00:42:27,808
to wipe away violent content
with automated precision,

846
00:42:27,808 --> 00:42:30,882
but decisions about what
constitutes pornography

847
00:42:30,882 --> 00:42:33,857
or hate speech are not made by algorithms.

848
00:42:33,857 --> 00:42:36,769
A facial recognition camera
seems of its own volition

849
00:42:36,769 --> 00:42:38,636
to spot a face in the crowd,

850
00:42:38,636 --> 00:42:41,887
an autonomous truck to drive
without human involvement.

851
00:42:41,887 --> 00:42:44,827
But in reality, the
magic of machine learning

852
00:42:44,827 --> 00:42:47,173
is the grind of data labeling."

853
00:42:47,173 --> 00:42:51,439
- And as jobs go, data
labelling is pretty sh*t.

854
00:42:51,439 --> 00:42:54,680
One of the reasons I love being
an actress is that my boss,

855
00:42:54,680 --> 00:42:56,894
the director, is always a person.

856
00:42:56,894 --> 00:42:59,166
If I wanna try something
different, I can ask.

857
00:42:59,166 --> 00:43:01,848
If they want me to try something
different, they just ask.

858
00:43:01,848 --> 00:43:03,518
I don't get like an
alert on my phone saying,

859
00:43:03,518 --> 00:43:04,911
"You were 4% less melancholy

860
00:43:04,911 --> 00:43:07,929
in the previous take,
please adjust performance."

861
00:43:07,929 --> 00:43:10,985
Hiring decisions are made
by casting directors,

862
00:43:10,985 --> 00:43:12,819
not by computers.

863
00:43:12,819 --> 00:43:14,845
That's also why I like working on Nebula.

864
00:43:14,845 --> 00:43:18,173
On YouTube, there are
content moderation algorithms

865
00:43:18,173 --> 00:43:19,006
and advertiser guidelines

866
00:43:19,006 --> 00:43:20,598
that are all enforced by computers.

867
00:43:20,598 --> 00:43:22,189
But on Nebula, there's none of that.

868
00:43:22,189 --> 00:43:23,051
If I have a problem,

869
00:43:23,051 --> 00:43:26,633
I text Dave the CEO and
he replies straightaway.

870
00:43:26,633 --> 00:43:31,076
As a worker, I am more free
when I work for humans.

871
00:43:31,076 --> 00:43:33,659
Data labelling is not like that.

872
00:43:35,201 --> 00:43:37,235
It's called 'microwork'

873
00:43:37,235 --> 00:43:40,268
because you get a tiny task
like 'label these pictures'

874
00:43:40,268 --> 00:43:43,210
and a fraction of a dollar for
every task that you complete.

875
00:43:43,210 --> 00:43:46,206
You might do tasks for 20
different companies in a day,

876
00:43:46,206 --> 00:43:48,224
all coordinated through an online platform

877
00:43:48,224 --> 00:43:50,260
like Amazon's Mechanical Turk.

878
00:43:50,260 --> 00:43:53,530
It's very poorly paid, has
almost no workers' rights,

879
00:43:53,530 --> 00:43:58,836
you're overseen and assessed by algorithms so if 
something goes wrong and the software fires you

880
00:43:58,836 --> 00:44:00,412
for no reason and you're locked out

881
00:44:00,412 --> 00:44:02,468
there's no one you can complain to.

882
00:44:02,468 --> 00:44:04,681
And because it's spread
all over the world,

883
00:44:04,681 --> 00:44:07,009
it's very difficult to unionise.

884
00:44:07,009 --> 00:44:09,167
In his book "Work Without the Worker",

885
00:44:09,167 --> 00:44:10,470
Scholar Phil Jones says

886
00:44:10,470 --> 00:44:13,630
the real AI jobs-pocalypse is this.

887
00:44:13,630 --> 00:44:17,130
Not jobs getting taken away
but becoming de-skilled,

888
00:44:17,130 --> 00:44:20,299
atomized, more surveillance,
less worker power,

889
00:44:20,299 --> 00:44:22,833
and generally sh*t.

890
00:44:22,833 --> 00:44:25,313
These jobs pay so little
and are so precarious

891
00:44:25,313 --> 00:44:26,720
that the main difference between them

892
00:44:26,720 --> 00:44:28,963
and unemployment is that
when you're unemployed,

893
00:44:28,963 --> 00:44:31,450
you at least have more
control over your time!

894
00:44:31,450 --> 00:44:33,633
So Jones coins a new term

895
00:44:33,633 --> 00:44:36,216
to describe them - 'subemployment.'

896
00:44:37,118 --> 00:44:38,677
[Voice of F.D. Signifier]
"Subemployment describes work

897
00:44:38,677 --> 00:44:41,788
that is highly temporary,
casual, and contingent,

898
00:44:41,788 --> 00:44:44,849
work that involves large
amounts of unpaid labor,

899
00:44:44,849 --> 00:44:47,482
significant underemployment or high levels

900
00:44:47,482 --> 00:44:51,363
of in-work poverty, or work
that more often than not,

901
00:44:51,363 --> 00:44:54,040
no longer guarantees
a life any better than

902
00:44:54,040 --> 00:44:56,720
the most abject forms of unemployment.

903
00:44:56,720 --> 00:44:58,867
Workers do a few hours here and there,

904
00:44:58,867 --> 00:45:01,457
not enough to qualify them
for employment rights,

905
00:45:01,457 --> 00:45:04,001
but enough to prevent them
from claiming benefits

906
00:45:04,001 --> 00:45:06,460
and appearing in unemployment statistics,

907
00:45:06,460 --> 00:45:10,367
a situation so uncertain
that workers are compelled

908
00:45:10,367 --> 00:45:12,803
to accept hours whenever offered."

909
00:45:12,803 --> 00:45:15,348
- In the global north, a
lot of the jobs created

910
00:45:15,348 --> 00:45:19,421
since the 2008 financial
crash are subemployment.

911
00:45:19,421 --> 00:45:22,202
That shift was achieved
through mass incarceration

912
00:45:22,202 --> 00:45:24,203
and gutting welfare.

913
00:45:24,203 --> 00:45:25,923
So people were moved from benefits

914
00:45:25,923 --> 00:45:29,009
to either jail, or working poverty.

915
00:45:29,009 --> 00:45:31,549
In the global south,
slums and refugee camps

916
00:45:31,549 --> 00:45:34,136
are kept full by wars and border policies

917
00:45:34,136 --> 00:45:36,728
that northern countries
often have a hand in.

918
00:45:36,728 --> 00:45:41,228
Just like mining, AI
only works as a business

919
00:45:41,228 --> 00:45:43,006
because the true costs are borne

920
00:45:43,006 --> 00:45:46,166
by people who are presumed
because of their race

921
00:45:46,166 --> 00:45:47,603
or their national origin or their age

922
00:45:47,603 --> 00:45:50,353
or whatever to simply not matter.

923
00:45:52,445 --> 00:45:55,864
You see, it's all well and good
to say that private property

924
00:45:55,864 --> 00:45:58,225
is the cornerstone of civilisation,

925
00:45:58,225 --> 00:46:00,050
but some people's private property

926
00:46:00,050 --> 00:46:02,852
has always been a bit more
cornerstone than others!

927
00:46:02,852 --> 00:46:07,036
When we remember things
like slavery, colonialism,

928
00:46:07,036 --> 00:46:10,466
the subjugation of women as
unpaid labourers in the home,

929
00:46:10,466 --> 00:46:13,765
it starts to look like for the
last few centuries at least

930
00:46:13,765 --> 00:46:15,647
the real cornerstone of civilisation

931
00:46:15,647 --> 00:46:19,296
has been private property
for some people, yes,

932
00:46:19,296 --> 00:46:24,243
propped up by theft and
exploitation of marginalised people.

933
00:46:24,243 --> 00:46:28,744
For example, John Locke said that when you put your labor
power into natural resources,

934
00:46:28,744 --> 00:46:30,542
you make them your property?

935
00:46:30,542 --> 00:46:32,750
Well, he also said that if you're not doing

936
00:46:32,750 --> 00:46:35,474
stationary agriculture on a piece of land,

937
00:46:35,474 --> 00:46:37,042
you're not really labouring on it

938
00:46:37,042 --> 00:46:39,287
and therefore it's not
really your property.

939
00:46:39,287 --> 00:46:40,827
And that argument was then used

940
00:46:40,827 --> 00:46:44,498
as the explicit legal
justification for stealing land

941
00:46:44,498 --> 00:46:47,200
from nomadic indigenous
people in the United States,

942
00:46:47,200 --> 00:46:52,447
a theft that John Locke himself directly oversaw and profited from!

943
00:46:53,978 --> 00:46:57,619
This is a trend that we in the 21st Century

944
00:46:57,619 --> 00:47:00,993
are starting to recognise
now and hopefully address,

945
00:47:00,993 --> 00:47:02,671
because if we really believe

946
00:47:02,671 --> 00:47:04,964
that all human beings are equal,

947
00:47:04,964 --> 00:47:08,613
then we cannot also do all of that stuff.

948
00:47:08,613 --> 00:47:12,863
But AI doesn't disrupt that
trend, it continues it.

949
00:47:15,397 --> 00:47:16,230
At the start of the video,

950
00:47:16,230 --> 00:47:19,299
we noted how 'AI' is quite
a promiscuous term,

951
00:47:19,299 --> 00:47:21,814
it's used to describe all
different kinds of systems.

952
00:47:21,814 --> 00:47:25,428
So in her book, "Atlas of
AI" researcher Kate Crawford

953
00:47:25,428 --> 00:47:29,178
uses a different term,
'Large-Scale Computing.'

954
00:47:31,100 --> 00:47:35,099
And I think that really brings
home how these technologies

955
00:47:35,099 --> 00:47:36,881
aren't a new type of thing.

956
00:47:36,881 --> 00:47:39,416
We've had tools like this for ages.

957
00:47:39,416 --> 00:47:41,236
These ones are just
bigger and more powerful,

958
00:47:41,236 --> 00:47:43,908
which means they use up more resources,

959
00:47:43,908 --> 00:47:47,867
emit more CO2, and exploit more people.

960
00:47:47,867 --> 00:47:51,914
And yet, even though these
systems are very large

961
00:47:51,914 --> 00:47:55,040
and powerful, they're also vulnerable.

962
00:47:55,040 --> 00:47:58,905
For one thing, they're very
vulnerable to climate change.

963
00:47:58,905 --> 00:48:02,849
If there's, for example, a
drought in Central America

964
00:48:02,849 --> 00:48:05,536
and ships can't get through
the Panama Canal anymore,

965
00:48:05,536 --> 00:48:06,630
then the flow of components

966
00:48:06,630 --> 00:48:10,211
that Large-Scale Computing
needs can slow down.

967
00:48:10,211 --> 00:48:11,506
The more CO2 these systems emit,

968
00:48:11,506 --> 00:48:14,664
the more that's gonna happen
until eventually it will stop.

969
00:48:14,664 --> 00:48:18,667
These systems are also
vulnerable to workers.

970
00:48:18,667 --> 00:48:20,954
If the people working
in mining or shipping

971
00:48:20,954 --> 00:48:23,232
or data labelling are suddenly unable

972
00:48:23,232 --> 00:48:26,150
(or unwilling!) to do their job,

973
00:48:26,150 --> 00:48:27,861
then the system can stop.

974
00:48:27,861 --> 00:48:31,290
Again, climate change is
really gonna accelerate this.

975
00:48:31,290 --> 00:48:32,660
Do you really think
people in the Philippines

976
00:48:32,660 --> 00:48:34,394
are still gonna be
working for Silicon Valley

977
00:48:34,394 --> 00:48:36,960
when half their country's underwater?

978
00:48:36,960 --> 00:48:37,793
They're gonna be like,

979
00:48:37,793 --> 00:48:40,039
"Abigail, I don't care that
your vibrator isn't charging.

980
00:48:40,039 --> 00:48:41,602
I've got bigger problems!"

981
00:48:41,602 --> 00:48:45,665
Deliberate action by workers
can jam the system too.

982
00:48:45,665 --> 00:48:47,622
When Google employees
found out their company

983
00:48:47,622 --> 00:48:49,475
was working with the US
Department of Defense,

984
00:48:49,475 --> 00:48:50,660
they kicked up such a fuss

985
00:48:50,660 --> 00:48:52,461
that Google cancelled the contract.

986
00:48:52,461 --> 00:48:54,588
Amazon employees staged walkouts

987
00:48:54,588 --> 00:48:56,559
over the company's carbon footprint.

988
00:48:56,559 --> 00:48:58,932
When actors and writers
discovered that studios

989
00:48:58,932 --> 00:49:03,749
wanted to replace us with
computers, we went on strike.

990
00:49:03,749 --> 00:49:07,535
The more physical and spread
out these systems are,

991
00:49:07,535 --> 00:49:12,535
the more opportunity there
is to take a hammer to them.

992
00:49:12,660 --> 00:49:16,417
That's why striking is
such a powerful tool

993
00:49:16,417 --> 00:49:17,789
and why governments who act on behalf

994
00:49:17,789 --> 00:49:20,137
of fossil fuel companies are so keen

995
00:49:20,137 --> 00:49:22,331
to suppress worker power.

996
00:49:22,331 --> 00:49:24,498
So can we make ethical AI?

997
00:49:27,458 --> 00:49:28,291
Well,

998
00:49:31,961 --> 00:49:35,211
who's we?

999
00:49:35,211 --> 00:49:40,549
[synth drone music like ooooooOOOOOoooo damn!]

1000
00:49:40,549 --> 00:49:44,132
Now we can return to
the alignment problem.

1001
00:49:46,512 --> 00:49:50,845
Suppose that a hostile AGI
came into being tomorrow.

1002
00:49:52,036 --> 00:49:56,864
We understand now that large
scale computing needs resources

1003
00:49:56,864 --> 00:49:59,996
and labour, and if our AGI
had an army of terminators

1004
00:49:59,996 --> 00:50:01,614
it could probably get
those things on its own,

1005
00:50:01,614 --> 00:50:03,967
but it's gonna have to build
those terminators first

1006
00:50:03,967 --> 00:50:06,205
and the factories that build them.

1007
00:50:06,205 --> 00:50:09,929
So more likely it would need people,

1008
00:50:09,929 --> 00:50:14,166
recruiting them through
trickery, force, or persuasion.

1009
00:50:14,166 --> 00:50:18,916
Which means the power of a
hostile AGI would still depend

1010
00:50:19,933 --> 00:50:22,930
on its relation to our existing political

1011
00:50:22,930 --> 00:50:24,908
and social systems.

1012
00:50:24,908 --> 00:50:28,257
It is a mistake to think
that a disembodied mind

1013
00:50:28,257 --> 00:50:31,035
could destroy the world
from inside the cloud.

1014
00:50:31,035 --> 00:50:33,626
Even if we built badly aligned AGI,

1015
00:50:33,626 --> 00:50:36,701
its power would come not from technology,

1016
00:50:36,701 --> 00:50:39,321
but from the human systems in which

1017
00:50:39,321 --> 00:50:42,307
that technology is embedded.

1018
00:50:42,307 --> 00:50:45,938
So when people talk about ethical AI,

1019
00:50:45,938 --> 00:50:48,960
maybe we shouldn't think about Skynet.

1020
00:50:48,960 --> 00:50:52,006
Maybe we should think instead
about working conditions,

1021
00:50:52,006 --> 00:50:56,011
climate change, and how to
make the economy serve humans

1022
00:50:56,011 --> 00:50:58,289
rather than the other way around.

1023
00:50:58,289 --> 00:51:01,214
We shouldn't assume that
AGI is inevitable and try

1024
00:51:01,214 --> 00:51:03,833
and patch ethics on as an afterthought,

1025
00:51:03,833 --> 00:51:07,453
but rather ask how we can make
justice and sustainability

1026
00:51:07,453 --> 00:51:11,608
and fairness the goal of
every piece of tech we build.

1027
00:51:11,608 --> 00:51:14,630
If we really want to make ethical AI,

1028
00:51:14,630 --> 00:51:18,630
then we might like to
consider this perspective.

1029
00:51:20,134 --> 00:51:25,966
There is no ethical computation under capitalism!

1030
00:51:28,795 --> 00:51:30,954
It feels a bit weird to do the Nebula ad

1031
00:51:30,954 --> 00:51:33,578
after that emotional
ending, but like I said,

1032
00:51:33,578 --> 00:51:37,195
the people who make this
show work really hard,

1033
00:51:37,195 --> 00:51:39,826
and if you wanna help us then you can!

1034
00:51:39,826 --> 00:51:43,505
Nebula is a streaming service
that's partially owned

1035
00:51:43,505 --> 00:51:46,234
by the creators on it, including me.

1036
00:51:46,234 --> 00:51:49,309
So we're trying not to be uber-capitalist.

1037
00:51:49,309 --> 00:51:53,209
A while ago, the Philosophy
Tube crew made a documentary

1038
00:51:53,209 --> 00:51:55,126
about how we make this show -

1039
00:51:55,126 --> 00:51:57,550
it goes into my research
and writing process -

1040
00:51:57,550 --> 00:51:59,812
and if you'd like to
see it, click the link

1041
00:51:59,812 --> 00:52:03,907
in the description,
Go.Nebula.TV/philosophytube.

1042
00:52:03,907 --> 00:52:05,795
If you use that link specifically

1043
00:52:05,795 --> 00:52:07,516
then I get a little
bit of your signup fee,

1044
00:52:07,516 --> 00:52:12,297
which helps us, and you get a
40% discount on an annual plan.

1045
00:52:12,297 --> 00:52:16,915
So it works out at $2.50
a month. Pretty good!

1046
00:52:16,915 --> 00:52:19,219
Every episode of Philosophy
Tube is released early

1047
00:52:19,219 --> 00:52:21,212
and uncensored on Nebula.

1048
00:52:21,212 --> 00:52:23,788
Other creators like F.D.
Signifier and Lindsay Ellis

1049
00:52:23,788 --> 00:52:25,683
make exclusive content on there.

1050
00:52:25,683 --> 00:52:28,044
Nebula are also stepping
up their production

1051
00:52:28,044 --> 00:52:29,604
of original content.

1052
00:52:29,604 --> 00:52:31,673
You might remember last year I wrote

1053
00:52:31,673 --> 00:52:34,591
an award-winning stage
play called "The Prince"?

1054
00:52:34,591 --> 00:52:36,527
A lot of you came to see it in London

1055
00:52:36,527 --> 00:52:39,118
and the recording is available on Nebula.

1056
00:52:39,118 --> 00:52:43,951
If you enjoyed that, then you
might be interested to know...

1057
00:52:44,342 --> 00:52:47,129
I'm writing something new!

1058
00:52:48,891 --> 00:52:50,999
See you on Nebula!

1059
00:52:50,999 --> 00:53:08,147
[sick-ass electronic music courtesy of Nina Richards]

1060
00:53:08,147 --> 00:53:19,600
[What do you reckon? I think this one turned out great!]

1061
00:53:19,600 --> 00:53:37,213
[I pushed the boat out a bit. The last one we did was the biggest hit we've ever had - a million views in under a week - so I wanted to push the boat out this time.]

1062
00:53:37,213 --> 00:53:53,470
[We filmed this over two days in two separate venues, whereas normally we film in one day! Two days' filming means double the cost but, yeah like I said - wanted to push the boat out a bit! Put the patrons' money on the screen, y'know?]

1063
00:53:53,470 --> 00:54:13,175
[I ended up making a really dumb mistake with it though. You know the Nebula signup link in the description? It wasn't clickable for the first 3 days of the video being out and we only just noticed!]

1064
00:54:13,175 --> 00:54:29,471
[So basically, half as many people signed up to Nebula for the first three days as they normally do. Since Nebula pays me according to how many people sign up, it basically means that on the NEXT episode I'll take a 50% pay cut. Such a dumb, dumb, mistake and it's gonna cost me bigtime.]

1065
00:54:29,471 --> 00:54:44,915
[I'll be fine, I have enough to cover the gap, I'm just kicking myself cause it's a big chunk of money to just set fire to over a tiny error. Won't make that mistake again!]

1066
00:54:44,915 --> 00:55:02,800
[Bad timing too. The Patreon took a hit cause student loan repayments started up again, and this video also cost 2x as much to make cause I pushed the boat out and filmed it over two days. I am absolutely kicking myself!]

1067
00:55:02,800 --> 00:55:27,285
[Anyway - uhhhh, sign up to Nebula if you can lol, and do use the link in the description. It's working now.]

1068
00:55:27,285 --> 00:55:47,586
[Ugh, I'm so mad at myself.]

1069
00:55:47,586 --> 00:56:04,192
[I hate letting people down, that's like my biggest fear. And Nebula put a lot of trust in me generally - and this time I blew it. It costs them and it costs me. Amateur mistake.]

1070
00:56:04,192 --> 00:56:19,674
[Anyway. We live and learn. I am writing a new thing that'll hopefully be going on Nebula - fingers crossed this won't affect my ability to pitch them stuff.]

1071
00:56:19,674 --> 00:56:32,334
[I hope we get to make it: I want it to be my next big creative project. We shall see!]

1072
00:56:32,334 --> 00:56:34,786
[Voice of Dan Olson]
"If you look at the progress from model to model,

1073
00:56:34,786 --> 00:56:37,702
even some of our biggest
critics are like, wow!"