WEBVTT

1
00:00:00.300 --> 00:00:01.610
<v ->Have you noticed that facial recognition</v>

2
00:00:01.610 --> 00:00:03.130
is everywhere at the moment?

3
00:00:03.130 --> 00:00:05.810
It's how we unlock our phones, tag people on Facebook,

4
00:00:05.810 --> 00:00:08.500
and at some US airports, it's being used to speed up

5
00:00:08.500 --> 00:00:10.830
the process of boarding an aircraft.

6
00:00:10.830 --> 00:00:13.450
But there are also concerns about this technology as well

7
00:00:13.450 --> 00:00:15.830
from lawmakers and civil rights groups.

8
00:00:15.830 --> 00:00:18.820
I'm here with Gretchen Greene, a computer vision expert

9
00:00:18.820 --> 00:00:21.530
and lawyer, who's going to help us understand

10
00:00:21.530 --> 00:00:23.770
this technology, thank you for joining us Gretchen.

11
00:00:23.770 --> 00:00:25.120
<v ->Hi Tom, great to be here.</v>

12
00:00:25.120 --> 00:00:27.460
<v ->The uses that are really driving</v>

13
00:00:27.460 --> 00:00:29.790
the public debate right now, are around

14
00:00:29.790 --> 00:00:32.480 line:15% 
the government uses, and so we've seen

15
00:00:32.480 --> 00:00:34.080 line:15% 
that the city of San Francisco where we are now

16
00:00:34.080 --> 00:00:36.340 line:15% 
recently banned use of facial recognition

17
00:00:36.340 --> 00:00:39.180 line:15% 
by it's agencies, Oakland across the bay,

18
00:00:39.180 --> 00:00:42.020
did so too, and so did some of lower Massachusetts.

19
00:00:42.020 --> 00:00:44.720
It's very unusual for a government at any level

20
00:00:44.720 --> 00:00:48.469 line:15% 
to completely ban a technology, what's driving this?

21
00:00:48.469 --> 00:00:51.490 line:15% 
<v ->Right, so one of the uses, or maybe the primary use</v>

22
00:00:51.490 --> 00:00:54.600 line:15% 
that I think drives that is law enforcement's use,

23
00:00:54.600 --> 00:00:58.220
and so besides it borders, it's also local government.

24
00:00:58.220 --> 00:01:00.540
You've seen it on TV episodes,

25
00:01:00.540 --> 00:01:04.980
where on crime shows, they just sort of search,

26
00:01:04.980 --> 00:01:07.030
and it's like, oh that's who that is,

27
00:01:07.030 --> 00:01:08.930
one of the issues is how that intersects

28
00:01:08.930 --> 00:01:11.320
with overall surveillance possibilities.

29
00:01:11.320 --> 00:01:13.810
The thing which is unprecedented right now is the number

30
00:01:13.810 --> 00:01:15.890
of surveillance cameras, both private and public.

31
00:01:15.890 --> 00:01:18.910
If you were to connect those into

32
00:01:18.910 --> 00:01:21.120
a network where it was easy to get that data

33
00:01:21.120 --> 00:01:23.780
in a continuous way, and then you combine it

34
00:01:23.780 --> 00:01:25.470
with tools like facial recognition

35
00:01:25.470 --> 00:01:28.320
that could allow the automated processing of the data.

36
00:01:28.320 --> 00:01:33.320
You could track everyone, potentially, all of the time,

37
00:01:34.100 --> 00:01:36.970
backwards to as far as you had surveillance data.

38
00:01:36.970 --> 00:01:40.334
Now we're not there, but if you've got a camera

39
00:01:40.334 --> 00:01:42.170
there's still the question of can the police

40
00:01:42.170 --> 00:01:44.730
get a search warrant, or can they just ask you for it.

41
00:01:44.730 --> 00:01:46.160
<v ->Tell us a little bit about this technology,</v>

42
00:01:46.160 --> 00:01:50.710
so how does it work and how are these systems being made?

43
00:01:50.710 --> 00:01:52.400
<v Gretchen>Ultimately, what it does</v>

44
00:01:52.400 --> 00:01:53.930
is it's finding patterns.

45
00:01:53.930 --> 00:01:55.910
And there's multiple layers that are finding

46
00:01:55.910 --> 00:01:58.220
different kinds of patterns, but you can imagine

47
00:01:58.220 --> 00:02:00.860
with your eyes that if you really simplified it,

48
00:02:00.860 --> 00:02:03.990
so you zoomed way out and you had this blurry image,

49
00:02:03.990 --> 00:02:05.880
there's kind of a dark band right here.

50
00:02:05.880 --> 00:02:09.190
So that's in one of the layers, one of the patterns

51
00:02:09.190 --> 00:02:12.860
that the model is picking up, and then you have to give it

52
00:02:12.860 --> 00:02:15.890
examples of pictures that you've labeled,

53
00:02:15.890 --> 00:02:18.670
either it can be a yes no, there's a face in the picture,

54
00:02:18.670 --> 00:02:21.410
there's not, or it can be where the face is,

55
00:02:21.410 --> 00:02:25.030
or it can be who the face is, and it might be millions

56
00:02:25.030 --> 00:02:27.140
of examples that you need to give it

57
00:02:27.140 --> 00:02:30.680
for it to figure out what are the patterns

58
00:02:30.680 --> 00:02:32.320
that it should be looking for.

59
00:02:32.320 --> 00:02:33.530
<v ->Where does that data come from?</v>

60
00:02:33.530 --> 00:02:35.630
<v ->Well it depends actually on who's using it.</v>

61
00:02:35.630 --> 00:02:37.880
Where does government get images,

62
00:02:37.880 --> 00:02:40.460
so I would say government databases,

63
00:02:40.460 --> 00:02:43.200
departments of motor vehicle, state department records,

64
00:02:43.200 --> 00:02:46.840
those kinds of records are starting to be assembled

65
00:02:46.840 --> 00:02:49.050
into databases that the FBI for instance

66
00:02:49.050 --> 00:02:51.350
has access to, Not all states but some.

67
00:02:51.350 --> 00:02:54.980
Where does, big tech get images from as a primary source?

68
00:02:54.980 --> 00:02:57.340
Well people post them all the time.

69
00:02:57.340 --> 00:03:00.200
I said what you would need is a picture of someone

70
00:03:00.200 --> 00:03:03.040
and a label that said who is this.

71
00:03:03.040 --> 00:03:05.310
So if you put up a picture or you see a picture

72
00:03:05.310 --> 00:03:08.110
and you say oh, well this is my friend Jack,

73
00:03:08.110 --> 00:03:10.520
and this is my friend Sue, now you've provided them

74
00:03:10.520 --> 00:03:12.930
that labeled training data that they can use

75
00:03:12.930 --> 00:03:15.270
for their facial recognition algorithms.

76
00:03:15.270 --> 00:03:17.340
It's also got much more accessible,

77
00:03:17.340 --> 00:03:19.840
so not only the big tech companies,

78
00:03:19.840 --> 00:03:22.810
with deep pockets and a lot of scientists

79
00:03:22.810 --> 00:03:26.510
can do this sort of thing, so it's

80
00:03:26.510 --> 00:03:28.610
been proliferating, absolutely.

81
00:03:28.610 --> 00:03:30.600
<v ->They're showing up in a lot of different places,</v>

82
00:03:30.600 --> 00:03:32.270
commercial, all kinds of uses.

83
00:03:32.270 --> 00:03:33.300
How are they being used?

84
00:03:33.300 --> 00:03:35.430
<v ->Private security, so it's being used</v>

85
00:03:35.430 --> 00:03:38.690
at large events, there was a report in China

86
00:03:38.690 --> 00:03:41.360
about someone who had an outstanding warrant

87
00:03:41.360 --> 00:03:44.700
or the equivalent going to a big concert

88
00:03:44.700 --> 00:03:48.380
and being identified by facial recognition.

89
00:03:48.380 --> 00:03:51.510
It's being used for advertisements,

90
00:03:51.510 --> 00:03:55.210
a closely related technology is a motion recognition,

91
00:03:55.210 --> 00:03:57.190
which can be done in a lot of ways,

92
00:03:57.190 --> 00:03:59.480
but one of them is through looking at the face.

93
00:03:59.480 --> 00:04:01.420
So if you can see how someone reacts

94
00:04:01.420 --> 00:04:04.210
to an electronic billboard outside a store,

95
00:04:04.210 --> 00:04:06.030
you can change the electronic billboard

96
00:04:06.030 --> 00:04:07.640
to try to make it more enticing to them.

97
00:04:07.640 --> 00:04:09.660
It's also being used in education

98
00:04:09.660 --> 00:04:12.830
with robot human interactions, having a robot

99
00:04:12.830 --> 00:04:15.970
that has a nice personality, but trying to interact,

100
00:04:15.970 --> 00:04:19.720
and so it has to understand when someone is looking at it.

101
00:04:19.720 --> 00:04:22.040
<v ->Yeah, so a really broad span, and lots of them</v>

102
00:04:22.040 --> 00:04:23.870
seem fairly innocuous, some of them

103
00:04:23.870 --> 00:04:24.980
I guess may seem a bit unusual,

104
00:04:24.980 --> 00:04:27.750
we don't expect billboards to scan our faces

105
00:04:27.750 --> 00:04:30.620
or look at us, and why does the government

106
00:04:30.620 --> 00:04:33.371
want this technology, so in law enforcement for example

107
00:04:33.371 --> 00:04:36.887
what benefits do they see from having the ability

108
00:04:36.887 --> 00:04:39.150
to search faces in a database?

109
00:04:39.150 --> 00:04:41.400
<v ->Law enforcement is interested in tools</v>

110
00:04:41.400 --> 00:04:42.930
to help them do their job better.

111
00:04:42.930 --> 00:04:46.260
If you've got a database full of pictures

112
00:04:46.260 --> 00:04:49.660
and their identities, and then you have a picture

113
00:04:49.660 --> 00:04:52.940
of somebody from the 7/11 that got robbed.

114
00:04:52.940 --> 00:04:56.460
Right now you would take that picture and pass it around

115
00:04:56.460 --> 00:04:59.360
to the officers who do work in that area,

116
00:04:59.360 --> 00:05:00.630
you'd maybe go around the neighborhood,

117
00:05:00.630 --> 00:05:02.460
say does anybody know who this is.

118
00:05:02.460 --> 00:05:04.140
Well one, that'll take longer,

119
00:05:04.140 --> 00:05:06.410
but the other thing is, maybe nobody knows.

120
00:05:06.410 --> 00:05:09.270
And maybe that picture is somewhere in this database,

121
00:05:09.270 --> 00:05:12.400
but your database has 10,000 pictures in it.

122
00:05:12.400 --> 00:05:15.940
It'll take a very long time to go one by one.

123
00:05:15.940 --> 00:05:18.580
It'll matter probably less in the case

124
00:05:18.580 --> 00:05:20.690
of a 7/11 that you find them quickly,

125
00:05:20.690 --> 00:05:23.140
but the longer that it takes to solve a case,

126
00:05:23.140 --> 00:05:25.030
the less likely that it is to be solved,

127
00:05:25.030 --> 00:05:28.060
and there are other kinds of crimes, kidnappings,

128
00:05:28.060 --> 00:05:30.760
where it's very important to solve it quickly.

129
00:05:30.760 --> 00:05:33.480
<v ->Raises a question that a lot of civil liberties</v>

130
00:05:33.480 --> 00:05:37.120
groups ask, which is about, what if the facial recognition

131
00:05:37.120 --> 00:05:40.450
software is wrong, what if it misidentifies someone,

132
00:05:40.450 --> 00:05:43.890
and I've heard concerns that there'll be

133
00:05:43.890 --> 00:05:45.390
different error rates for different

134
00:05:45.390 --> 00:05:47.900
demographics and different communities.

135
00:05:47.900 --> 00:05:50.420
<v ->Right, so there's two kinds of concerns to have.</v>

136
00:05:50.420 --> 00:05:52.780
What if it's right, should we be using something

137
00:05:52.780 --> 00:05:56.340
in a certain way at all, and then what if it's wrong,

138
00:05:56.340 --> 00:05:58.840
even if it's only wrong sometimes,

139
00:05:58.840 --> 00:06:00.770
what are the effects of it being wrong.

140
00:06:00.770 --> 00:06:02.520
It changes how the police officer

141
00:06:02.520 --> 00:06:05.250
will react, which could be good or bad.

142
00:06:05.250 --> 00:06:08.370
We want the officer to be safe,

143
00:06:08.370 --> 00:06:13.370
but they will be more likely to take more extreme action,

144
00:06:13.980 --> 00:06:15.520
and to think that they're in danger,

145
00:06:15.520 --> 00:06:16.930
when maybe they're actually not.

146
00:06:16.930 --> 00:06:19.430
So for instance if it's a misidentification,

147
00:06:19.430 --> 00:06:23.950
there is research out of MIT, where dark skin

148
00:06:23.950 --> 00:06:27.670
is not as likely to be correctly identified as light,

149
00:06:27.670 --> 00:06:29.690
particularly women with dark skin,

150
00:06:29.690 --> 00:06:31.970
not as likely to be correctly identified.

151
00:06:31.970 --> 00:06:34.920
<v ->There are also general concerns about</v>

152
00:06:34.920 --> 00:06:37.540
the effects on anyone from any community

153
00:06:37.540 --> 00:06:40.400
of just knowing that the government

154
00:06:40.400 --> 00:06:43.050
may be tracking your face, may be watching you.

155
00:06:43.050 --> 00:06:44.450
<v ->Even if I trust the government,</v>

156
00:06:44.450 --> 00:06:46.610
I do care, I would rather live in a world

157
00:06:46.610 --> 00:06:49.950
where I feel like I have some privacy,

158
00:06:49.950 --> 00:06:53.130
even in public spaces, that not all is known,

159
00:06:53.130 --> 00:06:55.710
because if people know where you are,

160
00:06:55.710 --> 00:06:59.040
you might not go there, you might not do those things,

161
00:06:59.040 --> 00:07:02.880
even though they're things that are bedrock of what we think

162
00:07:02.880 --> 00:07:04.950
people in this country should be able to do.

163
00:07:04.950 --> 00:07:08.730
For instance, coming out as gay

164
00:07:08.730 --> 00:07:12.050
is less problematic professionally now

165
00:07:12.050 --> 00:07:16.250
than it was in the US, but still potentially

166
00:07:16.250 --> 00:07:18.930
problematic, and so if an individual

167
00:07:18.930 --> 00:07:21.740
wants to make the choice when to publicly disclose that,

168
00:07:21.740 --> 00:07:26.090
then they don't want facial recognition technology

169
00:07:26.090 --> 00:07:29.410
identifying that they are walking down the street

170
00:07:29.410 --> 00:07:34.410
to the LGBTQ center, so there are kinds of membership

171
00:07:34.750 --> 00:07:37.280
issues of certain groups in society

172
00:07:37.280 --> 00:07:40.550
where we're not as a government trying to stop

173
00:07:40.550 --> 00:07:42.750
or as a society really trying to stop

174
00:07:42.750 --> 00:07:45.060
certain kinds of actions, we're not trying to stop

175
00:07:45.060 --> 00:07:47.200
people from going to church, we're not trying

176
00:07:47.200 --> 00:07:50.080
to stop them from going to community centers.

177
00:07:50.080 --> 00:07:53.580
But we will if they are afraid of

178
00:07:53.580 --> 00:07:56.330
what will the implications be in an environment

179
00:07:56.330 --> 00:07:59.180
that is hostile to, for instance,

180
00:07:59.180 --> 00:08:03.530
a certain ethnicity, or a certain religion.

181
00:08:03.530 --> 00:08:05.680
<v ->So really it's very difficult to opt out</v>

182
00:08:05.680 --> 00:08:09.150
of either government facial recgonition

183
00:08:09.150 --> 00:08:10.460
or commercial, it sounds like.

184
00:08:10.460 --> 00:08:12.850
<v ->It is very difficult, and that's one reason</v>

185
00:08:12.850 --> 00:08:16.190
that it is more controversial than some other things

186
00:08:16.190 --> 00:08:18.470
like fingerprints would be, because it can be done

187
00:08:18.470 --> 00:08:21.460
at a distance when you don't know it's being done,

188
00:08:21.460 --> 00:08:24.480
and in a way that it's very difficult to opt out.

189
00:08:24.480 --> 00:08:26.640
And some people might think that, well,

190
00:08:26.640 --> 00:08:28.220
there would be some kind of privacy laws,

191
00:08:28.220 --> 00:08:30.910
or something that might restrict ideas like that,

192
00:08:30.910 --> 00:08:33.180
I mean, is there any regulation at the federal

193
00:08:33.180 --> 00:08:35.050
or state level that specifically

194
00:08:36.200 --> 00:08:37.400
regulates facial recognition?

195
00:08:37.400 --> 00:08:40.640
<v ->We're seeing more happening on the local levels,</v>

196
00:08:40.640 --> 00:08:44.230
state or city, right now, than at the federal,

197
00:08:44.230 --> 00:08:46.210
the federal government has not said,

198
00:08:46.210 --> 00:08:48.700
we are regulating this and therefore

199
00:08:48.700 --> 00:08:51.860
local governments cannot, and then because of that

200
00:08:51.860 --> 00:08:54.700
we're seeing a patchwork of states and cities

201
00:08:54.700 --> 00:08:56.860
thinking about doing something.

202
00:08:56.860 --> 00:08:59.587
<v ->I wonder what do you think the future looks like?</v>

203
00:08:59.587 --> 00:09:02.120
Are we in a period that we will look back on

204
00:09:02.120 --> 00:09:05.170
and say, wow, this technology was really unfettered

205
00:09:05.170 --> 00:09:07.580
back then, but now we have some protections?

206
00:09:07.580 --> 00:09:08.840
Is that where things are going,

207
00:09:08.840 --> 00:09:12.800
or is it too late and this is just how it's gonna be?

208
00:09:12.800 --> 00:09:14.580
<v ->There aren't that many companies</v>

209
00:09:14.580 --> 00:09:17.550
where facial recognition is their quo business,

210
00:09:17.550 --> 00:09:22.000
and it's not deeply embedded in what the government is doing

211
00:09:22.000 --> 00:09:24.430
and how we're functioning, even if it were,

212
00:09:24.430 --> 00:09:26.670
I don't think it would be necessarily too late

213
00:09:26.670 --> 00:09:29.890
to say as a society, what are the implications

214
00:09:29.890 --> 00:09:33.280
and the effect that government use of technologies

215
00:09:33.280 --> 00:09:34.810
that can be used for surveillance,

216
00:09:34.810 --> 00:09:37.870
like facial recognition, can have on other rights.

217
00:09:37.870 --> 00:09:40.600
So freedom of speech, expression, religion,

218
00:09:40.600 --> 00:09:44.050
do we want this or not, this is a choice that we're making.

219
00:09:44.050 --> 00:09:46.840
And there are a number of ways to say that

220
00:09:46.840 --> 00:09:49.253
we don't wanna make that choice,

221
00:09:49.253 --> 00:09:53.410
or that we do, we should decide as a society.

222
00:09:53.410 --> 00:09:54.600
<v ->Thank you for joining us Gretchen,</v>

223
00:09:54.600 --> 00:09:57.130
it'll be interesting to see how all this plays out.

224
00:09:57.130 --> 00:09:58.230
<v ->Good to be here Tom.</v>

225
00:09:59.262 --> 00:10:02.179
[electronic music]