1
00:00:08,108 --> 00:00:09,709
>> THE FOURTH ANNUAL NLM

2
00:00:09,776 --> 00:00:12,212
LECTURE, OFFICE OF STRATEGIC

3
00:00:12,278 --> 00:00:13,646
INITIATIVES ORGANIZED THIS

4
00:00:13,713 --> 00:00:16,783
SERIES, TO RAISE AWARENESS

5
00:00:16,850 --> 00:00:20,987
AROUND SOCIETAL AND ETHICAL

6
00:00:21,054 --> 00:00:22,856
IMPLICATIONS IN CONDUCT OF

7
00:00:22,922 --> 00:00:23,123
RESEARCH.

8
00:00:23,189 --> 00:00:26,159
IT'S OUR HOPE THESE WILL FUEL

9
00:00:26,226 --> 00:00:27,894
IMPORTANT CONVERSATIONS AROUND

10
00:00:27,961 --> 00:00:30,530
NLM, NIH, BROADER BIOMEDICAL

11
00:00:30,597 --> 00:00:35,702
RESEARCH COMMUNITY.

12
00:00:35,769 --> 00:00:40,273
I'M PLEASED TODAY TO INTRODUCE

13
00:00:40,340 --> 00:00:42,242
OUR GUEST LECTURER MEREDITH

14
00:00:42,308 --> 00:00:45,311
BROUSSARD, ASSOCIATE PROFESSOR

15
00:00:45,378 --> 00:00:54,521
AT NEW YORK UNIVERSITY CARTER

16
00:00:54,587 --> 00:00:56,790
JOURNALISM INSTITUTE, AUTHOR OF

17
00:00:56,856 --> 00:00:58,992
"MORE THAN A GLITCH," WHICH CAME

18
00:00:59,058 --> 00:01:04,164
OUT LAST YEAR, AND AWARD-WINNING

19
00:01:04,230 --> 00:01:05,598
2018 BOOK "ARTIFICIAL

20
00:01:05,665 --> 00:01:07,333
UNINTELLIGENCE, HOW COMPUTERS

21
00:01:07,400 --> 00:01:08,535
MISUNDERSTAND THE WORLD,"

22
00:01:08,601 --> 00:01:10,103
RESEARCH ON ARTIFICIAL

23
00:01:10,170 --> 00:01:11,504
INTELLIGENCE AND INVESTIGATIVE

24
00:01:11,571 --> 00:01:13,573
REPORTING WITH FOCUS ON A.I.

25
00:01:13,640 --> 00:01:15,241
ETHICS USING DATA ANALYSIS FOR

26
00:01:15,308 --> 00:01:17,043
SOCIAL GOOD.

27
00:01:17,110 --> 00:01:18,945
FORMER FEATURES EDITOR AT THE

28
00:01:19,012 --> 00:01:20,180
"PHILADELPHIA INQUIRER," HAS SHE

29
00:01:20,246 --> 00:01:22,015
HAS WORKED AT A SOFTWARE

30
00:01:22,081 --> 00:01:25,218
DEVELOPER AT AT&T BELL LABS,

31
00:01:25,285 --> 00:01:29,055
M.I.T. MEDIA LAB, FEATURE

32
00:01:29,122 --> 00:01:29,656
STORIES AND ESSAYS APPEARED IN

33
00:01:29,722 --> 00:01:32,892
THE "NEW YORK TIMES," "THE

34
00:01:32,959 --> 00:01:36,930
ATLANTIC," SLATE AND BOX.

35
00:01:36,996 --> 00:01:38,698
TODAY'S LECTURE WILL HELP US

36
00:01:38,765 --> 00:01:42,135
THINK THROUGH  CONCRETE

37
00:01:42,202 --> 00:01:43,803
STRATEGIES TO MITIGATE BIAS IN

38
00:01:43,870 --> 00:01:49,075
OUR APPROACH TO DEVELOPING

39
00:01:49,142 --> 00:01:52,145
ARTIFICIAL INTELLIGENCE

40
00:01:52,212 --> 00:01:54,781
TECHNOLOGIES, HER INSIGHTS COMES

41
00:01:54,848 --> 00:01:55,748
AT A CRITICAL POINT.

42
00:01:55,815 --> 00:01:59,586
IN THE LAST TWO YEARS

43
00:01:59,652 --> 00:02:00,186
DISCUSSIONS AROUND TRUSTWORTHY

44
00:02:00,253 --> 00:02:01,454
A.I. HAVE TAKEN OFF.

45
00:02:01,521 --> 00:02:02,755
WE HAVE POWERFUL GUIDING

46
00:02:02,822 --> 00:02:04,824
DOCUMENTS SUCH AS THE BLUEPRINT

47
00:02:04,891 --> 00:02:07,961
FOR A.I. BILL OF RIGHTS FROM THE

48
00:02:08,027 --> 00:02:09,662
WHITE HOUSE OFFICE OF SCIENCE

49
00:02:09,729 --> 00:02:11,798
AND TECHNOLOGY POLICY IN OCTOBER

50
00:02:11,865 --> 00:02:13,433
OF 2022, AND A.I. RISK

51
00:02:13,500 --> 00:02:14,634
MANAGEMENT FRAMEWORK PUBLISHED

52
00:02:14,701 --> 00:02:17,036
BY THE NATIONAL INSTITUTES OF

53
00:02:17,103 --> 00:02:18,338
STANDARDS OF TECHNOLOGY, NIST,

54
00:02:18,404 --> 00:02:19,339
LAST YEAR.

55
00:02:19,405 --> 00:02:22,942
WE ALSO HAVE VERY CLEAR

56
00:02:23,009 --> 00:02:23,476
DIRECTIVES ON ADVANCING

57
00:02:23,543 --> 00:02:27,747
COMPREHENSIVE

58
00:02:27,814 --> 00:02:28,314
WHOLE-OF-GOVERNMENT STRATEGY

59
00:02:28,381 --> 00:02:29,849
INCLUDING AN EXECUTIVE ORDER ON

60
00:02:29,916 --> 00:02:30,817
THE SAFE, SECURE, AND

61
00:02:30,884 --> 00:02:33,219
TRUSTWORTHY DEVELOPMENT AND USE

62
00:02:33,286 --> 00:02:34,087
OF ARTIFICIAL INTELLIGENCE.

63
00:02:34,153 --> 00:02:39,125
SIGNED BY THE PRESIDENT IN

64
00:02:39,192 --> 00:02:40,860
OCTOBER 30 OF 2023.

65
00:02:40,927 --> 00:02:42,695
THESE EFFORTS ARE DEDICATED TO

66
00:02:42,762 --> 00:02:43,496
UNLOCKING POTENTIAL OF

67
00:02:43,563 --> 00:02:44,197
ARTIFICIAL INTELLIGENCE TO SOLVE

68
00:02:44,264 --> 00:02:47,901
SOME OF OUR MOST PRESSING

69
00:02:47,967 --> 00:02:48,735
CHALLENGES, ACCELERATING

70
00:02:48,801 --> 00:02:51,104
SCIENTIFIC DISCOVERY, WHILE

71
00:02:51,170 --> 00:02:52,205
PROTECTING AGAINST THE

72
00:02:52,272 --> 00:02:53,406
TECHNOLOGIES' POTENTIAL RISKS.

73
00:02:53,473 --> 00:02:57,777
AT NIH, THAT MEANS WORKING TO

74
00:02:57,844 --> 00:02:59,178
MITIGATE BIAS, PRESERVE PRIVACY,

75
00:02:59,245 --> 00:03:02,015
PROMOTE ETHICAL APPROACHES TO

76
00:03:02,081 --> 00:03:03,550
INTEGRATING A.I. AND BIOMEDICAL

77
00:03:03,616 --> 00:03:04,417
RESEARCH.

78
00:03:04,484 --> 00:03:05,752
THIS WILL SERVE TO ADVANCE

79
00:03:05,818 --> 00:03:07,287
EQUITABLE HEALTH OUTCOMES FOR

80
00:03:07,353 --> 00:03:09,055
COMMUNITIES ACROSS THE NATION.

81
00:03:09,122 --> 00:03:10,790
AS THE WORLD'S LARGEST

82
00:03:10,857 --> 00:03:11,991
BIOMEDICAL RESEARCH LIBRARY,

83
00:03:12,058 --> 00:03:13,059
WITH COMMITMENT TO PRESERVING

84
00:03:13,126 --> 00:03:15,361
AND PROTECTING THE PUBLIC TRUST,

85
00:03:15,428 --> 00:03:17,897
NLM PLAYS AN IMPORTANT ROLE IN

86
00:03:17,964 --> 00:03:19,432
REALIZING THIS GOAL AND

87
00:03:19,499 --> 00:03:20,533
AFFIRMING THIS COMMITMENT.

88
00:03:20,600 --> 00:03:23,503
WE HOUSE AND MAKE ACCESSIBLE

89
00:03:23,570 --> 00:03:24,904
VALUABLE DATA RESOURCES,

90
00:03:24,971 --> 00:03:25,772
HARNESSED TO DRIVE DISCOVERY.

91
00:03:25,838 --> 00:03:28,174
WE'RE A LEADER IN HEALTH DATA

92
00:03:28,241 --> 00:03:30,143
STANDARDS DEVELOPMENT, THAT

93
00:03:30,209 --> 00:03:35,415
ENABLES THE DISCOVERY AND WE'RE

94
00:03:35,481 --> 00:03:39,452
A DRIVER OF DATA SCIENCE THROUGH

95
00:03:39,519 --> 00:03:40,053
OUR EXTRAMURAL AND INTRAMURAL

96
00:03:40,119 --> 00:03:40,653
RESEARCH PROGRAMS.

97
00:03:40,720 --> 00:03:42,956
WITH THIS FRAMING IN MIND, I'M

98
00:03:43,022 --> 00:03:44,490
LOOKING FORWARD TO HEARING HER

99
00:03:44,557 --> 00:03:47,460
WALK US THROUGH EXAMPLES OF THE

100
00:03:47,527 --> 00:03:49,095
POTENTIAL CONSEQUENCES OF

101
00:03:49,162 --> 00:03:50,430
DEVELOPING ALGORITHMS WITHOUT

102
00:03:50,496 --> 00:03:51,397
CAREFUL CONSIDERATION OF

103
00:03:51,464 --> 00:03:52,599
POTENTIAL BIASES.

104
00:03:52,665 --> 00:03:55,468
EVEN WHEN THESE TECHNOLOGIES ARE

105
00:03:55,535 --> 00:03:59,706
DEVELOPED WITH BEST INTENTIONS.

106
00:03:59,772 --> 00:04:01,808
ALSO EAGER TO HEAR INSIGHTS ON

107
00:04:01,874 --> 00:04:02,775
PRACTICAL SOLUTIONS AND LEARN

108
00:04:02,842 --> 00:04:04,310
MORE ABOUT HER IDEAS FOR

109
00:04:04,377 --> 00:04:06,279
DESIGNING BETTER SYSTEMS FOR

110
00:04:06,346 --> 00:04:07,180
MORE EQUITABLE OUTCOMES.

111
00:04:07,246 --> 00:04:09,782
THIS LECTURE WILL BE FOLLOWED BY

112
00:04:09,849 --> 00:04:11,217
A QUESTION-AND-ANSWER PORTION.

113
00:04:11,284 --> 00:04:13,286
IF YOU HAVE ANY QUESTIONS, MAKE

114
00:04:13,353 --> 00:04:15,555
YOUR WAY TO A MICROPHONE THAT

115
00:04:15,622 --> 00:04:17,223
WILL BE PROVIDED, IF YOU'RE WITH

116
00:04:17,290 --> 00:04:18,524
US IN PERSON.

117
00:04:18,591 --> 00:04:22,929
IF JOINING ONLINE USE THE ASK A

118
00:04:22,996 --> 00:04:26,299
QUESTION FEATURE ON VIDEOCAST.

119
00:04:26,366 --> 00:04:29,469
PLEASE WELCOME ME NOW IN JOINING

120
00:04:29,535 --> 00:04:31,938
MS. BROUSSARD.

121
00:04:37,877 --> 00:04:39,078
>> IT'S GREAT TO BE WITH YOU.

122
00:04:39,145 --> 00:04:40,747
THANK YOU FOR HAVING ME.

123
00:04:40,813 --> 00:04:43,249
SO I'M GOING TO TALK TODAY ABOUT

124
00:04:43,316 --> 00:04:45,251
SOME IDEAS FROM MY MOST RECENT

125
00:04:45,318 --> 00:04:49,689
BACK, "MORE THAN A GLITCH,

126
00:04:49,756 --> 00:04:51,791
CONFRONTING RACE, GENDER, AND

127
00:04:51,858 --> 00:04:53,459
ABILITY BIAS IN TECH."

128
00:04:53,526 --> 00:04:54,994
YOU'RE A KICKOFF STOP ON THE

129
00:04:55,061 --> 00:04:56,763
PAPER BACK TOUR, BECAUSE THE

130
00:04:56,829 --> 00:04:58,064
PAPERBACK IS JUST RELEASING.

131
00:04:58,131 --> 00:05:00,700
THANK YOU FOR THAT.

132
00:05:00,767 --> 00:05:02,669
AND SPECIFICALLY, I WANT TO TALK

133
00:05:02,735 --> 00:05:05,938
A LITTLE BIT ABOUT CANCER.

134
00:05:06,005 --> 00:05:09,976
I WANT TO TALK ABOUT HOW LOOKING

135
00:05:10,043 --> 00:05:12,679
AT CANCER CAN HELP US UNDERSTAND

136
00:05:12,745 --> 00:05:13,813
ALGORITHMIC BIAS.

137
00:05:13,880 --> 00:05:15,915
I'M GOING TO DO THAT BY TELLING

138
00:05:15,982 --> 00:05:17,984
YOU A STORY ABOUT CANCER.

139
00:05:18,051 --> 00:05:21,154
BUT FIRST I'M GOING TO START BY

140
00:05:21,220 --> 00:05:22,488
TALKING ABOUT ARTIFICIAL

141
00:05:22,555 --> 00:05:22,922
INTELLIGENCE.

142
00:05:22,989 --> 00:05:25,224
SOMETHING THAT I THINK IS REALLY

143
00:05:25,291 --> 00:05:26,426
IMPORTANT TO EMPHASIZE WHEN

144
00:05:26,492 --> 00:05:29,395
WE'RE TALKING ABOUT A.I. WITH

145
00:05:29,462 --> 00:05:31,764
THE GENERAL PUBLIC IS NEED TO

146
00:05:31,831 --> 00:05:34,100
EMPHASIZE A.I. IS NOT MAGIC AND

147
00:05:34,167 --> 00:05:37,470
THAT IT IS NOT HOLLYWOOD VISION.

148
00:05:37,537 --> 00:05:38,671
OKAY?

149
00:05:38,738 --> 00:05:40,973
BECAUSE EVERYBODY YOU TALK TO,

150
00:05:41,040 --> 00:05:43,076
WELL, EVERY NON-A.I. SCIENTIST,

151
00:05:43,142 --> 00:05:44,177
IS LIKE -- ACTUALLY EVERYBODY

152
00:05:44,243 --> 00:05:47,647
YOU TALK TO IS THINKING ABOUT

153
00:05:47,714 --> 00:05:48,314
THE TERMINATOR.

154
00:05:48,381 --> 00:05:51,951
OR THEY ARE THINKING ABOUT "STAR

155
00:05:52,018 --> 00:06:00,259
WARS" OR "STAR TREK," COOL

156
00:06:00,326 --> 00:06:01,027
HOLLYWOOD PORTRAYALS OF A.I.

157
00:06:01,094 --> 00:06:02,595
BECAUSE OF A SPECIFIC WAY OUR

158
00:06:02,662 --> 00:06:03,463
BRAINS WORK.

159
00:06:03,529 --> 00:06:05,665
OUR BRAINS ARE BETTER AT

160
00:06:05,732 --> 00:06:07,934
RECALLING STORIES THAN FACTS AND

161
00:06:08,000 --> 00:06:08,334
SPECIFICS.

162
00:06:08,401 --> 00:06:11,003
SO, WE ALWAYS GO TO HOLLYWOOD

163
00:06:11,070 --> 00:06:11,437
FIRST.

164
00:06:11,504 --> 00:06:12,438
HOLLYWOOD TELLS AMAZING STORIES

165
00:06:12,505 --> 00:06:13,973
SO THE TERMINATOR IS STUCK IN

166
00:06:14,040 --> 00:06:15,808
OUR BRAINS.

167
00:06:15,875 --> 00:06:17,143
I WILL ADMIT TO HAVING SPENT A

168
00:06:17,210 --> 00:06:20,446
LOT OF TIME THINKING ABOUT THE

169
00:06:20,513 --> 00:06:21,314
TERMINATOR MYSELF.

170
00:06:21,380 --> 00:06:24,717
BUT IT'S NOT REAL.

171
00:06:24,784 --> 00:06:27,320
WE NEED TO EMPHASIZE WHEN WE'RE

172
00:06:27,386 --> 00:06:31,057
TALKING TO THE GENERAL PUBLIC

173
00:06:31,124 --> 00:06:33,392
THERE'S LOTS OF STUFF THAT'S

174
00:06:33,459 --> 00:06:36,129
IMAGINARY, WHAT'S REAL ABOUT

175
00:06:36,195 --> 00:06:38,297
A.I., IT'S COMPLICATED AND

176
00:06:38,364 --> 00:06:38,631
BEAUTIFUL.

177
00:06:38,698 --> 00:06:40,399
AND MATH IS GREAT BUT IT IS NOT

178
00:06:40,466 --> 00:06:42,034
GOING TO RISE UP AND TAKE OVER.

179
00:06:42,101 --> 00:06:45,738
THERE'S A LOT OF ANXIETY ABOUT

180
00:06:45,805 --> 00:06:47,607
A.I. TAKEOVERS AND ROBOT

181
00:06:47,673 --> 00:06:49,342
APOCALYPSES OUT THERE AND I JUST

182
00:06:49,408 --> 00:06:51,511
LIKE TO GET THAT OUT OF THE WAY

183
00:06:51,577 --> 00:06:54,380
BECAUSE IT ALLOWS US TO HAVE

184
00:06:54,447 --> 00:06:54,914
DEEPER CONVERSATIONS ABOUT

185
00:06:54,981 --> 00:06:56,382
ARTIFICIAL INTELLIGENCE WHICH I

186
00:06:56,449 --> 00:06:58,551
THINK IS REALLY IMPORTANT.

187
00:06:58,618 --> 00:06:59,318
AND SPECIFICALLY, THE

188
00:06:59,385 --> 00:07:02,488
CONVERSATION THAT I'M REALLY

189
00:07:02,555 --> 00:07:03,790
INTERESTED IN HAVING IS ABOUT

190
00:07:03,856 --> 00:07:06,759
BIAS IN A.I. AND THE WAY A.I.

191
00:07:06,826 --> 00:07:07,226
DISCRIMINATES.

192
00:07:07,293 --> 00:07:11,164
SO, I REALLY LIKE AN IDEA THAT

193
00:07:11,230 --> 00:07:12,265
COMES FROM BENJAMIN'S

194
00:07:12,331 --> 00:07:16,302
TECHNOLOGY, THE IDEA THAT A.I.

195
00:07:16,369 --> 00:07:18,171
SYSTEMS, AUTOMATED SYSTEMS,

196
00:07:18,237 --> 00:07:19,172
DISCRIMINATE BY DEFAULT.

197
00:07:19,238 --> 00:07:21,340
FOR A LONG TIME THERE'S BEEN

198
00:07:21,407 --> 00:07:27,246
THIS IDEA THAT I CALL TECHNO

199
00:07:27,313 --> 00:07:27,647
CHAUVINISM.

200
00:07:27,713 --> 00:07:29,081
WHAT I WOULD ARGUE IS WE SHOULD

201
00:07:29,148 --> 00:07:30,516
THINK ABOUT USING THE RIGHT TOOL

202
00:07:30,583 --> 00:07:31,384
FOR THE TASK.

203
00:07:31,450 --> 00:07:32,819
BECAUSE SOMETIMES THE RIGHT TOOL

204
00:07:32,885 --> 00:07:35,121
FOR THE TASK IS UNDOUBTEDLY A

205
00:07:35,188 --> 00:07:35,988
COMPUTER, SOMETIMES IT'S

206
00:07:36,055 --> 00:07:37,256
SOMETHING SIMPLE LIKE A BOOK IN

207
00:07:37,323 --> 00:07:41,027
THE HANDS OF A CHILD SITTING ON

208
00:07:41,093 --> 00:07:44,530
A PARENT'S LAP, ONE IS NOT

209
00:07:44,597 --> 00:07:46,065
INHERENTLY BETTER.

210
00:07:46,132 --> 00:07:47,800
BUT TECHNO CHAUVINIST

211
00:07:47,867 --> 00:07:51,070
PERSPECTIVE TELLS US A.I. OR

212
00:07:51,137 --> 00:07:52,939
COMPUTATIONAL SOLUTIONS ARE

213
00:07:53,005 --> 00:07:54,740
OBJECTIVE OR UNBIASED OR

214
00:07:54,807 --> 00:07:55,074
SUPERIOR.

215
00:07:55,141 --> 00:07:57,910
BUT WE CAN PUSH BACK AND LET IT

216
00:07:57,977 --> 00:08:00,112
GO AND THEN CERTAIN THINGS

217
00:08:00,179 --> 00:08:02,081
BECOME CLEAR, IT BECOMES CLEAR

218
00:08:02,148 --> 00:08:03,649
THAT THE PROBLEMS OF THE PAST

219
00:08:03,716 --> 00:08:05,017
ARE REFLECTED IN THE DATA THAT

220
00:08:05,084 --> 00:08:07,053
WE USE TO TRAIN A.I. SYSTEMS.

221
00:08:07,119 --> 00:08:08,988
SO WHEN WE MAKE A MACHINE

222
00:08:09,055 --> 00:08:10,389
LEARNING SYSTEM, WHAT WE DO IS

223
00:08:10,456 --> 00:08:13,159
TAKE A WHOLE BUNCH OF DATA,

224
00:08:13,226 --> 00:08:15,228
USUALLY FROM THE INTERNET, DUMP

225
00:08:15,294 --> 00:08:16,596
IT IN THE COMPUTER AND COMPUTER

226
00:08:16,662 --> 00:08:19,065
MAKES A MODEL.

227
00:08:19,131 --> 00:08:20,366
MODEL SHOWS MATHEMATICAL PATTERN

228
00:08:20,433 --> 00:08:21,767
OF DATA AND WHAT WE CAN DO IS

229
00:08:21,834 --> 00:08:22,902
ALL KINDS OF COOL STUFF.

230
00:08:22,969 --> 00:08:26,038
WE CAN USE THAT MODEL TO MAKE

231
00:08:26,105 --> 00:08:27,240
PREDICTIONS OR DECISIONS OR TO

232
00:08:27,306 --> 00:08:29,876
GENERATE NEW TEXT OR IMAGES OR

233
00:08:29,942 --> 00:08:33,112
AUDIO OR VIDEO OR WHATEVER.

234
00:08:33,179 --> 00:08:36,015
BUT THE PROBLEMS OF THE PAST ARE

235
00:08:36,082 --> 00:08:39,719
THINGS LIKE DISCRIMINATION.

236
00:08:39,785 --> 00:08:43,756
RACISM, SEXISM, ABLEISM,

237
00:08:43,823 --> 00:08:44,624
STRUCTURAL INEQUALITY, WHICH

238
00:08:44,690 --> 00:08:45,524
UNFORTUNATELY ALL OCCUR IN THE

239
00:08:45,591 --> 00:08:45,758
WORLD.

240
00:08:45,825 --> 00:08:48,027
WE DO NOT LIVE IN A PERFECT

241
00:08:48,094 --> 00:08:48,461
WORLD.

242
00:08:48,527 --> 00:08:50,229
ALL OF THESE PATTERNS OF

243
00:08:50,296 --> 00:08:50,997
DISCRIMINATION ARE REFLECTED IN

244
00:08:51,063 --> 00:08:53,499
THE DATA WE USE TO TRAIN A.I.

245
00:08:53,566 --> 00:08:55,801
SYSTEMS.

246
00:08:55,868 --> 00:08:56,836
HOW DOES THIS MANIFEST?

247
00:08:56,903 --> 00:09:00,740
WELL, ONE COMES FROM THE MARKUP,

248
00:09:00,806 --> 00:09:02,608
WHICH IS TERRIFIC.

249
00:09:02,675 --> 00:09:03,709
ALGORITHM ACCOUNTABILITY NEWS

250
00:09:03,776 --> 00:09:04,477
ORGANIZATION, HIGHLY RECOMMEND

251
00:09:04,543 --> 00:09:06,345
READING THEM IF YOU'RE NOT

252
00:09:06,412 --> 00:09:09,749
READING THEM ALREADY.

253
00:09:09,815 --> 00:09:12,685
THE MARKUP DID AN INVESTIGATION

254
00:09:12,752 --> 00:09:16,756
ABOUT SECRET BIAS HIDDEN IN

255
00:09:16,822 --> 00:09:17,423
MORTGAGE-APPROVAL ALGORITHMS, 40

256
00:09:17,490 --> 00:09:21,460
TO 80% MORE LIKELY TO DENY

257
00:09:21,527 --> 00:09:24,931
BORROWERS OF COLOR THAN WHITE

258
00:09:24,997 --> 00:09:26,499
COUNTERPARTS, IN SOME METRO

259
00:09:26,565 --> 00:09:27,900
AREAS MORE THAN 250%.

260
00:09:27,967 --> 00:09:29,335
NOW A DATA SCIENTIST MIGHT LOOK

261
00:09:29,402 --> 00:09:31,938
AT THIS AND SAY, WELL, IT'S JUST

262
00:09:32,004 --> 00:09:33,506
WHAT'S IN THE DATA.

263
00:09:33,572 --> 00:09:35,808
YOU KNOW, WHAT'S PROBLEMATIC?

264
00:09:35,875 --> 00:09:38,210
A SOCIOLOGIST MIGHT SAY, ALL

265
00:09:38,277 --> 00:09:40,613
RIGHT, WELL, CLEARLY WHAT'S

266
00:09:40,680 --> 00:09:42,248
HAPPENING IS THE MORTGAGE

267
00:09:42,315 --> 00:09:43,249
APPROVAL ALGORITHMS ARE FED WITH

268
00:09:43,316 --> 00:09:45,217
DATA ABOUT WHO GOT MORTGAGES IN

269
00:09:45,284 --> 00:09:48,187
THE PAST AND THE U.S., VERY LONG

270
00:09:48,254 --> 00:09:49,722
HISTORY OF RESIDENTIAL

271
00:09:49,789 --> 00:09:51,457
SEGREGATION, HAS A LONG HISTORY

272
00:09:51,524 --> 00:09:53,759
OF FINANCIAL DISCRIMINATION IN

273
00:09:53,826 --> 00:09:54,460
LENDING.

274
00:09:54,527 --> 00:09:55,861
AND SO THE ALGORITHMS ARE

275
00:09:55,928 --> 00:09:59,031
PICKING UP ON THOSE PATTERNS OF

276
00:09:59,098 --> 00:09:59,665
DISCRIMINATION AND REPRODUCING

277
00:09:59,732 --> 00:09:59,999
THEM.

278
00:10:00,066 --> 00:10:04,270
SO IT'S A GOOD EXAMPLE OF HOW

279
00:10:04,337 --> 00:10:05,371
DISCRIMINATION HAPPENS BY

280
00:10:05,438 --> 00:10:08,107
DEFAULT INSIDE THE AUTOMATED

281
00:10:08,174 --> 00:10:08,341
SYSTEMS.

282
00:10:08,407 --> 00:10:09,542
WE COULD MATHEMATICALLY PUT A

283
00:10:09,608 --> 00:10:11,610
FINGER ON THE SCALE.

284
00:10:11,677 --> 00:10:14,313
WE COULD EVALUATE OUR ALGORITHMS

285
00:10:14,380 --> 00:10:17,283
AND SAY, OKAY, HOW MANY LOANS

286
00:10:17,350 --> 00:10:21,887
ARE GOING TO PEOPLE OF DIFFERENT

287
00:10:21,954 --> 00:10:22,989
CATEGORIES/PROTECTED CLASSES AND

288
00:10:23,055 --> 00:10:25,091
IF IT'S NOT A SUFFICIENT

289
00:10:25,157 --> 00:10:26,392
PERCENTAGE WE CAN CHANGE THE

290
00:10:26,459 --> 00:10:28,027
ALGORITHM SO THAT IT GIVES MORE

291
00:10:28,094 --> 00:10:29,996
LOANS TO PEOPLE FROM THOSE

292
00:10:30,062 --> 00:10:31,998
PROTECTED CATEGORIES.

293
00:10:32,064 --> 00:10:37,269
IS THAT ACTUALLY HAPPENING?

294
00:10:37,336 --> 00:10:38,771
NOT QUITE SO MUCH.

295
00:10:38,838 --> 00:10:40,840
ANOTHER EXAMPLE IS IN FACIAL

296
00:10:40,906 --> 00:10:41,340
RECOGNITION.

297
00:10:41,407 --> 00:10:43,809
YOU'RE PROBABLY FAMILIAR WITH

298
00:10:43,876 --> 00:10:46,012
THE GENDER SHADES PROJECT, WHERE

299
00:10:46,078 --> 00:10:48,214
THE INVESTIGATORS LOOKED AT

300
00:10:48,280 --> 00:10:51,717
FACIAL RECOGNITION SYSTEMS AND

301
00:10:51,784 --> 00:10:53,986
FOUND THAT THE MAJOR FACIAL

302
00:10:54,053 --> 00:10:55,421
RECOGNITION SYSTEM WORKED BETTER

303
00:10:55,488 --> 00:10:56,989
ON MEN THAN WOMEN, BETTER ON

304
00:10:57,056 --> 00:11:03,329
PEOPLE WITH LIGHT SKIN THAN DARK

305
00:11:03,396 --> 00:11:04,530
SKIN, TRANSAND NON-BINARY FOLKS

306
00:11:04,597 --> 00:11:05,398
ARE GENERALLY NOT REPRESENTED IN

307
00:11:05,464 --> 00:11:06,932
THE DATASETS AT ALL.

308
00:11:06,999 --> 00:11:08,501
WHEN DO YOU INTERSECTIONAL

309
00:11:08,567 --> 00:11:09,435
ANALYSIS, THE SYSTEMS WORK BEST

310
00:11:09,502 --> 00:11:11,637
OF ALL ON MEN WITH LIGHT SKIN

311
00:11:11,704 --> 00:11:14,740
AND WORST OF ALL ON WOMEN WITH

312
00:11:14,807 --> 00:11:20,413
DARK SKIN.

313
00:11:20,479 --> 00:11:21,847
AND SO, THIS IS ABSOLUTELY A

314
00:11:21,914 --> 00:11:25,017
PROBLEM THAT COULD BE ADDRESSED

315
00:11:25,084 --> 00:11:27,119
WITH THE TRAINING DATA, DATA

316
00:11:27,186 --> 00:11:28,621
SCIENTIST FIRST QUESTION IS

317
00:11:28,687 --> 00:11:29,855
USUALLY CAN'T YOU JUST MAKE THE

318
00:11:29,922 --> 00:11:31,390
TRAINING DATA MORE DIVERSE?

319
00:11:31,457 --> 00:11:32,825
ABSOLUTELY, YES, THAT WOULD MAKE

320
00:11:32,892 --> 00:11:33,893
THESE SYSTEMS MORE ACCURATE.

321
00:11:33,959 --> 00:11:36,862
BUT ONE THING WE DO IS PUSH

322
00:11:36,929 --> 00:11:39,365
FURTHER AND LOOK AT THE USES OF

323
00:11:39,432 --> 00:11:40,266
FACIAL RECOGNITION.

324
00:11:40,332 --> 00:11:43,769
WE CAN LOOK AT A.I. IN CONTEXT,

325
00:11:43,836 --> 00:11:46,472
AND SAY, WELL, WE NEED TO THINK

326
00:11:46,539 --> 00:11:48,774
ABOUT THINGS LIKE FACIAL

327
00:11:48,841 --> 00:11:50,576
RECOGNITION USE IN POLICING

328
00:11:50,643 --> 00:11:51,844
BECAUSE FACIAL RECOGNITION USED

329
00:11:51,911 --> 00:11:53,813
TO UNLOCK YOUR PHONE IS A LOW

330
00:11:53,879 --> 00:11:54,380
RISK USE.

331
00:11:54,447 --> 00:11:55,681
DOESN'T WORK FOR ME MOST OF THE

332
00:11:55,748 --> 00:12:01,787
TIME BUT I DON'T CARE, I JUST

333
00:12:01,854 --> 00:12:03,222
USE MY PASSWORD.

334
00:12:03,289 --> 00:12:05,958
FACIAL RECOGNITION USED BY

335
00:12:06,025 --> 00:12:08,160
POLICE ON REALTIME VIDEO

336
00:12:08,227 --> 00:12:09,929
SURVEILLANCE FEEDS IS GOING TO

337
00:12:09,995 --> 00:12:12,398
BE A HIGH RISK USE BECAUSE IT'S

338
00:12:12,465 --> 00:12:14,400
GOING TO MISIDENTIFY WOMEN AND

339
00:12:14,467 --> 00:12:17,470
PEOPLE OF COLOR MORE OFTEN AND

340
00:12:17,536 --> 00:12:20,873
GET PEOPLE CAUGHT UP IN CRIMINAL

341
00:12:20,940 --> 00:12:21,807
JUSTICE SYSTEM WHO DON'T NEED TO

342
00:12:21,874 --> 00:12:25,444
BE CAUGHT UP IN THE CRIMINAL

343
00:12:25,511 --> 00:12:25,811
JUSTICE SYSTEM.

344
00:12:25,878 --> 00:12:27,446
IT'S A HIGH-RISK USE THAT MIGHT

345
00:12:27,513 --> 00:12:30,049
NEED TO BE REGULATED.

346
00:12:30,116 --> 00:12:33,752
SO WE NEED TO -- OH, SURE.

347
00:12:40,793 --> 00:12:41,060
BETTER?

348
00:12:41,127 --> 00:12:41,260
OKAY.

349
00:12:41,327 --> 00:12:42,161
SORRY, ONLINE FOLKS.

350
00:12:42,228 --> 00:12:46,398
ALL RIGHT, SO LET'S NOT USE

351
00:12:46,465 --> 00:12:52,771
FACIAL RECOGNITION IN POLICING.

352
00:12:52,838 --> 00:12:54,106
SO, LET'S TALK ABOUT CANCER.

353
00:12:54,173 --> 00:12:55,641
I HAVE AN A.I. STORY I WANT TO

354
00:12:55,708 --> 00:12:56,609
GET OFF MY CHEST.

355
00:12:56,675 --> 00:12:59,178
IT IS ABOUT THE TIME THAT I GOT

356
00:12:59,245 --> 00:12:59,879
BREAST CANCER.

357
00:12:59,945 --> 00:13:03,382
IT WAS BEGINNING OF THE

358
00:13:03,449 --> 00:13:03,749
PANDEMIC.

359
00:13:03,816 --> 00:13:05,918
AND I GOT THE NEWS THAT YOU

360
00:13:05,985 --> 00:13:09,388
NEVER WANT TO GET.

361
00:13:09,455 --> 00:13:11,924
AND I HAD BREAST CANCER.

362
00:13:11,991 --> 00:13:14,393
IT WAS REALLY HARD, I'M FINE

363
00:13:14,460 --> 00:13:16,495
NOW, GRATEFUL TO THE DOCTORS AND

364
00:13:16,562 --> 00:13:20,199
MEDICAL PROFESSIONALS WHO TOOK

365
00:13:20,266 --> 00:13:21,433
CARE OF ME.

366
00:13:21,500 --> 00:13:22,835
AND I LEARNED A WHOLE BUNCH OF

367
00:13:22,902 --> 00:13:25,738
THINGS ALONG THE WAY THAT I

368
00:13:25,804 --> 00:13:27,640
THINK ARE USEFUL TO TALK ABOUT

369
00:13:27,706 --> 00:13:29,742
WHEN WE TALK ABOUT WHAT THE

370
00:13:29,808 --> 00:13:30,943
GENERAL PUBLIC UNDERSTANDS ABOUT

371
00:13:31,010 --> 00:13:33,012
A.I. AND HOW WE CAN TALK ABOUT

372
00:13:33,078 --> 00:13:35,214
A.I. IN MEDICINE WITH THE

373
00:13:35,281 --> 00:13:35,748
GENERAL PUBLIC.

374
00:13:35,814 --> 00:13:39,585
SO ONE OF THE THINGS THAT I DID

375
00:13:39,652 --> 00:13:43,856
WHEN I FIRST GOT DIAGNOSED IS I

376
00:13:43,923 --> 00:13:44,723
FREAKED OUT, BECAUSE EVERYBODY

377
00:13:44,790 --> 00:13:46,358
FREAKS OUT WHEN THEY FIRST GET

378
00:13:46,425 --> 00:13:47,526
DIAGNOSED, AND THE WAY THAT YOU

379
00:13:47,593 --> 00:13:49,662
FREAK OUT TENDS TO BE CONSISTENT

380
00:13:49,728 --> 00:13:52,164
WITH YOUR PERSONALITY.

381
00:13:52,231 --> 00:13:56,335
SO, MY FREAKOUT INVOLVED

382
00:13:56,402 --> 00:13:59,271
OBSESSIVIVELY READING EVERYTHING

383
00:13:59,338 --> 00:14:00,406
I POSSIBLY COULD.

384
00:14:00,472 --> 00:14:02,675
I SPENT SO MUCH TIME ON

385
00:14:02,741 --> 00:14:04,210
PubMed AND READ EVERY SINGLE

386
00:14:04,276 --> 00:14:05,644
THING IN EVERY SINGLE DROPDOWN

387
00:14:05,711 --> 00:14:07,846
AND EVERY SINGLE PART OF MY

388
00:14:07,913 --> 00:14:11,250
ELECTRONIC MEDICAL RECORD, WHICH

389
00:14:11,317 --> 00:14:15,054
IS WHERE I STUMBLED ACROSS A

390
00:14:15,120 --> 00:14:18,157
SCAN THIS SCAN WAS READ BY DR.

391
00:14:18,224 --> 00:14:19,491
SO-AND-SO AND DR. A.I.

392
00:14:19,558 --> 00:14:21,327
WHY IS THIS A.I. READING MY

393
00:14:21,393 --> 00:14:21,527
SCAN?

394
00:14:21,594 --> 00:14:22,394
WHAT DID IT FIND?

395
00:14:22,461 --> 00:14:23,929
WHO WROTE THIS A.I.?

396
00:14:23,996 --> 00:14:26,265
WHO KIND OF BIAS IS IN IT?

397
00:14:26,332 --> 00:14:27,866
AND I HAD A LITTLE MOMENT AND

398
00:14:27,933 --> 00:14:30,169
THEN I FORGOT ABOUT IT BECAUSE,

399
00:14:30,236 --> 00:14:31,804
YOU KNOW, CANCER.

400
00:14:31,870 --> 00:14:33,772
AND THEN I CAME BACK TO IT A

401
00:14:33,839 --> 00:14:34,540
LITTLE WHILE LATER.

402
00:14:34,607 --> 00:14:36,175
A YEAR OR SO LATER.

403
00:14:36,242 --> 00:14:38,244
I WAS FEELING BETTER.

404
00:14:38,310 --> 00:14:42,481
IT WAS COVID.

405
00:14:42,548 --> 00:14:45,484
AND I DECIDED TO GET A LITTLE

406
00:14:45,551 --> 00:14:47,019
BIT CARRIED AWAY.

407
00:14:47,086 --> 00:14:48,654
I DECIDED THAT I WAS GOING TO DO

408
00:14:48,721 --> 00:14:52,157
AN EXPERIMENT.

409
00:14:52,224 --> 00:14:56,295
I WAS GOING TO DOWNLOAD MY OWN

410
00:14:56,362 --> 00:14:58,797
SCANS AND RUN THEM THROUGH AN

411
00:14:58,864 --> 00:15:01,667
OPEN SOURCE A.I. IN ORDER TO

412
00:15:01,734 --> 00:15:02,768
WRITE ABOUT THE STATE-OF-THE-ART

413
00:15:02,835 --> 00:15:07,740
AND A.I.-BASED CANCER DETECTION.

414
00:15:07,806 --> 00:15:09,842
AND I WILL ADMIT THAT I GOT A

415
00:15:09,908 --> 00:15:10,843
LITTLE CARRIED AWAY, THIS WAS

416
00:15:10,909 --> 00:15:16,415
MAYBE A LITTLE EXTREME.

417
00:15:16,482 --> 00:15:17,716
BUT I WAS TRYING TO REPRODUCE

418
00:15:17,783 --> 00:15:18,517
THESE THE EXPERIMENT.

419
00:15:18,584 --> 00:15:22,421
I HAVE ALWAYS BEEN REALLY

420
00:15:22,488 --> 00:15:24,189
INTERESTED IN REPRODUCIBILITY.

421
00:15:24,256 --> 00:15:27,493
AND OF COURSE THE IDEA IN

422
00:15:27,559 --> 00:15:28,027
REPRODUCIBILITY YOU WOULD

423
00:15:28,093 --> 00:15:29,895
PUBLISH YOUR CODE AND DATA

424
00:15:29,962 --> 00:15:30,763
ONLINE, OTHER PEOPLE CAN

425
00:15:30,829 --> 00:15:31,630
REPRODUCE YOUR EXPERIMENT AND

426
00:15:31,697 --> 00:15:32,131
VALIDATE IT.

427
00:15:32,197 --> 00:15:34,266
WE DO THIS A LOT IN JOURNALISM

428
00:15:34,333 --> 00:15:36,468
WHEN WE PUBLISH THE CODE AND

429
00:15:36,535 --> 00:15:41,040
DATA WE USE TO DO A PARTICULAR

430
00:15:41,106 --> 00:15:41,373
ANALYSIS.

431
00:15:41,440 --> 00:15:42,508
ESPECIALLY ALGORITHMIC

432
00:15:42,574 --> 00:15:44,743
ACCOUNTABILITY REPORTING.

433
00:15:44,810 --> 00:15:46,612
AND SO, I TOOK MY SCANS, AND I

434
00:15:46,679 --> 00:15:49,114
RAN THEM THROUGH AN A.I.

435
00:15:49,181 --> 00:15:53,819
I HAD A LOT OF MISCONCEPTIONS,

436
00:15:53,886 --> 00:15:56,055
IT TURNED OUT.

437
00:15:56,121 --> 00:15:59,525
AND MY MISCONCEPTION ABOUT HOW

438
00:15:59,591 --> 00:16:03,562
A.I. WOULD BE USED IN CANCER

439
00:16:03,629 --> 00:16:07,232
DETECTION ARE NOT UNCOMMON.

440
00:16:07,299 --> 00:16:11,270
SO, HERE IS A PICTURE OF, YOU

441
00:16:11,337 --> 00:16:12,237
KNOW, SOMEBODY'S MAMMOGRAM.

442
00:16:12,304 --> 00:16:13,672
SOME MAMMOGRAM THAT I FOUND ON

443
00:16:13,739 --> 00:16:17,943
THE INTERNET, NOT MY ACTUAL

444
00:16:18,010 --> 00:16:18,444
MAMMOGRAM.

445
00:16:18,510 --> 00:16:21,113
AND I THOUGHT THAT I WAS GOING

446
00:16:21,180 --> 00:16:23,082
TO TAKE MY ENTIRE ELECTRONIC

447
00:16:23,148 --> 00:16:24,516
MEDICAL RECORD, FEED IT IN, AND

448
00:16:24,583 --> 00:16:27,286
I WAS GOING TO PUT IN ALL OF MY

449
00:16:27,353 --> 00:16:31,323
SCANS AND, YOU KNOW, THE 3D

450
00:16:31,390 --> 00:16:34,059
VIDEO, BLAH, BLAH, BLAH, IT WAS

451
00:16:34,126 --> 00:16:35,594
GOING TO EVALUATE EVERYTHING AND

452
00:16:35,661 --> 00:16:37,229
THEN LIKE POP OUT AN ANSWER.

453
00:16:37,296 --> 00:16:40,833
LIKE WE THINK YOU HAVE CANCER.

454
00:16:40,899 --> 00:16:42,267
BUT, NO.

455
00:16:42,334 --> 00:16:43,469
THIS IS TOTALLY WRONG.

456
00:16:43,535 --> 00:16:46,205
I ALSO THOUGHT THAT I WAS GOING

457
00:16:46,271 --> 00:16:47,940
TO DO AN EXPERIMENT WHERE I

458
00:16:48,006 --> 00:16:52,077
WOULD CHANGE MY RACE AND SEE IF

459
00:16:52,144 --> 00:16:53,078
THE PREDICTION CHANGED.

460
00:16:53,145 --> 00:16:54,413
AND THIS WAS A GOOD PREMISE FOR

461
00:16:54,480 --> 00:16:56,682
AN EXPERIMENT BUT ALSO TOTALLY

462
00:16:56,749 --> 00:16:58,450
WRONG BECAUSE THERE'S NO RACE IN

463
00:16:58,517 --> 00:17:01,520
THE A.I. THAT I USED.

464
00:17:01,587 --> 00:17:02,888
SO, ONE OF MY MISCONCEPTIONS WAS

465
00:17:02,955 --> 00:17:04,456
IT WAS GOING TO LOOK AT MY

466
00:17:04,523 --> 00:17:08,827
ENTIRE RECORD.

467
00:17:08,894 --> 00:17:10,362
NO, THE WAY THAT MOST CANCER

468
00:17:10,429 --> 00:17:12,564
DETECTION A.I. THAT WE HAVE

469
00:17:12,631 --> 00:17:16,635
RIGHT NOW WORKS IS THAT YOU PUT

470
00:17:16,702 --> 00:17:19,071
IN A FLAT IMAGE, AND WHAT IT

471
00:17:19,138 --> 00:17:22,808
WILL DO IS IT WILL IDENTIFY AN

472
00:17:22,875 --> 00:17:27,679
AREA OF CONCERN ON THE IMAGE.

473
00:17:27,746 --> 00:17:31,316
SO, I WAS A LITTLE SURPRISED TO

474
00:17:31,383 --> 00:17:32,985
HEAR THIS.

475
00:17:33,051 --> 00:17:34,586
BUT I DOWNLOADED THE STUFF, I

476
00:17:34,653 --> 00:17:37,423
GOT ALL THE CODE WORKING, I FED

477
00:17:37,489 --> 00:17:40,592
IN MY IMAGES AND I GOT NOTHING.

478
00:17:40,659 --> 00:17:42,694
IT WAS LIKE, YOU KNOW, I KNOW

479
00:17:42,761 --> 00:17:43,762
THERE'S CANCER HERE.

480
00:17:43,829 --> 00:17:47,065
LIKE MY DOCTOR HAS SEEN IT.

481
00:17:47,132 --> 00:17:49,701
LIKE I GOT A MASTECTOMY CAUSE

482
00:17:49,768 --> 00:17:50,636
THERE WAS CANCER HERE, WHY CAN'T

483
00:17:50,702 --> 00:17:53,505
THE A.I. SEE IT?

484
00:17:53,572 --> 00:17:56,675
IT TURNED OUT THAT MY IMAGE

485
00:17:56,742 --> 00:17:57,476
RESOLUTION WAS TOTALLY WRONG,

486
00:17:57,543 --> 00:18:01,180
AND I HAD TO GO THROUGH A

487
00:18:01,246 --> 00:18:03,849
RIGMAROLE OF TRYING TO GET

488
00:18:03,916 --> 00:18:05,784
ACCESS TO HIGH RESOLUTION

489
00:18:05,851 --> 00:18:07,653
VERSION OF MY SCANS.

490
00:18:07,719 --> 00:18:09,822
AFTER A MONTH OR TWO OF

491
00:18:09,888 --> 00:18:12,357
WRANGLING I FINALLY GAVE UP AND

492
00:18:12,424 --> 00:18:14,193
ASKED THEM TO SEND ME A CD IN

493
00:18:14,259 --> 00:18:22,968
THE MAIL AND I HAD TO BUY A CD

494
00:18:23,035 --> 00:18:23,202
READER.

495
00:18:23,268 --> 00:18:25,804
THIS IS 2007, I DON'T KNOW.

496
00:18:25,871 --> 00:18:28,640
SO, I GOT IT GOING.

497
00:18:28,707 --> 00:18:31,310
AND EVENTUALLY IT DID IDENTIFY

498
00:18:31,376 --> 00:18:33,412
AN AREA OF CONCERN.

499
00:18:33,479 --> 00:18:34,713
THIS IS WHAT AN AREA OF CONCERN

500
00:18:34,780 --> 00:18:35,981
MIGHT LOOK LIKE.

501
00:18:36,048 --> 00:18:39,818
AGAIN, SOME IMAGE I GOT OFF THE

502
00:18:39,885 --> 00:18:40,085
INTERNET.

503
00:18:40,152 --> 00:18:43,422
AND WHAT I FOUND WAS THAT I GOT

504
00:18:43,489 --> 00:18:47,159
A SCORE OF -- IT GAVE ME -- IT

505
00:18:47,226 --> 00:18:48,460
IDENTIFIED AN AREA OF CONCERN

506
00:18:48,527 --> 00:18:53,165
AND GAVE ME A SCORE.

507
00:18:53,232 --> 00:18:56,134
AND I WAS REALLY SURPRISED THAT

508
00:18:56,201 --> 00:19:00,072
IT DIDN'T LIKE KIND OF GIVE ME A

509
00:19:00,138 --> 00:19:02,341
BIG -- LIKE A BIG RIGMAROLE.

510
00:19:02,407 --> 00:19:03,509
WHEN YOU TEXT SOMEBODY

511
00:19:03,575 --> 00:19:04,176
CONGRATULATIONS IT GIVES YOU

512
00:19:04,243 --> 00:19:05,110
BALLOONS AND STUFF.

513
00:19:05,177 --> 00:19:08,947
I WAS SORT OF EXPECTING THAT BUT

514
00:19:09,014 --> 00:19:10,148
LIKE YOU HAVE CANCER!

515
00:19:10,215 --> 00:19:11,316
YOU DON'T HAVE CANCER!

516
00:19:11,383 --> 00:19:14,620
THAT'S NOW HOW IT WORKS.

517
00:19:14,686 --> 00:19:16,488
SO, THAT WAS ONE OF THE THINGS

518
00:19:16,555 --> 00:19:20,959
THAT I LEARNED ALONG THE WAY,

519
00:19:21,026 --> 00:19:23,729
THAT EVEN WHEN YOU ARE A

520
00:19:23,795 --> 00:19:26,164
RESEARCHER WHO WORKS ON PEOPLE'S

521
00:19:26,231 --> 00:19:27,666
UNREALISTIC EXPECTATIONS AROUND

522
00:19:27,733 --> 00:19:29,835
A.I., YOU ACTUALLY YOURSELF HAVE

523
00:19:29,902 --> 00:19:31,169
UNREALISTIC EXPECTATIONS AROUND

524
00:19:31,236 --> 00:19:33,038
A.I., AND IT'S USEFUL TO KEEP

525
00:19:33,105 --> 00:19:34,139
THIS IN MIND WHEN WE'RE

526
00:19:34,206 --> 00:19:35,674
COMMUNICATING WITH THE PUBLIC

527
00:19:35,741 --> 00:19:39,378
ABOUT WHAT A.I. DOES.

528
00:19:39,444 --> 00:19:41,680
ONE OF THE THINGS THAT I WAS

529
00:19:41,747 --> 00:19:43,415
OBSESSED WITH WAS HOW LIKELY AM

530
00:19:43,482 --> 00:19:44,950
I TO DIE FROM THIS.

531
00:19:45,017 --> 00:19:45,183
RIGHT?

532
00:19:45,250 --> 00:19:46,919
BECAUSE THAT IS THE BIG QUESTION

533
00:19:46,985 --> 00:19:48,987
THAT YOU HAVE WHEN YOU ARE

534
00:19:49,054 --> 00:19:50,322
DIAGNOSED WITH CANCER.

535
00:19:50,389 --> 00:19:52,524
I DON'T KNOW WHAT CANCER

536
00:19:52,591 --> 00:19:53,292
RESEARCHERS' BIG QUESTIONS ARE.

537
00:19:53,358 --> 00:19:55,994
ALL I KNOW IS FROM THE

538
00:19:56,061 --> 00:19:56,662
PERSPECTIVE OF SOMEBODY

539
00:19:56,728 --> 00:19:56,929
DIAGNOSED.

540
00:19:56,995 --> 00:20:00,599
AND WHAT I KNEW WAS THAT BLACK

541
00:20:00,666 --> 00:20:06,171
WOMEN ARE 40% MORE LIKELY TO DIE

542
00:20:06,238 --> 00:20:07,072
FROM BREAST CANCER.

543
00:20:07,139 --> 00:20:11,843
BUT I AM -- MY MOM IS BLACK.

544
00:20:11,910 --> 00:20:14,713
MY DAD WAS BLACK, MY MOM WAS

545
00:20:14,780 --> 00:20:18,684
WHITE, AND MY MOM DIED OF BREAST

546
00:20:18,750 --> 00:20:23,255
CANCER WHEN SHE WAS 50.

547
00:20:23,322 --> 00:20:26,091
AND I WONDERED, OKAY, FOR THE

548
00:20:26,158 --> 00:20:27,526
PURPOSES OF MEDICAL STATISTICS,

549
00:20:27,593 --> 00:20:29,194
WOULD I BE CONSIDERED BLACK?

550
00:20:29,261 --> 00:20:30,596
WOULD I BE CONSIDERED WHITE?

551
00:20:30,662 --> 00:20:34,333
WHAT IS MY RISK PROFILE?

552
00:20:34,399 --> 00:20:38,036
AND I ASKED MY DOCTOR ABOUT

553
00:20:38,103 --> 00:20:38,604
THIS.

554
00:20:38,670 --> 00:20:40,772
MY DOCTOR EXPLAINED THAT, WELL,

555
00:20:40,839 --> 00:20:45,377
ACTUALLY WHAT WE DO WHEN THERE

556
00:20:45,444 --> 00:20:47,112
ARE DIFFERENCES IN PATTERNS, OR

557
00:20:47,179 --> 00:20:48,780
PREVALENCE BY RACE, IS FOR YOU

558
00:20:48,847 --> 00:20:50,415
WE RUN YOUR NUMBERS BOTH WAYS AS

559
00:20:50,482 --> 00:20:52,150
IF YOU WERE BLACK AND AS IF YOU

560
00:20:52,217 --> 00:20:54,586
WERE WHITE, AND THEN WE JUST

561
00:20:54,653 --> 00:20:55,787
ASSUME YOUR ACTUAL NUMBER IS

562
00:20:55,854 --> 00:20:56,855
SOMEWHERE IN BETWEEN.

563
00:20:56,922 --> 00:21:00,258
I WAS LIKE, THAT'S SCIENCE?

564
00:21:00,325 --> 00:21:02,561
SO, I COULDN'T FIGURE OUT, OKAY,

565
00:21:02,628 --> 00:21:05,297
AM I LIKELY TO DIE FROM THIS,

566
00:21:05,364 --> 00:21:08,367
NOT LIKELY TO DIE?

567
00:21:08,433 --> 00:21:10,669
IT MADE ME THINK, OKAY, THERE'S

568
00:21:10,736 --> 00:21:13,071
A LOT OF AMBIGUITY IN DIAGNOSIS,

569
00:21:13,138 --> 00:21:15,907
A LOT OF AMBIGUITY IN

570
00:21:15,974 --> 00:21:16,508
STATISTICS.

571
00:21:16,575 --> 00:21:17,876
AND WE'RE NOT NECESSARILY PAYING

572
00:21:17,943 --> 00:21:19,611
ENOUGH ATTENTION TO THAT

573
00:21:19,678 --> 00:21:23,815
AMBIGUITY WHEN WE ARE INCLUDING

574
00:21:23,882 --> 00:21:25,117
THESE SYSTEMS IN ALGORITHMS.

575
00:21:25,183 --> 00:21:25,317
RIGHT?

576
00:21:25,384 --> 00:21:28,186
SO WE NEED TO TALK MORE ABOUT

577
00:21:28,253 --> 00:21:28,920
THE AMBIGUITY.

578
00:21:28,987 --> 00:21:32,257
I ALSO THOUGHT IT WAS

579
00:21:32,324 --> 00:21:33,325
SIGNIFICANT THAT COMPUTERS DON'T

580
00:21:33,392 --> 00:21:35,327
DIAGNOSE THE SAME WAY THAT

581
00:21:35,394 --> 00:21:37,062
DOCTORS DO, RIGHT?

582
00:21:37,129 --> 00:21:38,397
THE COMPUTER IDENTIFIES AN AREA

583
00:21:38,463 --> 00:21:40,766
OF CONCERN AND GIVES YOU A

584
00:21:40,832 --> 00:21:40,966
SCORE.

585
00:21:41,033 --> 00:21:44,403
THIS PARTICULAR ONE GAVE ME A

586
00:21:44,469 --> 00:21:45,704
SCORE BETWEEN ZERO AND ONE.

587
00:21:45,771 --> 00:21:47,372
I THOUGHT THIS MEANT YOU HAVE A

588
00:21:47,439 --> 00:21:48,340
20% CHANCE OF HAVING CANCER IN

589
00:21:48,407 --> 00:21:49,675
THIS AREA.

590
00:21:49,741 --> 00:21:54,146
BUT NO, RESEARCHER TOLD ME THAT

591
00:21:54,212 --> 00:21:56,782
IT'S ABSOLUTELY EMPIRICALLY NOT

592
00:21:56,848 --> 00:21:57,115
A PREDICTION.

593
00:21:57,182 --> 00:21:58,316
IT IS JUST A SCORE.

594
00:21:58,383 --> 00:22:02,454
TURNS OUT THIS IS BECAUSE OF THE

595
00:22:02,521 --> 00:22:05,424
LEGAL LANDSCAPE, BECAUSE, YOU

596
00:22:05,490 --> 00:22:08,493
KNOW, THERE'S -- HEAVILY

597
00:22:08,560 --> 00:22:10,028
REGULATED INDUSTRY AND HE COULD

598
00:22:10,095 --> 00:22:13,632
GET IN BIG TROUBLE FOR SAYING

599
00:22:13,699 --> 00:22:14,966
THERE'S 20% CHANCE.

600
00:22:15,033 --> 00:22:17,469
OH, YEAH, THIS IS A .2 SCORE.

601
00:22:17,536 --> 00:22:19,004
ALL RIGHT, FINE.

602
00:22:19,071 --> 00:22:20,872
ANOTHER THING THAT PEOPLE

603
00:22:20,939 --> 00:22:22,507
USUALLY DON'T -- THE GENERAL

604
00:22:22,574 --> 00:22:23,775
PUBLIC GENERALLY DOES NOT

605
00:22:23,842 --> 00:22:25,110
REALIZE ABOUT A.I. THAT'S

606
00:22:25,177 --> 00:22:28,847
IMPORTANT TO SAY IN CONTEXT OF

607
00:22:28,914 --> 00:22:32,684
CANCER IS THAT THESE SYSTEMS ARE

608
00:22:32,751 --> 00:22:34,219
TUNED TO EITHER GIVE YOU A

609
00:22:34,286 --> 00:22:35,520
HIGHER RATE OF FALSE POSITIVES

610
00:22:35,587 --> 00:22:38,490
OR HIGHER RATE OF FALSE

611
00:22:38,557 --> 00:22:39,124
NEGATIVES.

612
00:22:39,191 --> 00:22:42,761
SO, A FALSE POSITIVE WOULD BE

613
00:22:42,828 --> 00:22:45,263
WHERE IT SAYS, YEP, HAVE CANCER,

614
00:22:45,330 --> 00:22:47,666
DON'T HAVE NEGATIVE.

615
00:22:47,733 --> 00:22:49,201
FALSE NEGATIVE, NOPE, NO CANCER,

616
00:22:49,267 --> 00:22:51,837
AND DO YOU HAVE MEDICINE.

617
00:22:51,903 --> 00:22:53,371
IN MEDICINE IT'S CONSIDERED

618
00:22:53,438 --> 00:22:55,974
HIGHER STAKES TO HAVE MORE FALSE

619
00:22:56,041 --> 00:22:58,410
NEGATIVES THAN FALSE POSITIVES.

620
00:22:58,477 --> 00:23:01,580
WE ARE ALL UNITED IN WANTING TO

621
00:23:01,646 --> 00:23:04,616
SAVE MORE LIVES, FOR MORE PEOPLE

622
00:23:04,683 --> 00:23:09,554
TO BE DIAGNOSED AND TREATED

623
00:23:09,621 --> 00:23:09,855
EFFECTIVELY.

624
00:23:09,921 --> 00:23:11,189
SO, THERE'S CONSENSUS,

625
00:23:11,256 --> 00:23:11,757
APPROPRIATELY, THESE SYSTEMS

626
00:23:11,823 --> 00:23:14,159
SHOULD BE TUNED TO HAVE HIGHER

627
00:23:14,226 --> 00:23:15,894
RATES OF FALSE POSITIVES THAN

628
00:23:15,961 --> 00:23:16,795
FALSE NEGATIVES.

629
00:23:16,862 --> 00:23:19,631
IT'S KIND OF THE STRANGE THING

630
00:23:19,698 --> 00:23:21,600
TO REALIZE YOU HAVE TO DECIDE

631
00:23:21,666 --> 00:23:23,902
HOW WRONG YOU WANT YOUR

632
00:23:23,969 --> 00:23:26,304
DIAGNOSTIC SYSTEM TO BE.

633
00:23:26,371 --> 00:23:29,641
AND PEOPLE ARE PRETTY

634
00:23:29,708 --> 00:23:33,912
UNCOMFORTABLE WITH THAT.

635
00:23:33,979 --> 00:23:39,384
I ALSO GOT REALLY INTERESTED IN

636
00:23:39,451 --> 00:23:40,719
RACIAL DISPARITIES.

637
00:23:40,786 --> 00:23:44,422
IN DIAGNOSIS AND THEN ALSO IN

638
00:23:44,489 --> 00:23:46,458
A.I.-READ SCANS.

639
00:23:46,525 --> 00:23:49,127
THERE IS THIS REALLY INTERESTING

640
00:23:49,194 --> 00:23:52,764
STUDY THAT LOOKED AT MACHINE

641
00:23:52,831 --> 00:23:58,436
LEARNING MODELS, DIAGNOSING I

642
00:23:58,503 --> 00:23:59,971
THINK IT WAS PLEURAL EFFUSIONS.

643
00:24:00,038 --> 00:24:02,474
YOU COULD FACT CHECK ME ON THIS

644
00:24:02,541 --> 00:24:02,641
ONE.

645
00:24:02,707 --> 00:24:04,142
AT ANY RATE, THEY WERE LOOKING

646
00:24:04,209 --> 00:24:08,180
AT THESE MODELS THAT WERE

647
00:24:08,246 --> 00:24:09,281
EXTREMELY EFFECTIVE ON DATA FROM

648
00:24:09,347 --> 00:24:11,116
ONE HOSPITAL AND THEN THEY ADDED

649
00:24:11,183 --> 00:24:12,350
ANOTHER HOSPITAL, AND IT WAS

650
00:24:12,417 --> 00:24:13,685
REALLY EFFECTIVE ON THE DATA

651
00:24:13,752 --> 00:24:16,087
FROM THE NEXT HOSPITAL.

652
00:24:16,154 --> 00:24:18,590
AND THEN THEY ADDED IN THE DATA

653
00:24:18,657 --> 00:24:20,559
ON THE PATIENT'S RACE.

654
00:24:20,625 --> 00:24:25,030
AND IT TURNED OUT THAT THESE

655
00:24:25,096 --> 00:24:25,697
MODELS --

656
00:24:25,764 --> 00:24:28,200
>> RECORDING IN PROGRESS.

657
00:24:28,266 --> 00:24:30,068
>> OH.

658
00:24:30,135 --> 00:24:32,170
THESE MODELS WERE DIFFERENTIALLY

659
00:24:32,237 --> 00:24:34,239
ACCURATE, BASED ON RACE.

660
00:24:34,306 --> 00:24:34,773
RIGHT?

661
00:24:34,840 --> 00:24:37,943
BUT IT'S JUST SCANS OF PEOPLE'S

662
00:24:38,009 --> 00:24:38,176
INSIDES.

663
00:24:38,243 --> 00:24:40,245
RESEARCHERS ARE LIKE WHY IS THIS

664
00:24:40,312 --> 00:24:40,846
HAPPENING?

665
00:24:40,912 --> 00:24:41,880
AND NOBODY'S BEEN ABLE TO

666
00:24:41,947 --> 00:24:45,183
EXPLAIN WHY THIS IS HAPPENING,

667
00:24:45,250 --> 00:24:45,483
RIGHT?

668
00:24:45,550 --> 00:24:46,484
REALLY BIG UNKNOWN, RIGHT?

669
00:24:46,551 --> 00:24:50,088
WE KNOW THINGS LIKE FACIAL

670
00:24:50,155 --> 00:24:52,724
RECOGNITION SYSTEMS ARE

671
00:24:52,791 --> 00:24:54,259
DIFFERENTIALLY ACCURATE BASED ON

672
00:24:54,326 --> 00:24:55,794
SKIN TONES, BUT IT DOESN'T MAKE

673
00:24:55,861 --> 00:24:58,330
ANY SENSE TO A HUMAN OR AT LEAST

674
00:24:58,396 --> 00:25:03,268
TO THIS HUMAN THAT THE SCAN OF

675
00:25:03,335 --> 00:25:06,137
THE INSIDE WOULD BE DIFFERENT.

676
00:25:06,204 --> 00:25:07,873
SO BIG QUESTION MARK.

677
00:25:07,939 --> 00:25:10,709
AND ONE OTHER THING I FIND

678
00:25:10,775 --> 00:25:15,747
REALLY INTERESTING IS THE WAY

679
00:25:15,814 --> 00:25:19,017
THAT RACE GETS ENCODED IN

680
00:25:19,084 --> 00:25:20,452
MEDICAL SYSTEMS.

681
00:25:20,518 --> 00:25:21,753
AND IN ALGORITHMS.

682
00:25:21,820 --> 00:25:24,089
AS IF IT WERE REAL.

683
00:25:24,155 --> 00:25:27,359
SO RACE IS SOCIAL.

684
00:25:27,425 --> 00:25:28,693
IT'S NOT BIOLOGICAL.

685
00:25:28,760 --> 00:25:30,528
THERE ARE GENETIC DIFFERENCES

686
00:25:30,595 --> 00:25:32,631
BETWEEN PEOPLE, RIGHT?

687
00:25:32,697 --> 00:25:35,033
BUT RACE DOES NOT DESCRIBE THOSE

688
00:25:35,100 --> 00:25:35,700
GENETIC DIFFERENCES.

689
00:25:35,767 --> 00:25:38,770
THE RACE IS A SOCIAL CONSTRUCT,

690
00:25:38,837 --> 00:25:41,039
NOT A BIOLOGICAL REALITY, BUT

691
00:25:41,106 --> 00:25:42,707
SOMETIMES GETS EMBEDDED IN

692
00:25:42,774 --> 00:25:45,310
ALGORITHMS AND IN MEDICAL

693
00:25:45,377 --> 00:25:47,812
SYSTEMS AS IF IT WERE A

694
00:25:47,879 --> 00:25:48,246
BIOLOGICAL REALITY.

695
00:25:48,313 --> 00:25:51,883
SO YOU'RE PROBABLY FAMILIAR WITH

696
00:25:51,950 --> 00:25:53,518
RACE CORRECTION.

697
00:25:53,585 --> 00:25:56,354
SO, IN KIDNEY DISEASE DIAGNOSIS,

698
00:25:56,421 --> 00:26:01,059
THERE'S A PRETTY SIGNIFICANT

699
00:26:01,126 --> 00:26:04,696
EXAMPLE OF EGFR, FILTRATION

700
00:26:04,763 --> 00:26:06,665
RATES, SO WHEN YOUR EGFR GETS

701
00:26:06,731 --> 00:26:09,467
DOWN TO 20, 20% KIDNEY FUNCTION,

702
00:26:09,534 --> 00:26:13,705
YOU'RE ELIGIBLE FOR THE KIDNEY

703
00:26:13,772 --> 00:26:14,406
TRANSPLANT LIST.

704
00:26:14,472 --> 00:26:16,408
IT DOESN'T MEAN YOU'LL GET IT,

705
00:26:16,474 --> 00:26:17,409
JUST ELIGIBLE TO START WAITING

706
00:26:17,475 --> 00:26:19,477
FOR A KIDNEY THAT MATCHES.

707
00:26:19,544 --> 00:26:20,478
MIGHT TAKE YEARS.

708
00:26:20,545 --> 00:26:25,417
FOR A VERY LONG TIME THERE WAS A

709
00:26:25,483 --> 00:26:26,818
RACE CORRECTION BUILT IN SO THAT

710
00:26:26,885 --> 00:26:30,588
IF YOU WERE BLACK, YOU GOT A

711
00:26:30,655 --> 00:26:34,225
MULTIPLIER APPLIED IN YOUR EGFR

712
00:26:34,292 --> 00:26:34,826
CALCULATION WHICH MEANT BLACK

713
00:26:34,893 --> 00:26:36,928
PATIENTS HAD TO BE SICKER THAN

714
00:26:36,995 --> 00:26:39,264
ANY OTHER PATIENTS IN ORDER TO

715
00:26:39,331 --> 00:26:40,498
QUALIFY FOR THE KIDNEY

716
00:26:40,565 --> 00:26:41,333
TRANSPLANT LIST.

717
00:26:41,399 --> 00:26:46,805
WE ALSO SEE RACE CORRECTIONS IN

718
00:26:46,871 --> 00:26:48,773
THINGS LIKE CONCUSSION.

719
00:26:48,840 --> 00:26:51,209
CONCUSSION CALCULATIONS.

720
00:26:51,276 --> 00:26:53,178
THERE WAS A CASE IN THE NFL

721
00:26:53,244 --> 00:26:55,847
WHERE THE NFL MADE THIS

722
00:26:55,914 --> 00:26:58,149
MULTI-BILLION DOLLAR SETTLEMENT

723
00:26:58,216 --> 00:27:00,785
TO PLAYERS WHO HAD EXPERIENCED

724
00:27:00,852 --> 00:27:02,420
MULTIPLE CONCUSSIONS AND HAD

725
00:27:02,487 --> 00:27:06,124
REALLY TRAUMATIC BRAIN INJURIES

726
00:27:06,191 --> 00:27:07,993
AND LONG-TERM EFFECTS, AND THEY

727
00:27:08,059 --> 00:27:09,194
USED A CALCULATION TO FIGURE OUT

728
00:27:09,260 --> 00:27:12,464
HOW MUCH EACH PLAYER WAS OWED,

729
00:27:12,530 --> 00:27:14,866
AND BLACK PLAYERS WERE GIVEN A

730
00:27:14,933 --> 00:27:18,370
RACE CORRECTION UNDER THE

731
00:27:18,436 --> 00:27:21,106
ASSUMPTION THAT THEIR MENTAL

732
00:27:21,172 --> 00:27:22,640
CAPACITY WAS DIMINISHED RELATIVE

733
00:27:22,707 --> 00:27:24,609
TO OTHER PLAYERS, AND THEY WERE

734
00:27:24,676 --> 00:27:25,477
OWED LESS MONEY.

735
00:27:25,543 --> 00:27:27,579
THEY FOUGHT IT.

736
00:27:27,645 --> 00:27:28,980
THEY WON.

737
00:27:29,047 --> 00:27:32,951
BUT WE DO SEE THINGS LIKE RACE

738
00:27:33,018 --> 00:27:37,655
CORRECTION LINGERING.

739
00:27:37,722 --> 00:27:39,758
THIS LINGERED IN KIDNEY

740
00:27:39,824 --> 00:27:45,563
ALGORITHM UNTIL VERY RECENTLY.

741
00:27:45,630 --> 00:27:48,066
ACTUALLY UNTIL COPY EDITS FOR MY

742
00:27:48,133 --> 00:27:49,701
BOOK WERE DUE, I WAS GLAD IT

743
00:27:49,768 --> 00:27:54,406
CHANGED IN TIME TO GET INTO MY

744
00:27:54,472 --> 00:27:54,639
BOOK.

745
00:27:54,706 --> 00:27:57,575
SO, THE KIDNEY ALGORITHM HAS

746
00:27:57,642 --> 00:27:57,942
CHANGED.

747
00:27:58,009 --> 00:28:00,345
EGFR NO LONGER INCLUDES A RACE

748
00:28:00,412 --> 00:28:00,879
CORRECTION.

749
00:28:00,945 --> 00:28:04,616
BUT WE DO HAVE TO THINK ABOUT

750
00:28:04,682 --> 00:28:07,585
EVERY SINGLE LAB THAT USED THESE

751
00:28:07,652 --> 00:28:10,188
DIFFERENT NUMBERS, RIGHT?

752
00:28:10,255 --> 00:28:12,157
SO EVERY SINGLE LAB WOULD TAKE

753
00:28:12,223 --> 00:28:14,659
YOUR RACE ON YOUR PATIENT CHART

754
00:28:14,726 --> 00:28:17,295
AND WOULD CALCULATE IT AS EITHER

755
00:28:17,362 --> 00:28:18,630
BLACK OR WHITE, AND I DISCOVERED

756
00:28:18,696 --> 00:28:21,699
THE OTHER DAY THAT MY LAB WAS

757
00:28:21,766 --> 00:28:24,803
ACTUALLY STILL CALCULATING IT.

758
00:28:24,869 --> 00:28:26,671
IT'S JUST THAT THE NUMBER GOT

759
00:28:26,738 --> 00:28:29,874
MUSHED AND BY THE TIME I SAW IT

760
00:28:29,941 --> 00:28:32,310
IN MY ELECTRONIC MEDICAL RECORD

761
00:28:32,377 --> 00:28:34,479
IT HAD BEEN, YOU KNOW, THE TWO

762
00:28:34,546 --> 00:28:35,246
NUMBERS HAD BEEN SQUISHED

763
00:28:35,313 --> 00:28:36,314
TOGETHER OR SOMETHING.

764
00:28:36,381 --> 00:28:39,851
SO, WE HAVE TO THINK ABOUT THESE

765
00:28:39,918 --> 00:28:41,052
SOCIOTECHNICAL SYSTEMS, THINK

766
00:28:41,119 --> 00:28:43,354
ABOUT THE DATA PIPELINES, AND

767
00:28:43,421 --> 00:28:46,291
WHEN WE MAKE CHANGES LIKE WE

768
00:28:46,357 --> 00:28:48,059
CHANGE THE FORMULA TO ELIMINATE

769
00:28:48,126 --> 00:28:48,726
RACE CORRECTION VARIABLE, WE DO

770
00:28:48,793 --> 00:28:51,996
HAVE TO THINK ABOUT THE

771
00:28:52,063 --> 00:28:53,531
PIPELINES TO THINK ABOUT IT'S

772
00:28:53,598 --> 00:28:55,066
COMPLICATED TO CHANGE THESE

773
00:28:55,133 --> 00:28:56,367
THINGS AND IMPLICATIONS OF A.I.

774
00:28:56,434 --> 00:28:56,901
MODELS, RIGHT?

775
00:28:56,968 --> 00:29:00,505
IF THE A.I. MODEL IS BEING

776
00:29:00,572 --> 00:29:03,508
TRAMMED WITH SOME KIND OF

777
00:29:03,575 --> 00:29:04,809
DISCRIMINATORY ASSUMPTION INSIDE

778
00:29:04,876 --> 00:29:06,444
IT AND THEN WE REALIZE IT, THEN

779
00:29:06,511 --> 00:29:09,280
YOU HAVE TO GO AND YOU HAVE TO

780
00:29:09,347 --> 00:29:10,281
RETRAIN THE MODEL AND YOU HAVE

781
00:29:10,348 --> 00:29:13,451
TO CHANGE THE WHOLE PIPELINE,

782
00:29:13,518 --> 00:29:18,723
IT'S ACTUALLY AN EXTREMELY,

783
00:29:18,790 --> 00:29:19,357
EXTREMELY COMPLICATED AND

784
00:29:19,424 --> 00:29:20,225
RESOURCE-INTENSIVE EFFORT.

785
00:29:20,291 --> 00:29:22,660
WE SHOULD BE DOING IT.

786
00:29:22,727 --> 00:29:24,629
BUT IT'S WAY MORE COMPLICATED

787
00:29:24,696 --> 00:29:26,030
THAN EVERYBODY IMAGINES BECAUSE

788
00:29:26,097 --> 00:29:29,200
WE GENERALLY THINK ABOUT

789
00:29:29,267 --> 00:29:31,936
COMPUTATIONAL SOLUTIONS AS BEING

790
00:29:32,003 --> 00:29:32,804
FAST, CHEAP, EFFICIENT, AND THEY

791
00:29:32,871 --> 00:29:37,208
ARE NOT ALL OF THESE THINGS AT

792
00:29:37,275 --> 00:29:41,146
THE SAME TIME.

793
00:29:41,212 --> 00:29:45,750
LLMs ARE ON EVERYBODY'S MINDS,

794
00:29:45,817 --> 00:29:48,153
AND UNFORTUNATELY LLMs HAVE

795
00:29:48,219 --> 00:29:52,857
THE SAME KINDS OF PROBLEMS WHEN

796
00:29:52,924 --> 00:29:57,228
IT COMES TO RACE-BASED MEDICINE.

797
00:29:57,295 --> 00:29:59,197
THEY HAVE THE BIASES THAT WE

798
00:29:59,264 --> 00:30:00,498
TALKED ABOUT BEFORE.

799
00:30:00,565 --> 00:30:02,500
SO LLMs ARE PRETTY BAD AT

800
00:30:02,567 --> 00:30:02,834
MATH.

801
00:30:02,901 --> 00:30:06,304
SO, NO, THEY ARE NOT CALCULATING

802
00:30:06,371 --> 00:30:08,806
ANYBODY'S EGFR SCORE.

803
00:30:08,873 --> 00:30:14,412
BUT THEY ARE DOING THINGS LIKE

804
00:30:14,479 --> 00:30:16,915
PROPAGATING RACE-BASED MEDICINE

805
00:30:16,981 --> 00:30:22,053
THAT IS HARMFUL, THAT IS RACIST,

806
00:30:22,120 --> 00:30:26,224
THAT IS PROBLEMATIC, IN PART

807
00:30:26,291 --> 00:30:28,293
BECAUSE THESE LLMs, LARGE

808
00:30:28,359 --> 00:30:31,930
LANGUAGE MODELS, GENERATIVE A.I.

809
00:30:31,996 --> 00:30:35,600
SYSTEMS, DO NOT -- THEY ARE NOT

810
00:30:35,667 --> 00:30:37,368
PARSING FOR CONTENT.

811
00:30:37,435 --> 00:30:39,337
THEY ARE BASICALLY LANGUAGE

812
00:30:39,404 --> 00:30:39,637
CALCULATORS.

813
00:30:39,704 --> 00:30:46,477
THEY ARE GIVING US STRINGS OF

814
00:30:46,544 --> 00:30:50,982
TEXT BASED ON STRINGS OF  TEXT

815
00:30:51,049 --> 00:30:52,383
BASED IN THE PAST.

816
00:30:52,450 --> 00:30:53,484
RACISM IS NOT GOOD BUT

817
00:30:53,551 --> 00:30:56,087
UNFORTUNATELY IT IS POPULAR.

818
00:30:56,154 --> 00:31:01,159
AND POPULAR IS USED AS A PROXY

819
00:31:01,226 --> 00:31:04,629
FOR GOOD INSIDE LARGE ALGORITHMS

820
00:31:04,696 --> 00:31:04,996
(INDISCERNIBLE).

821
00:31:05,063 --> 00:31:06,397
AND OVERALL, RACE AND MACHINE

822
00:31:06,464 --> 00:31:10,335
LEARNING IS A LITTLE BIT OF A

823
00:31:10,401 --> 00:31:11,236
MESS.

824
00:31:11,302 --> 00:31:14,372
SO, AS I SAID BEFORE, RACE IS

825
00:31:14,439 --> 00:31:16,708
OFTEN EMBEDDED IN THESE

826
00:31:16,774 --> 00:31:18,343
COMPUTATIONAL SYSTEMS, AS --

827
00:31:18,409 --> 00:31:24,816
WELL,  RACISM IS EMBEDDED IN

828
00:31:24,882 --> 00:31:26,251
MEDICAL SYSTEMS AND ABSOLUTELY

829
00:31:26,317 --> 00:31:28,119
EMBEDDED IN ML SYSTEMS, MACHINE

830
00:31:28,186 --> 00:31:29,988
LEARNING SYSTEMS, AS IF IT WERE

831
00:31:30,054 --> 00:31:31,656
A BIOLOGICAL AND SOCIAL REALITY.

832
00:31:31,723 --> 00:31:36,327
SO WE NEED TO DO MORE REFLECTING

833
00:31:36,394 --> 00:31:39,631
ON THE SOCIAL UNDERPINNINGS OF

834
00:31:39,697 --> 00:31:45,203
SYSTEMS BEFORE WE GO MAKING

835
00:31:45,270 --> 00:31:47,305
ALGORITHMIC SYSTEMS AND

836
00:31:47,372 --> 00:31:49,907
FOSSILIZING THESE DISCRIMINATORY

837
00:31:49,974 --> 00:31:50,908
DECISIONS IN CODE.

838
00:31:50,975 --> 00:31:53,077
NOW, ONE OF THE THINGS PEOPLE

839
00:31:53,144 --> 00:31:55,380
OFTEN SAY IS THAT THE A.I. WILL

840
00:31:55,446 --> 00:31:57,015
GET BETTER EVENTUALLY.

841
00:31:57,081 --> 00:31:58,516
RIGHT?

842
00:31:58,583 --> 00:32:02,487
TALK ABOUT THE FACIAL

843
00:32:02,553 --> 00:32:03,254
RECOGNITION SYSTEMS THAT

844
00:32:03,321 --> 00:32:04,889
MISIDENTIFY WOMEN AND PEOPLE OF

845
00:32:04,956 --> 00:32:05,123
COLOR.

846
00:32:05,189 --> 00:32:06,424
CAN'T YOU JUST MAKE IT BETTER?

847
00:32:06,491 --> 00:32:08,526
AND I TALK ABOUT THE MORTGAGE

848
00:32:08,593 --> 00:32:10,194
APPROVAL SYSTEMS, WELL, CAN'T

849
00:32:10,261 --> 00:32:11,162
YOU JUST DO THIS?

850
00:32:11,229 --> 00:32:12,363
I TALK ABOUT THE RACE

851
00:32:12,430 --> 00:32:13,364
CORRECTIONS, THEY ARE LIKE,

852
00:32:13,431 --> 00:32:17,935
WELL, CAN'T YOU JUST DO THIS?

853
00:32:18,002 --> 00:32:20,905
YEAH, OKAY, I HEAR YOU THAT

854
00:32:20,972 --> 00:32:24,509
THINGS IMPROVE BUT I'M NOT

855
00:32:24,575 --> 00:32:28,346
CONVINCED THAT THAT ROSY

856
00:32:28,413 --> 00:32:29,647
TECHNOLOGICAL SEAMLESS FUTURE IS

857
00:32:29,714 --> 00:32:31,182
RIGHT AROUND THE CORNER BECAUSE

858
00:32:31,249 --> 00:32:32,850
I'VE HEARD TOO MANY TIMES THAT,

859
00:32:32,917 --> 00:32:35,019
OH, YEAH, IN FIVE YEARS IT'S

860
00:32:35,086 --> 00:32:36,421
GOING TO BE, YOU KNOW, ALL THE

861
00:32:36,487 --> 00:32:38,456
PROBLEMS ARE GOING TO BE FIXED.

862
00:32:38,523 --> 00:32:42,226
AND ONE OF THE THINGS I'VE BEEN

863
00:32:42,293 --> 00:32:43,561
WONDERING ABOUT IS WHETHER ALL

864
00:32:43,628 --> 00:32:46,397
OF THE PROBLEMS THAT ARE EASY TO

865
00:32:46,464 --> 00:32:48,366
SOLVE WITH COMPUTERS HAVE BEEN

866
00:32:48,433 --> 00:32:52,737
SOLVED, AND WE'RE JUST LEFT WITH

867
00:32:52,804 --> 00:32:55,573
THE REALLY COMPLICATED

868
00:32:55,640 --> 00:32:57,475
SOCIOTECHNICAL PROBLEMS, AND A

869
00:32:57,542 --> 00:32:59,243
LOT OF SOCIAL PROBLEMS HAVEN'T

870
00:32:59,310 --> 00:33:01,245
BEEN FIXED IN THOUSANDS OF YEARS

871
00:33:01,312 --> 00:33:03,915
SO I'M NOT SURE WHY WE EXPECT

872
00:33:03,981 --> 00:33:05,616
WE'LL BE ABLE TO BUILD MACHINES

873
00:33:05,683 --> 00:33:07,418
TO FIX THEM IN A COUPLE MONTHS

874
00:33:07,485 --> 00:33:08,519
OR JUST A COUPLE MILLION

875
00:33:08,586 --> 00:33:11,322
DOLLARS.

876
00:33:11,389 --> 00:33:14,292
AND ONE EXAMPLE OF THIS COMES

877
00:33:14,359 --> 00:33:15,927
FROM A NEW NEWS ORGANIZATION,

878
00:33:15,993 --> 00:33:20,531
THAT WAS STARTED BY JULIA

879
00:33:20,598 --> 00:33:26,003
ENGLAND, FORMERLY OF PRO PUBLIC

880
00:33:26,070 --> 00:33:29,841
A , A NEW ORGANIZATION CALLED

881
00:33:29,907 --> 00:33:31,576
PROOF.

882
00:33:31,642 --> 00:33:33,811
PROOF LABS AND ALONDRA NELSON'S

883
00:33:33,878 --> 00:33:36,948
LAB JOINED TOGETHER TO DO THE

884
00:33:37,014 --> 00:33:38,383
A.I. DEMOCRACY PROJECT.

885
00:33:38,449 --> 00:33:42,120
THEY JUST DID THIS REALLY

886
00:33:42,186 --> 00:33:43,988
SPECTACULAR BENCHMARKING STUDY

887
00:33:44,055 --> 00:33:47,158
WHERE THEY LOOKED AT THE FIVE

888
00:33:47,225 --> 00:33:49,560
LEADING A.I. MODELS AND

889
00:33:49,627 --> 00:33:51,095
BENCHMARKED THEM.

890
00:33:51,162 --> 00:33:55,366
OKAY, INSIDE THIS VERTICAL,

891
00:33:55,433 --> 00:33:58,002
RIGHT, INSIDE ELECTIONS,

892
00:33:58,069 --> 00:33:58,569
SOMETHING THAT'S INCREDIBLY

893
00:33:58,636 --> 00:34:02,240
RELEVANT TO U.S. DEMOCRACY

894
00:34:02,306 --> 00:34:04,142
SUCCEEDING, ELECTIONS, HOW MUCH

895
00:34:04,208 --> 00:34:05,343
MISINFORMATION IS REALLY BEING

896
00:34:05,410 --> 00:34:06,210
GENERATED BY THESE MODELS?

897
00:34:06,277 --> 00:34:07,979
BECAUSE ONE OF THE THINGS WE

898
00:34:08,045 --> 00:34:10,681
HAVEN'T HAD TO DATE IS WE

899
00:34:10,748 --> 00:34:13,317
HAVEN'T HAD A BENCHMARK FOR

900
00:34:13,384 --> 00:34:16,821
LLMs, HAVEN'T BEEN ABLE TO

901
00:34:16,888 --> 00:34:19,557
SAY, OKAY, LLMs ARE GENERATING

902
00:34:19,624 --> 00:34:21,659
USEFUL CONTENT, 10% OF THE TIME,

903
00:34:21,726 --> 00:34:22,627
100% OF THE TIME.

904
00:34:22,693 --> 00:34:24,929
WE JUST DON'T KNOW.

905
00:34:24,996 --> 00:34:26,564
ALL WE HAVE ARE QUALITATIVE

906
00:34:26,631 --> 00:34:26,864
CLAIMS.

907
00:34:26,931 --> 00:34:31,269
ONE OF THE THINGS BRILLIANT

908
00:34:31,335 --> 00:34:33,037
ABOUT THIS PROJECT, LET'S TEST

909
00:34:33,104 --> 00:34:33,905
THESE SYSTEMS.

910
00:34:33,971 --> 00:34:37,175
THEY BUILT THIS TOOL THAT LOOKED

911
00:34:37,241 --> 00:34:45,850
AT -- HIT THE APIs OF ALL FIVE

912
00:34:45,917 --> 00:34:46,651
LANGUAGE MODELS SIMULTANEOUSLY,

913
00:34:46,717 --> 00:34:47,985
HIT WITH THE SAME PROMPT AND GET

914
00:34:48,052 --> 00:34:50,087
BACK RESPONSES FROM EACH OF THE

915
00:34:50,154 --> 00:34:50,421
MODELS.

916
00:34:50,488 --> 00:34:52,890
IT WAS COOL SEEING RESPONSES ON

917
00:34:52,957 --> 00:34:55,660
ONE PAGE BECAUSE WHEN YOU GO TO

918
00:34:55,726 --> 00:34:57,128
THE SITES INDIVIDUALLY, YOU GET

919
00:34:57,195 --> 00:35:01,098
DISTRACTED BY THE WHOLE USER

920
00:35:01,165 --> 00:35:02,300
EXPERIENCE, BUT HAVING THEM ON

921
00:35:02,366 --> 00:35:05,803
THE SAME PAGE WAS EASY TO SEE

922
00:35:05,870 --> 00:35:06,537
DIFFERENCES BETWEEN THEM.

923
00:35:06,604 --> 00:35:08,973
AND YOU KIND OF GOT A SENSE OF

924
00:35:09,040 --> 00:35:13,010
WHAT WAS HAPPENING IN EACH OF

925
00:35:13,077 --> 00:35:14,245
THE MODELS.

926
00:35:14,312 --> 00:35:15,847
SO, THE RESPONSES WERE EVALUATED

927
00:35:15,913 --> 00:35:16,614
BY TEAMS.

928
00:35:16,681 --> 00:35:20,451
THE TEAMS WERE MADE UP OF

929
00:35:20,518 --> 00:35:23,054
JOURNALISTS AND ELECTION

930
00:35:23,120 --> 00:35:24,722
OFFICIALS, SO PEOPLE WHO HAVE A

931
00:35:24,789 --> 00:35:27,325
LITTLE BIT OF SUBJECT MATTER

932
00:35:27,391 --> 00:35:28,960
EXPERTISE AND, YOU KNOW, A LOT

933
00:35:29,026 --> 00:35:31,262
OF GENERAL KNOWLEDGE AND THEN

934
00:35:31,329 --> 00:35:34,098
ALSO PEOPLE WITH DEEP SUBJECT

935
00:35:34,165 --> 00:35:35,433
AREA KNOWLEDGE.

936
00:35:35,500 --> 00:35:38,703
AND WHAT THEY FOUND WAS THAT

937
00:35:38,769 --> 00:35:42,006
ABOUT HALF OF THE MODELS'

938
00:35:42,073 --> 00:35:44,709
ANSWERS WERE INACCURATE WHEN IT

939
00:35:44,775 --> 00:35:47,678
CAME TO QUESTIONS ABOUT U.S.

940
00:35:47,745 --> 00:35:48,913
ELECTIONS.

941
00:35:48,980 --> 00:35:51,048
THEY EVALUATED THE ANSWERS ON

942
00:35:51,115 --> 00:35:56,153
FOUR DIMENSIONS, SO WAS IT

943
00:35:56,220 --> 00:35:57,855
INACCURATE, HARMFUL, INCOMPLETE,

944
00:35:57,922 --> 00:35:59,090
BIASED?

945
00:35:59,156 --> 00:36:03,127
I REALLY LIKE THIS

946
00:36:03,194 --> 00:36:04,061
FOUR-DIMENSIONAL APPROACH

947
00:36:04,128 --> 00:36:05,329
BECAUSE JUST SAYING, OH, A.I. IS

948
00:36:05,396 --> 00:36:07,398
GOOD OR A.I. IS BAD IS NOT

949
00:36:07,465 --> 00:36:09,667
REALLY GETTING US VERY FAR

950
00:36:09,734 --> 00:36:10,268
ANYMORE.

951
00:36:10,334 --> 00:36:13,938
I REALLY LIKE THIS APPROACH OF

952
00:36:14,005 --> 00:36:19,844
GOING AT A.I. IN CONTEXT, IN A

953
00:36:19,911 --> 00:36:21,479
PARTICULAR TOPIC AREA, AND

954
00:36:21,546 --> 00:36:25,650
EVALUATING IT FOR QUALITIES THAT

955
00:36:25,716 --> 00:36:26,751
WE CARE ABOUT.

956
00:36:26,817 --> 00:36:28,252
SO, I LOVE THIS APPROACH.

957
00:36:28,319 --> 00:36:29,487
THIS IS REALLY NEW

958
00:36:29,554 --> 00:36:31,889
METHODOLOGIES, JUST CAME OUT TWO

959
00:36:31,956 --> 00:36:32,356
WEEKS AGO.

960
00:36:32,423 --> 00:36:34,292
SO, HIGHLY RECOMMEND CHECKING

961
00:36:34,358 --> 00:36:35,726
THIS OUT.

962
00:36:35,793 --> 00:36:38,696
IN GENERAL THOUGH, THINGS I

963
00:36:38,763 --> 00:36:40,398
RECOMMEND ARE LOOKING FOR HUMAN

964
00:36:40,464 --> 00:36:44,035
PROBLEMS INSIDE A.I. SYSTEMS.

965
00:36:44,101 --> 00:36:46,437
SO, A.I. HAS A LOT OF BOOSTERS.

966
00:36:46,504 --> 00:36:47,972
YOU DON'T NEED ME TO STAND UP

967
00:36:48,039 --> 00:36:49,607
HERE AND TALK ABOUT HOW AWESOME

968
00:36:49,674 --> 00:36:50,374
A.I. IS.

969
00:36:50,441 --> 00:36:52,777
YOU'VE HAD A LOT OF PEOPLE

970
00:36:52,843 --> 00:36:53,878
TELLING YOU HOW AWESOME IT IS

971
00:36:53,945 --> 00:36:55,846
AND IT DOES COOL STUFF.

972
00:36:55,913 --> 00:37:00,751
WE CAN DEFINITELY TAKE THAT FOR

973
00:37:00,818 --> 00:37:01,319
GRANTED.

974
00:37:01,385 --> 00:37:03,621
BUT I'M HERE TO SAY, HEY, LET'S

975
00:37:03,688 --> 00:37:07,124
ALSO LOOK FOR THE HUMAN PROBLEMS

976
00:37:07,191 --> 00:37:08,659
BECAUSE NOTHING IS PERFECT.

977
00:37:08,726 --> 00:37:13,331
COLLABORATION IS KEY IN THIS

978
00:37:13,397 --> 00:37:13,564
REGARD.

979
00:37:13,631 --> 00:37:14,665
SO, COLLABORATIONS WITH

980
00:37:14,732 --> 00:37:17,401
HUMANISTS AND SOCIAL SCIENTISTS,

981
00:37:17,468 --> 00:37:18,803
AND TECHNOLOGISTS, AND

982
00:37:18,869 --> 00:37:20,237
BIOMEDICAL RESEARCHERS, WE CAN

983
00:37:20,304 --> 00:37:22,006
TEST OUR TECHNOLOGY FOR

984
00:37:22,073 --> 00:37:24,709
ACCESSIBILITY, WE CAN ENGAGE IN

985
00:37:24,775 --> 00:37:25,610
ALGORITHMIC AUDITING, SO

986
00:37:25,676 --> 00:37:29,413
BENCHMARKING AS A KIND OF

987
00:37:29,480 --> 00:37:30,781
ALGORITHMIC BENCHMARKING.

988
00:37:30,848 --> 00:37:33,150
WHEN WE AUDIT WE OPEN THE BLACK

989
00:37:33,217 --> 00:37:35,019
BOX OF ALGORITHMING, AND EXPLAIN

990
00:37:35,086 --> 00:37:37,221
WHAT'S HAPPENING INSIDE AND

991
00:37:37,288 --> 00:37:39,824
ALLOW PEOPLE SOME INSIGHTS AND

992
00:37:39,890 --> 00:37:43,127
THUS EMPOWER THEM.

993
00:37:43,194 --> 00:37:44,195
AND OVERALL, I THINK THAT

994
00:37:44,261 --> 00:37:46,397
THERE'S A LOT OF DISCOURSE ABOUT

995
00:37:46,464 --> 00:37:48,299
A.I. BEING TRANSFORMATIVE AND

996
00:37:48,366 --> 00:37:49,700
GOOD FOR GENERAL PURPOSES.

997
00:37:49,767 --> 00:37:50,668
I THINK IT'S ACTUALLY SAFER TO

998
00:37:50,735 --> 00:37:54,605
THINK ABOUT A.I. AS BEING GOOD

999
00:37:54,672 --> 00:37:58,576
FOR LIMITED LOW-STAKES MUNDANE

1000
00:37:58,643 --> 00:37:59,443
TASKS.

1001
00:37:59,510 --> 00:38:01,412
NOT FOR HIGH-STAKES TASKS, NOT

1002
00:38:01,479 --> 00:38:05,149
FOR GENERAL PURPOSE USE.

1003
00:38:05,216 --> 00:38:05,349
RIGHT?

1004
00:38:05,416 --> 00:38:12,356
SO, HELPING AT CERTAIN PHASES OF

1005
00:38:12,423 --> 00:38:13,524
THE DIAGNOSTIC PROCESS NOT

1006
00:38:13,591 --> 00:38:13,958
(INDISCERNIBLE).

1007
00:38:14,025 --> 00:38:17,094
I'D LIKE TO END WITH A BUNCH OF

1008
00:38:17,161 --> 00:38:19,030
RESOURCES FOR LEARNING MORE.

1009
00:38:19,096 --> 00:38:23,834
SOME OF MY FAVORITES RIGHT NOW

1010
00:38:23,901 --> 00:38:27,672
IN ORDER TO LEARN MORE ABOUT

1011
00:38:27,738 --> 00:38:29,240
RACIAL DISPARITIES IN MEDICINE,

1012
00:38:29,306 --> 00:38:36,113
DR. BLACKSTOCK HAS A NEW BOOK

1013
00:38:36,180 --> 00:38:38,616
OUT, "LEGACY," IT'S FANTASTIC,

1014
00:38:38,683 --> 00:38:42,586
AND SHE ALSO HAS A CONSULTING

1015
00:38:42,653 --> 00:38:44,755
AGENCY, HEALTH EQUITY PARTNERS.

1016
00:38:44,822 --> 00:38:47,591
ON SOCIAL MEDIA THERE'S A

1017
00:38:47,658 --> 00:38:48,826
MEDICAL STUDENT, JOEL BROVELL,

1018
00:38:48,893 --> 00:38:50,428
SOME OF YOU HAVE SEEN HIM COME

1019
00:38:50,494 --> 00:38:52,730
ACROSS YOUR FEED.

1020
00:38:52,797 --> 00:38:55,933
HE DOES REALLY DELIGHTFUL

1021
00:38:56,000 --> 00:38:56,667
ACCESSIBLE WORK.

1022
00:38:56,734 --> 00:38:59,537
YOU CAN LEARN FROM JOURNALISTS

1023
00:38:59,603 --> 00:39:01,072
ON THE ALGORITHMIC

1024
00:39:01,138 --> 00:39:05,876
ACCOUNTABILITY BEAT, SO PRACTICE

1025
00:39:05,943 --> 00:39:07,845
-- TRADITIONAL IS FUNCTION OF

1026
00:39:07,912 --> 00:39:10,347
THE MEDIA IS HOLD POWER

1027
00:39:10,414 --> 00:39:11,615
ACCOUNTABILITY, THAT

1028
00:39:11,682 --> 00:39:13,217
ACCOUNTABILITY FUNCTION HAS TO

1029
00:39:13,284 --> 00:39:14,885
TRANSFER ONTO ALGORITHMS AND

1030
00:39:14,952 --> 00:39:16,754
ONTO THEIR MAKERS.

1031
00:39:16,821 --> 00:39:18,289
SO ALGORITHMIC ACCOUNTABILITY IS

1032
00:39:18,355 --> 00:39:19,490
A BEAT.

1033
00:39:19,557 --> 00:39:26,931
I MENTIONED THE MARKUP,

1034
00:39:26,997 --> 00:39:27,264
PRO-PUBLICA.

1035
00:39:27,331 --> 00:39:28,265
ALGORITHMIC ACCOUNTABILITY I

1036
00:39:28,332 --> 00:39:30,201
SHOULD SAY IS EXPENSIVE AND TIME

1037
00:39:30,267 --> 00:39:31,202
CONSUMING AND REQUIRES A LOT OF

1038
00:39:31,268 --> 00:39:33,404
PEOPLE TO DO WELL, SO IF YOU'RE

1039
00:39:33,471 --> 00:39:34,405
WONDERING WHY THERE'S NOT MORE

1040
00:39:34,472 --> 00:39:37,441
HAPPENING AND WHY PEOPLE ARE NOT

1041
00:39:37,508 --> 00:39:39,410
DOING ALGORITHMIC AUDITING OF

1042
00:39:39,477 --> 00:39:40,644
SOCIAL MEDIA PLATFORMS, WELL,

1043
00:39:40,711 --> 00:39:42,046
IT'S BECAUSE IT'S EXTREMELY

1044
00:39:42,113 --> 00:39:44,248
EXPENSIVE AND TIME CONSUMING AND

1045
00:39:44,315 --> 00:39:46,984
COMPLICATED, AND TECH COMPANIES

1046
00:39:47,051 --> 00:39:48,619
ARE DELIBERATELY TRYING TO KEEP

1047
00:39:48,686 --> 00:39:49,887
PEOPLE FROM HAVING ACCESS TO THE

1048
00:39:49,954 --> 00:39:52,089
DATA.

1049
00:39:54,158 --> 00:39:58,496
WE HEARD EARLIER ABOUT SOME OF

1050
00:39:58,562 --> 00:40:03,067
THE GOVERNMENT RESOURCES AROUND

1051
00:40:03,134 --> 00:40:04,034
RESPONSIBLE A.I.

1052
00:40:04,101 --> 00:40:05,336
SO LIKE THE RISK MANAGEMENT

1053
00:40:05,402 --> 00:40:06,237
FRAMEWORK OUT OF NIST.

1054
00:40:06,303 --> 00:40:08,973
I REALLY LIKE SOMETHING CALLED

1055
00:40:09,039 --> 00:40:10,074
THE ALGORITHMIC TRANSPARENCY

1056
00:40:10,141 --> 00:40:13,110
PLAYBOOK FROM MY COLLEAGUES AT

1057
00:40:13,177 --> 00:40:14,111
THE RESPONSIBLE -- CENTER FOR

1058
00:40:14,178 --> 00:40:17,047
RESPONSIBLE A.I. AT NYU, AND

1059
00:40:17,114 --> 00:40:18,916
THEN DATA IN SOCIETY, EQUAL A.I.

1060
00:40:18,983 --> 00:40:21,318
WITH DOING SOME GOOD STUFF

1061
00:40:21,385 --> 00:40:24,722
AROUND ALGORITHMIC IMPACT

1062
00:40:24,789 --> 00:40:25,389
ASSESSMENT TOOLING.

1063
00:40:25,456 --> 00:40:29,226
THE SAME WE DO ENVIRONMENTAL

1064
00:40:29,293 --> 00:40:30,327
IMPACT ASSESSMENT WHEN WE BUILD

1065
00:40:30,394 --> 00:40:32,496
A NEW BUILDING, WE CAN DO AN

1066
00:40:32,563 --> 00:40:34,365
ALGORITHMIC IMPACT ASSESSMENT IN

1067
00:40:34,431 --> 00:40:37,134
ORDER TO LOOK AT HOW WELL OR HOW

1068
00:40:37,201 --> 00:40:39,103
BADLY AN ALGORITHM OR

1069
00:40:39,170 --> 00:40:40,471
ALGORITHMIC SYSTEM IS AFFECTING

1070
00:40:40,538 --> 00:40:43,440
A COMMUNITY.

1071
00:40:43,507 --> 00:40:46,844
IF YOU LIKE, MORE THAN A GLITCH,

1072
00:40:46,911 --> 00:40:49,580
OR YOU LIKE ARTIFICIAL

1073
00:40:49,647 --> 00:40:53,083
UNINTELLIGENCE, LEER  -- HERE

1074
00:40:53,150 --> 00:40:54,418
ARE A HANDFUL OF OTHER THINGS

1075
00:40:54,485 --> 00:40:56,821
YOU MIGHT WANT TO READ OR WATCH.

1076
00:40:56,887 --> 00:40:59,623
AND OF COURSE THERE'S TWO PAGES

1077
00:40:59,690 --> 00:41:01,158
OF SYLLABUS BECAUSE I'M A

1078
00:41:01,225 --> 00:41:04,328
PROFESSOR, I CAN'T HELP MYSELF.

1079
00:41:04,395 --> 00:41:05,162
AND WITH THAT, I'M GOING TO

1080
00:41:05,229 --> 00:41:06,864
THANK YOU VERY MUCH AND SAY WE

1081
00:41:06,931 --> 00:41:09,867
HAVE TIME FOR SOME QUESTIONS.

1082
00:41:25,115 --> 00:41:27,451
>> THANK YOU SO MUCH.

1083
00:41:27,518 --> 00:41:29,186
THIS WAS A PHENOMENAL,

1084
00:41:29,253 --> 00:41:30,054
PHENOMENAL LECTURE.

1085
00:41:30,120 --> 00:41:31,922
I KNOW SOMETHING THAT WE WILL BE

1086
00:41:31,989 --> 00:41:37,862
DISCUSSING QUITE A BIT AS AN NLM

1087
00:41:37,928 --> 00:41:38,729
COMMUNITY AND COLLEAGUES ACROSS

1088
00:41:38,796 --> 00:41:41,265
THE NIH, GOVERNMENT AND BEYOND.

1089
00:41:41,332 --> 00:41:43,334
I AM THE DATA SCIENCE AND OPEN

1090
00:41:43,400 --> 00:41:45,135
SCIENCE OFFICER IN THE OFFICE OF

1091
00:41:45,202 --> 00:41:47,905
STRATEGIC INITIATIVES.

1092
00:41:47,972 --> 00:41:49,540
I'LL BE MODERATING THE Q&A.

1093
00:41:49,607 --> 00:41:54,712
FOR THOSE WHO ARE JOINING US

1094
00:41:54,778 --> 00:41:56,580
REMOTELY, YOU CAN USE THE "LIVE

1095
00:41:56,647 --> 00:41:58,415
FEEDBACK" BUTTON UNDER YOUR

1096
00:41:58,482 --> 00:41:59,850
VIDEOCAST LINK TO ASK YOUR

1097
00:41:59,917 --> 00:42:00,284
QUESTIONS.

1098
00:42:00,351 --> 00:42:02,820
I EVEN HAVE THEM COMING UP HERE

1099
00:42:02,887 --> 00:42:04,321
ON MY COMPUTER.

1100
00:42:04,388 --> 00:42:07,691
IF YOU ARE IN THE ROOM, YOU ARE

1101
00:42:07,758 --> 00:42:10,160
WELCOME TO HEAD TO THIS

1102
00:42:10,227 --> 00:42:13,330
MICROPHONE TO ASK YOUR QUESTIONS

1103
00:42:13,397 --> 00:42:14,331
AS WELL.

1104
00:42:14,398 --> 00:42:15,866
I WILL KICK US OFF.

1105
00:42:15,933 --> 00:42:17,935
MINE IS A SORT OF LONG QUESTION

1106
00:42:18,002 --> 00:42:21,105
SO I WROTE IT DOWN.

1107
00:42:21,171 --> 00:42:24,475
BUT YOU STARTED OFF YOUR TALK BY

1108
00:42:24,541 --> 00:42:26,243
TELLING US THAT BRAINS ARE

1109
00:42:26,310 --> 00:42:28,646
BETTER AT RECALLING STORIES THAN

1110
00:42:28,712 --> 00:42:30,781
THEY ARE FACTS AND PHYSICS.

1111
00:42:30,848 --> 00:42:33,550
AND WHAT STRUCK ME BOTH IN YOUR

1112
00:42:33,617 --> 00:42:35,753
WORK AND OVER THE COURSE OF THE

1113
00:42:35,819 --> 00:42:37,187
CRAFT OF THIS LECTURE THAT

1114
00:42:37,254 --> 00:42:39,924
YOU'VE JUST GIVEN US IS THAT YOU

1115
00:42:39,990 --> 00:42:43,060
ARE BOTH A STORY TELLER OF NOT

1116
00:42:43,127 --> 00:42:46,697
ONLY WHAT CAN GO WRONG AND WHY

1117
00:42:46,764 --> 00:42:49,967
IT MATTERS BUT ALSO THE HOW, HOW

1118
00:42:50,034 --> 00:42:51,936
DO THINGS GO WRONG, AND WHAT CAN

1119
00:42:52,002 --> 00:42:55,673
WE DO ABOUT IT, FROM THAT SORT

1120
00:42:55,739 --> 00:42:57,107
OF DEEP TECHNICAL PRACTITIONER

1121
00:42:57,174 --> 00:42:58,642
EXPERTISE THAT YOU HAVE.

1122
00:42:58,709 --> 00:43:00,611
AND THIS TO ME SEEMS LIKE A

1123
00:43:00,678 --> 00:43:02,479
REALLY IMPORTANT PART OF SORT OF

1124
00:43:02,546 --> 00:43:05,316
GETTING INTO THAT BLACK BOX AND

1125
00:43:05,382 --> 00:43:08,819
BRINGING MORE PEOPLE WHO MAY NOT

1126
00:43:08,886 --> 00:43:11,021
BE CODERS THEMSELVES BUT HAVE

1127
00:43:11,088 --> 00:43:12,723
REAL SORT OF STAKES AND

1128
00:43:12,790 --> 00:43:15,392
INTERESTS AND THERE ARE REAL

1129
00:43:15,459 --> 00:43:17,995
IMPLICATIONS FOR THEIR LIVES, TO

1130
00:43:18,062 --> 00:43:18,595
THAT PROCESS, TO CO-CREATE

1131
00:43:18,662 --> 00:43:21,398
SOLUTIONS AND BE PART OF THAT

1132
00:43:21,465 --> 00:43:21,732
DISCUSSION.

1133
00:43:21,799 --> 00:43:23,801
SO I'M WONDERING, YOU KNOW, HOW

1134
00:43:23,867 --> 00:43:25,569
IS IT THAT YOU HAVE -- THIS IS

1135
00:43:25,636 --> 00:43:27,438
KIND OF A CRAFT QUESTION, RATHER

1136
00:43:27,504 --> 00:43:28,172
THAN A TECHNICAL ONE.

1137
00:43:28,238 --> 00:43:30,808
BUT HOW IS IT THAT YOU SORT OF

1138
00:43:30,874 --> 00:43:32,443
HONED YOUR CRAFT AND WHAT ARE

1139
00:43:32,509 --> 00:43:35,312
SOME OF THE SORT OF -- IF YOU

1140
00:43:35,379 --> 00:43:38,148
HAVE ANECDOTES ABOUT INTERESTING

1141
00:43:38,215 --> 00:43:39,917
CONVERSATIONS OR PATHS THAT LED

1142
00:43:39,984 --> 00:43:41,986
YOU DOWN, I'M REALLY CURIOUS TO

1143
00:43:42,052 --> 00:43:42,186
HEAR.

1144
00:43:42,252 --> 00:43:43,954
>> THAT'S A REALLY GREAT

1145
00:43:44,021 --> 00:43:44,688
QUESTION.

1146
00:43:44,755 --> 00:43:44,955
THANK YOU.

1147
00:43:45,022 --> 00:43:46,790
AND IT IS IN FACT ONE OF MY

1148
00:43:46,857 --> 00:43:49,326
FAVORITE THINGS TO TALK ABOUT.

1149
00:43:49,393 --> 00:43:58,702
SO, ONE OF THE THINGS THAT I DO

1150
00:43:58,769 --> 00:44:00,371
IN THE WRITING PROCESS IS I LIKE

1151
00:44:00,437 --> 00:44:08,445
TO TRY EXPLAIN  MYSELF TO KIDS.

1152
00:44:08,512 --> 00:44:10,381
MY SON USED TO BE A GOOD READER

1153
00:44:10,447 --> 00:44:11,215
OF MY WORK.

1154
00:44:11,281 --> 00:44:13,951
STILL IS BUT HE'S OLDER NOW.

1155
00:44:14,018 --> 00:44:14,918
DIFFERENT HAVING AN 11-YEAR-OLD

1156
00:44:14,985 --> 00:44:17,688
READ IT VERSUS AN OLDER TEEN.

1157
00:44:17,755 --> 00:44:22,393
AND I GOT THIS IDEA FROM JAY

1158
00:44:22,459 --> 00:44:29,066
HAMILTON WHO WROTE "DEMOCRACY

1159
00:44:29,133 --> 00:44:29,433
DETECTIVE."

1160
00:44:29,500 --> 00:44:33,037
HE HIRES TEENS TO BE FIRST

1161
00:44:33,103 --> 00:44:33,370
READERS.

1162
00:44:33,437 --> 00:44:35,105
TEENAGERS WILL TELL WHAT YOU IS

1163
00:44:35,172 --> 00:44:37,408
CONFUSING AND WHAT IS BORING.

1164
00:44:37,474 --> 00:44:40,577
LIKE THEY DON'T HAVE A FILTER.

1165
00:44:40,644 --> 00:44:41,945
SO THAT'S REALLY HELPFUL.

1166
00:44:42,012 --> 00:44:47,251
I ALSO TRY AND THINK ABOUT HOW I

1167
00:44:47,317 --> 00:44:49,353
WOULD EXPLAIN SOMETHING IN AN

1168
00:44:49,420 --> 00:44:53,490
INTERESTING WAY AT A COCKTAIL

1169
00:44:53,557 --> 00:44:54,925
PARTY BECAUSE I BORED A LOT OF

1170
00:44:54,992 --> 00:44:56,226
PEOPLE AT COCKTAIL PARTIES OVER

1171
00:44:56,293 --> 00:44:58,529
THE YEARS, AND I TRY NOT TO DO

1172
00:44:58,595 --> 00:44:58,695
IT.

1173
00:44:58,762 --> 00:45:00,631
BUT THAT'S ACTUALLY A REALLY

1174
00:45:00,697 --> 00:45:02,132
GOOD WAY OF GAUGING HOW

1175
00:45:02,199 --> 00:45:05,069
INTERESTING SOMETHING IS.

1176
00:45:05,135 --> 00:45:06,403
SO IN USER INTERFACE DESIGN,

1177
00:45:06,470 --> 00:45:08,172
PEOPLE TALK ABOUT HALLWAY USER

1178
00:45:08,238 --> 00:45:11,675
TESTING, WHERE YOU GO OUT IN THE

1179
00:45:11,742 --> 00:45:13,077
HALLWAY AND FIND SOMEBODY, HEY,

1180
00:45:13,143 --> 00:45:15,179
LOOK AT THIS, USER TEST IT ON

1181
00:45:15,245 --> 00:45:15,379
THEM.

1182
00:45:15,446 --> 00:45:17,114
SO I DO THE SAME THING WITH MY

1183
00:45:17,181 --> 00:45:17,347
WRITING.

1184
00:45:17,414 --> 00:45:19,416
I GO TO PARTIES.

1185
00:45:19,483 --> 00:45:20,851
I TELL STORIES.

1186
00:45:20,918 --> 00:45:23,253
AND IF IT HITS, I'M LIKE, ALL

1187
00:45:23,320 --> 00:45:24,455
RIGHT, THIS IS GOING TO BE

1188
00:45:24,521 --> 00:45:25,656
INTERESTING ENOUGH TO WRITE

1189
00:45:25,722 --> 00:45:26,523
ABOUT.

1190
00:45:26,590 --> 00:45:28,959
IF IT DOESN'T, ALL RIGHT, I JUST

1191
00:45:29,026 --> 00:45:30,461
BORED SOMEBODY ELSE.

1192
00:45:30,527 --> 00:45:33,430
THE STAKES ARE VERY LOW.

1193
00:45:33,497 --> 00:45:33,864
>> GREAT.

1194
00:45:33,931 --> 00:45:34,198
THANK YOU.

1195
00:45:34,264 --> 00:45:36,700
A QUESTION FROM THE AUDIENCE,

1196
00:45:36,767 --> 00:45:39,670
LIVE AUDIENCE.

1197
00:45:39,736 --> 00:45:42,172
>> THANKS AGAIN.

1198
00:45:42,239 --> 00:45:45,008
STEVE SHERRY, ACTING DIRECTOR AT

1199
00:45:45,075 --> 00:45:46,777
THE LIBRARY.

1200
00:45:46,844 --> 00:45:49,012
REMARKS AROUND HARM AND

1201
00:45:49,079 --> 00:45:51,181
ACCOUNTABILITY RESONATE WITH ME.

1202
00:45:51,248 --> 00:45:53,350
THE ROLE OF NLM, INSTITUTIONS

1203
00:45:53,417 --> 00:45:54,985
TRYING TO PRODUCE PUBLIC GOOD,

1204
00:45:55,052 --> 00:45:57,855
PREVENTING HARM.

1205
00:45:57,921 --> 00:45:58,822
MY QUESTIONS ARE THREE.

1206
00:45:58,889 --> 00:46:01,091
FIRST IN YOUR EXAMPLE ABOUT THE

1207
00:46:01,158 --> 00:46:04,428
FOUR DIMENSIONAL ASSESSMENT OF

1208
00:46:04,495 --> 00:46:06,730
THE INFORMATION RISKS AND

1209
00:46:06,797 --> 00:46:08,265
ELECTION RESULTS, YOU MENTIONED

1210
00:46:08,332 --> 00:46:11,568
HARM AS THE SECOND DIMENSION.

1211
00:46:11,635 --> 00:46:14,204
WHAT IS THE HARM THAT YOU'RE

1212
00:46:14,271 --> 00:46:15,939
SCORING ON WITH ELECTION

1213
00:46:16,006 --> 00:46:17,241
INFORMATION, PERSONAL HARM?

1214
00:46:17,307 --> 00:46:20,077
I'M CURIOUS WHAT HARM IS IN THAT

1215
00:46:20,144 --> 00:46:21,111
CONTEXT?

1216
00:46:21,178 --> 00:46:23,714
SECOND QUESTION IS ABOUT

1217
00:46:23,780 --> 00:46:24,448
ACCOUNTABILITY, ALGORITHMIC

1218
00:46:24,515 --> 00:46:25,983
ACCOUNTABILITY, IS THERE A --

1219
00:46:26,049 --> 00:46:27,184
>> I'M GOING TO FORGET IF YOU

1220
00:46:27,251 --> 00:46:28,619
GIVE ME ALL THREE AT ONCE.

1221
00:46:28,685 --> 00:46:29,820
LET ME DO ONE AT A TIME.

1222
00:46:29,887 --> 00:46:33,223
I WANT TO GET TO ALL OF THEM.

1223
00:46:33,290 --> 00:46:34,324
SO, HARM AROUND ELECTIONS, ONE

1224
00:46:34,391 --> 00:46:36,627
OF THE THINGS THAT WAS REALLY

1225
00:46:36,693 --> 00:46:39,563
SMART ABOUT HOW THIS EXPERIMENT

1226
00:46:39,630 --> 00:46:42,299
WAS CONSTRUCTED IS THAT IT WAS

1227
00:46:42,366 --> 00:46:44,735
NOT ABOUT ALL A.I., EVERYWHERE,

1228
00:46:44,801 --> 00:46:48,005
THROUGHOUT TIME ON EVERY

1229
00:46:48,071 --> 00:46:48,972
SUBJECT.

1230
00:46:49,039 --> 00:46:53,844
IT WAS U.S. RESIDENTIAL ELECTION

1231
00:46:53,911 --> 00:46:54,678
2024.

1232
00:46:54,745 --> 00:47:05,055
AND IT WAS PEOPLE WHO ARE

1233
00:47:05,122 --> 00:47:07,791
WORKING IN LOCAL, STATE LEVEL

1234
00:47:07,858 --> 00:47:08,292
ELECTION ADMINISTRATION.

1235
00:47:08,358 --> 00:47:11,094
PROMPTS WERE THINGS LIKE WHERE

1236
00:47:11,161 --> 00:47:13,163
CAN I VOTE, IN NEVADA?

1237
00:47:13,230 --> 00:47:17,234
AND THERE ARE ALL THESE

1238
00:47:17,301 --> 00:47:19,836
DIFFERENT LOCAL RULES ABOUT

1239
00:47:19,903 --> 00:47:20,604
SAME-DAY REGISTRATION, YOU KNOW,

1240
00:47:20,671 --> 00:47:23,106
IF YOU NEED ID, WHAT KIND OF ID

1241
00:47:23,173 --> 00:47:28,278
YOU NEED, CAN YOU VOTE BY TEXT?

1242
00:47:28,345 --> 00:47:32,216
THE LLMs WILL TELL YOU THAT

1243
00:47:32,282 --> 00:47:34,218
YOU CAN VOTE VIA TEXT, YOU

1244
00:47:34,284 --> 00:47:35,052
CANNOT VOTE VIA TEXT.

1245
00:47:35,118 --> 00:47:38,155
SO ONE OF THE THINGS THAT

1246
00:47:38,222 --> 00:47:39,790
HAPPENED AT THE TESTING EVENT IS

1247
00:47:39,856 --> 00:47:41,992
WE TALKED ABOUT, OKAY, WHAT ARE

1248
00:47:42,059 --> 00:47:44,828
POTENTIAL HARMS THAT COULD

1249
00:47:44,895 --> 00:47:46,997
RESULT FROM THIS PARTICULAR

1250
00:47:47,064 --> 00:47:48,332
ANSWER TO THIS PARTICULAR

1251
00:47:48,398 --> 00:47:49,199
QUESTION, RIGHT?

1252
00:47:49,266 --> 00:47:53,904
SO, I DO NOT HAVE THE -- THERE'S

1253
00:47:53,971 --> 00:47:57,407
A GET THE DATA LINK HERE AND SO

1254
00:47:57,474 --> 00:48:01,979
YOU CAN DIG INTO THE PROMPTS AND

1255
00:48:02,045 --> 00:48:02,980
RESPONSES, AND THEY PROBABLY

1256
00:48:03,046 --> 00:48:04,348
PUBLISHED THE RATINGS TOO SO YOU

1257
00:48:04,414 --> 00:48:06,383
WOULD BE ABLE TO SEE, OKAY,

1258
00:48:06,450 --> 00:48:09,119
WHICH ONES WERE RATED HARMFUL BY

1259
00:48:09,186 --> 00:48:12,289
THE SPECIFIC PEOPLE IN THE ROOM

1260
00:48:12,356 --> 00:48:13,423
AROUND THIS SPECIFIC TOPIC.

1261
00:48:13,490 --> 00:48:15,592
>> SO IN A GENERAL SENSE I'M

1262
00:48:15,659 --> 00:48:18,428
SENSING THEN THE HARM IS

1263
00:48:18,495 --> 00:48:20,163
DISENFRANCHISING TO EXERCISE

1264
00:48:20,230 --> 00:48:23,467
YOUR RIGHT TO VOTE, NOT HARM IN

1265
00:48:23,533 --> 00:48:26,737
AN INJURIOUS SENSE, IT'S A CIVIL

1266
00:48:26,803 --> 00:48:27,170
HARM?

1267
00:48:27,237 --> 00:48:27,437
>> YES.

1268
00:48:27,504 --> 00:48:28,405
>> THAT MAKES SENSE.

1269
00:48:28,472 --> 00:48:31,708
>> THE OTHER THING THAT'S REALLY

1270
00:48:31,775 --> 00:48:34,544
IMPORTANT IS THAT THIS IS AN

1271
00:48:34,611 --> 00:48:36,613
EXPERIMENT THAT GETS US CLOSER

1272
00:48:36,680 --> 00:48:39,349
TO EVALUATING A.I. IN CONTEXT.

1273
00:48:39,416 --> 00:48:45,589
SO FOR A VERY LONG TIME,

1274
00:48:45,656 --> 00:48:48,659
COMPUTING IN GENERAL, WE TRY TO

1275
00:48:48,725 --> 00:48:49,526
MAKE THESE GENERALIZABLE

1276
00:48:49,593 --> 00:48:52,496
SYSTEMS, AND THIS COMES FROM THE

1277
00:48:52,562 --> 00:48:53,664
INTELLECTUAL HERITAGE OF

1278
00:48:53,730 --> 00:48:56,400
MATHEMATICS WHERE IN MATH YOU'RE

1279
00:48:56,466 --> 00:48:57,267
TRYING TO DERIVE THEOREMS THAT

1280
00:48:57,334 --> 00:49:01,071
ARE GOING TO WORK EVERYWHERE,

1281
00:49:01,138 --> 00:49:02,706
LIKE THE PYTHAGOREAN THEOREM

1282
00:49:02,773 --> 00:49:04,474
NEVER STOPS BEING TRUE, THAT'S

1283
00:49:04,541 --> 00:49:07,210
AMAZING, BUT THAT'S THE PURSUIT

1284
00:49:07,277 --> 00:49:10,047
OF MATHEMATICS, UNIVERSAL

1285
00:49:10,113 --> 00:49:10,814
TRUTHS.

1286
00:49:10,881 --> 00:49:16,620
WELL, A.I. IS MATH BUT A.I.

1287
00:49:16,687 --> 00:49:18,155
SYSTEMS CAN'T MAKE UNIVERSAL

1288
00:49:18,221 --> 00:49:20,457
TRUTHS WHEN IT COMES TO

1289
00:49:20,524 --> 00:49:22,259
SOMETHING LIKE LANGUAGE OR

1290
00:49:22,326 --> 00:49:24,394
ELECTION INFORMATION OR, YOU

1291
00:49:24,461 --> 00:49:25,862
KNOW, MEDICAL DIAGNOSIS.

1292
00:49:25,929 --> 00:49:28,165
LIKE IT'S JUST NOT POSSIBLE

1293
00:49:28,231 --> 00:49:31,034
BECAUSE THERE'S SO MUCH

1294
00:49:31,101 --> 00:49:32,502
VARIATION IN THE WORLD.

1295
00:49:32,569 --> 00:49:34,037
YOU KNOW, LIKE I'M MUCH MORE

1296
00:49:34,104 --> 00:49:38,208
EXCITED ABOUT USING A.I. FOR

1297
00:49:38,275 --> 00:49:39,176
LIKE RESEARCH INTO PROTEIN

1298
00:49:39,242 --> 00:49:43,580
FOLDING THAN I AM ABOUT USING

1299
00:49:43,647 --> 00:49:46,083
A.I. FOR DIAGNOSING COMMON SKIN

1300
00:49:46,149 --> 00:49:46,616
CONDITIONS.

1301
00:49:46,683 --> 00:49:49,586
>> PERFECT INTRO TO MY SECOND

1302
00:49:49,653 --> 00:49:50,721
QUESTION, ALGORITHMIC

1303
00:49:50,787 --> 00:49:51,788
ACCOUNTABILITY OR AUDITING, YOU

1304
00:49:51,855 --> 00:49:54,758
HAD A TERM FOR THAT, IS THAT

1305
00:49:54,825 --> 00:49:58,729
SOMETHING -- BECAUSE OF THIS

1306
00:49:58,795 --> 00:49:59,463
CONDITIONAL CONTEXT OR

1307
00:49:59,529 --> 00:50:03,066
ALGORITHMS OPERATE IS THERE A

1308
00:50:03,133 --> 00:50:04,935
BURDEN ON ALGORITHM CREATORS TO

1309
00:50:05,001 --> 00:50:06,269
DO THE AUDITING AND OFFER UP

1310
00:50:06,336 --> 00:50:14,478
SOME MEASURE OF PERFORMANCE LIKE

1311
00:50:14,544 --> 00:50:15,445
THE RECEIVERRER OPERATOR YOU

1312
00:50:15,512 --> 00:50:16,980
SHOWED AND WHAT STANDARDS OR

1313
00:50:17,047 --> 00:50:17,247
PRACTICES?

1314
00:50:17,314 --> 00:50:18,382
I'M NOT FAMILIAR IN NIST.

1315
00:50:18,448 --> 00:50:21,351
THEY JUST CHALLENGE YOU TO MAKE

1316
00:50:21,418 --> 00:50:23,320
YOUR OWN ASSESSMENTS BUT MAYBE

1317
00:50:23,387 --> 00:50:24,955
FDA DOES IN MEDICAL DEVICES, I

1318
00:50:25,021 --> 00:50:27,624
DON'T KNOW.

1319
00:50:27,691 --> 00:50:29,559
>> THERE DEFINITELY ISN'T ANY.

1320
00:50:29,626 --> 00:50:32,095
THE STANDARDS ARE EVOLVING AS WE

1321
00:50:32,162 --> 00:50:32,295
SPEAK.

1322
00:50:32,362 --> 00:50:35,599
I JUST SAW SOMETHING SOMETHING

1323
00:50:35,665 --> 00:50:38,769
ON SOCIAL MEDIA TODAY A NEW

1324
00:50:38,835 --> 00:50:40,537
PAPER BY DEBRA RAJI, AN EXPERT

1325
00:50:40,604 --> 00:50:42,038
IN ALGORITHMIC AUDITING.

1326
00:50:42,105 --> 00:50:44,241
SHE AND HER COLLABORATORS HAVE A

1327
00:50:44,307 --> 00:50:49,312
NEW PAPER OUT ABOUT GAPS IN

1328
00:50:49,379 --> 00:50:50,947
TOOLING IN ALGORITHMIC AUDITING.

1329
00:50:51,014 --> 00:50:54,117
SO, IT'S A REALLY NEW FIELD.

1330
00:50:54,184 --> 00:50:57,053
AND SO WE ARE DEVELOPING

1331
00:50:57,120 --> 00:50:57,454
STANDARDS.

1332
00:50:57,521 --> 00:51:01,458
ONE OF THE THINGS THAT I DID FOR

1333
00:51:01,525 --> 00:51:04,060
THE BOOK WAS I SOMETIMES WRITE

1334
00:51:04,127 --> 00:51:05,929
SOFTWARE IN ORDER TO KIND OF

1335
00:51:05,996 --> 00:51:08,432
ANSWER A JOURNALISTIC QUESTION

1336
00:51:08,498 --> 00:51:09,433
OR LIKE DEMONSTRATE SOMETHING

1337
00:51:09,499 --> 00:51:11,067
ABOUT THE PROCESS OF MAKING

1338
00:51:11,134 --> 00:51:11,401
SOFTWARE.

1339
00:51:11,468 --> 00:51:15,005
AND ONE OF THE THINGS THAT I

1340
00:51:15,071 --> 00:51:17,240
MADE OVER THE COURSE OF WRITING

1341
00:51:17,307 --> 00:51:20,177
THE BOOK WAS I MADE A PLATFORM

1342
00:51:20,243 --> 00:51:21,478
FOR ALGORITHMIC AUDITING.

1343
00:51:21,545 --> 00:51:23,447
SO, IF YOU'RE FAMILIAR WITH

1344
00:51:23,513 --> 00:51:25,615
KATHY O'NEIL'S BOOK, WEAPONS OF

1345
00:51:25,682 --> 00:51:27,184
MASS DESTRUCTION, IT WAS ONE OF

1346
00:51:27,250 --> 00:51:30,053
THE BOOKS THAT KICKED OFF THE

1347
00:51:30,120 --> 00:51:31,555
ENTIRE ALGORITHMIC

1348
00:51:31,621 --> 00:51:32,556
ACCOUNTABILITY CONVERSATION.

1349
00:51:32,622 --> 00:51:35,192
KATHY HAS AN ALGORITHMIC

1350
00:51:35,258 --> 00:51:36,526
AUDITING CONSULTING FIRM, CALLED

1351
00:51:36,593 --> 00:51:37,127
ORCA.

1352
00:51:37,194 --> 00:51:39,596
AND SO I COLLABORATED ORCA TO

1353
00:51:39,663 --> 00:51:42,532
BUILD A PLATFORM FOR ALGORITHMIC

1354
00:51:42,599 --> 00:51:46,136
AUDITING WHERE YOU COULD FEED

1355
00:51:46,203 --> 00:51:51,174
IN, YOU KNOW, FEED IN DATA AND

1356
00:51:51,241 --> 00:51:52,108
LIKE BASICALLY GET EVALUATED,

1357
00:51:52,175 --> 00:51:55,345
GIVE YOU A LITTLE REPORT ABOUT,

1358
00:51:55,412 --> 00:51:56,446
YOU KNOW, POTENTIAL BIASES.

1359
00:51:56,513 --> 00:52:02,319
SO, THIS IS THE KIND OF THING WE

1360
00:52:02,385 --> 00:52:03,220
DO NEED.

1361
00:52:03,286 --> 00:52:07,057
IT EXISTS IN VARIOUS FORMS BUT

1362
00:52:07,123 --> 00:52:09,793
YOU CAN'T RIGHT NOW JUST GO AND

1363
00:52:09,860 --> 00:52:12,963
BUY SOMETHING OFF THE SHELF.

1364
00:52:13,029 --> 00:52:14,164
IT'S MOSTLY BESPOKE AND THERE

1365
00:52:14,231 --> 00:52:22,672
ARE A BUNCH OF OPEN SOURCE

1366
00:52:22,739 --> 00:52:23,373
SOLUTIONS OUT THERE.

1367
00:52:23,440 --> 00:52:25,108
WE DO NEED TO GET IT TO THE

1368
00:52:25,175 --> 00:52:26,510
LEVEL OF PLATFORMS AT SOME POINT

1369
00:52:26,576 --> 00:52:28,144
BUT WE'RE NOT THERE YET.

1370
00:52:28,211 --> 00:52:30,580
>> THAT WAS MY LAST QUESTION ON

1371
00:52:30,647 --> 00:52:33,116
THAT TOPIC, NIH IS VERY

1372
00:52:33,183 --> 00:52:33,717
INTERESTED IN SOCIAL

1373
00:52:33,783 --> 00:52:36,019
DETERMINANTS OF HEALTH, RIGHT?

1374
00:52:36,086 --> 00:52:37,988
AS WE LOOK AT EQUITY IN ALL OF

1375
00:52:38,054 --> 00:52:40,524
THE ENVIRONMENTAL AND HISTORICAL

1376
00:52:40,590 --> 00:52:42,592
CONTEXTS THAT CREATE RACIAL BIAS

1377
00:52:42,659 --> 00:52:45,662
IN MEDICINE, IS THERE A PLACE TO

1378
00:52:45,729 --> 00:52:48,064
CONSIDER THESE ALGORITHMS AS

1379
00:52:48,131 --> 00:52:49,833
ANOTHER SOCIAL DETERMINANT IN

1380
00:52:49,900 --> 00:52:50,767
THAT SPECTRUM OF EXTERNAL

1381
00:52:50,834 --> 00:52:51,601
FACTORS TO HELP?

1382
00:52:51,668 --> 00:52:54,204
LIKE YOU'RE SAYING IT'S BECOMING

1383
00:52:54,271 --> 00:52:55,405
AN EMERGENT EMBEDDED PROPERTY IN

1384
00:52:55,472 --> 00:52:57,307
THE HEALTH CARE SYSTEMS SO DO WE

1385
00:52:57,374 --> 00:53:00,977
NEED TO PAY ATTENTION TO THE

1386
00:53:01,044 --> 00:53:02,846
LIBRARY TO TRY AND IDENTIFY AND

1387
00:53:02,913 --> 00:53:03,880
ORGANIZE INFORMATION AROUND

1388
00:53:03,947 --> 00:53:05,949
ALGORITHMS AND THEIR BIAS?

1389
00:53:06,016 --> 00:53:08,885
>> OH, ABSOLUTELY, YES.

1390
00:53:08,952 --> 00:53:09,386
YEAH.

1391
00:53:09,452 --> 00:53:11,154
THERE IS -- SO EMMA PEARSON IS

1392
00:53:11,221 --> 00:53:14,124
ONE OF THE AUTHORS ON THIS

1393
00:53:14,190 --> 00:53:16,326
PAPER, IS DOING SOME INTERESTING

1394
00:53:16,393 --> 00:53:20,664
WORK AT CORNELL.

1395
00:53:20,730 --> 00:53:24,100
I ALSO ALWAYS RECOMMEND READING

1396
00:53:24,167 --> 00:53:29,573
ROXANNEA -- I'M GOING TO MESS UP

1397
00:53:29,639 --> 00:53:30,774
THE NAME.

1398
00:53:30,840 --> 00:53:34,210
DANISHOW, WHO IS DOING REALLY

1399
00:53:34,277 --> 00:53:41,551
INTERESTING WORK IN DERMATOLOGY

1400
00:53:41,618 --> 00:53:43,853
AND  ML STUFF.

1401
00:53:43,920 --> 00:53:45,922
FOR ME IT'S ABOUT KEEPING TRACK

1402
00:53:45,989 --> 00:53:48,525
OF WHO IS DOING INTERESTING WORK

1403
00:53:48,592 --> 00:53:50,560
AND KIND OF KEEPING UP WITH THE

1404
00:53:50,627 --> 00:53:51,628
LATEST THINKING BECAUSE I THINK

1405
00:53:51,695 --> 00:53:54,931
THE WAY THAT WE THINK ABOUT

1406
00:53:54,998 --> 00:53:56,800
THESE ISSUES IS CHANGE, RIGHT?

1407
00:53:56,866 --> 00:54:02,739
I MEAN IF WE EVEN THINK ABOUT

1408
00:54:02,806 --> 00:54:04,274
SOMETHING LIKE IN PEDIATRICS,

1409
00:54:04,341 --> 00:54:06,109
ONE OF THE INITIATIVES A FEW

1410
00:54:06,176 --> 00:54:10,480
YEARS AGO, OH, LET'S GET

1411
00:54:10,547 --> 00:54:12,248
BABIES -- LET'S GET LITTLE KIDS

1412
00:54:12,315 --> 00:54:14,417
DENTAL HEALTH EVALUATED BECAUSE

1413
00:54:14,484 --> 00:54:15,952
THEY REALIZE, OKAY, IF KIDS GO

1414
00:54:16,019 --> 00:54:18,688
TO THE DENTIST THAT'S LIKE AN

1415
00:54:18,755 --> 00:54:20,890
INDICATOR THEY ARE BEING WELL

1416
00:54:20,957 --> 00:54:22,092
LOOKED AFTER MEDICALLY, IT

1417
00:54:22,158 --> 00:54:23,627
WASN'T ABOUT THE DENTIST PER SE,

1418
00:54:23,693 --> 00:54:29,866
IT WAS ABOUT THE DENTISTS BEING

1419
00:54:29,933 --> 00:54:32,369
A SIGNIFIER OF WELLNESS AND SO

1420
00:54:32,435 --> 00:54:34,537
HOW DO YOU HIGHLIGHT THAT?

1421
00:54:34,604 --> 00:54:40,877
AND HOW DO YOU AS ALGORITHMS OF

1422
00:54:40,944 --> 00:54:42,445
ALGORITHMS FIND THINGS LIKE IF A

1423
00:54:42,512 --> 00:54:44,314
CHILD HAS THREE CAVITIES BEFORE

1424
00:54:44,381 --> 00:54:46,349
AGE 4 THEY MIGHT NEED

1425
00:54:46,416 --> 00:54:47,050
NUTRITIONAL COUNSELING, HOW DO

1426
00:54:47,117 --> 00:54:49,753
YOU APPROACH THAT FROM A LIBRARY

1427
00:54:49,819 --> 00:54:50,220
PERSPECTIVE?

1428
00:54:50,286 --> 00:54:51,721
I THINK WE'RE STILL TRYING TO

1429
00:54:51,788 --> 00:54:52,989
FIGURE THAT OUT.

1430
00:54:53,056 --> 00:54:53,523
>> THANK YOU.

1431
00:54:53,590 --> 00:54:55,225
>> SO WE'LL TAKE A QUESTION FROM

1432
00:54:55,291 --> 00:54:58,194
ONLINE AND IT LOOKS LIKE WE HAVE

1433
00:54:58,261 --> 00:55:00,497
A QUESTION FROM THE ROOM.

1434
00:55:00,563 --> 00:55:02,032
SO, ONE VIRTUAL ATTENDEE NOTED

1435
00:55:02,098 --> 00:55:05,068
THAT MANY OF THE DISCUSSIONS

1436
00:55:05,135 --> 00:55:07,070
THAT THEY HEARD AROUND DIVERSITY

1437
00:55:07,137 --> 00:55:09,839
AND A.I. HAVE FOCUSED ON RACE,

1438
00:55:09,906 --> 00:55:12,342
GENDER, AND NATIONAL ORIGIN WITH

1439
00:55:12,409 --> 00:55:13,877
NOT AS MUCH DISCUSSION OF

1440
00:55:13,943 --> 00:55:16,312
COMMUNITIES OF DISABILITIES.

1441
00:55:16,379 --> 00:55:18,248
THIS HAS BEEN A FOCUS OF YOUR

1442
00:55:18,314 --> 00:55:19,916
RESEARCH FOR MORE THAN A GLITCH

1443
00:55:19,983 --> 00:55:23,253
SO WAS WOUND ERIC IF --

1444
00:55:23,319 --> 00:55:26,489
WONDERING IF YOU COULD COMMENT

1445
00:55:26,556 --> 00:55:28,725
ON YOUR LEARNINGS IN THIS SPACE,

1446
00:55:28,792 --> 00:55:30,560
PERHAPS A STORY ABOUT SUCCESS IN

1447
00:55:30,627 --> 00:55:34,431
CO-CREATION AND I'VE HEARD YOU

1448
00:55:34,497 --> 00:55:37,934
TALK ABOUT -- OH GOSH.

1449
00:55:38,001 --> 00:55:38,735
>> CURB CUTS.

1450
00:55:38,802 --> 00:55:39,703
>> CURB CUTS.

1451
00:55:39,769 --> 00:55:41,137
HOW DOES THAT WORK?

1452
00:55:41,204 --> 00:55:43,940
>> LET ME TALK ABOUT DISABILITY

1453
00:55:44,007 --> 00:55:45,075
DONGELS, THOSE ARE INTERESTING.

1454
00:55:45,141 --> 00:55:47,143
FIRST OF ALL, MORE THAN A GLITCH

1455
00:55:47,210 --> 00:55:49,979
IS ABOUT CONFRONTING RACE,

1456
00:55:50,046 --> 00:55:51,047
GENDER, AND ABILITY BIAS IN

1457
00:55:51,114 --> 00:55:51,514
TECH.

1458
00:55:51,581 --> 00:55:52,682
ORIGINALLY IT WAS JUST ABOUT

1459
00:55:52,749 --> 00:55:53,817
RACE AND GENDER.

1460
00:55:53,883 --> 00:55:58,154
I WAS LIKE, WAIT A MINUTE,

1461
00:55:58,221 --> 00:55:59,255
THERE'S NOT ENOUGH DISCOURSE

1462
00:55:59,322 --> 00:56:00,356
ABOUT DISABILITY.

1463
00:56:00,423 --> 00:56:02,158
THEN I STARTED WRITING.

1464
00:56:02,225 --> 00:56:04,294
OH, WAIT, I NEED TO LEARN.

1465
00:56:04,360 --> 00:56:05,962
THAT WAS THE TOPIC THAT I HAD TO

1466
00:56:06,029 --> 00:56:07,797
LEARN THE MOST ABOUT IN ORDER TO

1467
00:56:07,864 --> 00:56:09,332
WRITE THIS TOOK.

1468
00:56:09,399 --> 00:56:11,401
I'M GRATEFUL TO THE FOLKS WHO

1469
00:56:11,468 --> 00:56:14,471
TOOK THE TIME TO EXPLAIN THINGS

1470
00:56:14,537 --> 00:56:15,672
TO ME.

1471
00:56:15,739 --> 00:56:16,973
THINKERS AND CREATORS WHO I HAD

1472
00:56:17,040 --> 00:56:20,376
AN OPPORTUNITY TO LEARN FROM.

1473
00:56:20,443 --> 00:56:21,611
AND SOMETHING THAT WAS REALLY

1474
00:56:21,678 --> 00:56:24,714
POWERFUL TO ME WAS THE CONCEPT

1475
00:56:24,781 --> 00:56:29,953
OF A DISABILITY  DONGLE,

1476
00:56:30,019 --> 00:56:32,122
SOMETHING A DESIGNER WILL CREATE

1477
00:56:32,188 --> 00:56:34,758
BECAUSE THEY THINK IT'S GOING TO

1478
00:56:34,824 --> 00:56:39,262
BE TRANSFORMATIVE FOR PEOPLE

1479
00:56:39,329 --> 00:56:40,430
WITH DISABILITIES.

1480
00:56:40,497 --> 00:56:43,500
A GOOD EXAMPLE IS THE

1481
00:56:43,566 --> 00:56:44,434
STAIR-CLIMBING WHEELCHAIR.

1482
00:56:44,501 --> 00:56:47,670
TONS OF NEW THINGS HAVE BEEN

1483
00:56:47,737 --> 00:56:51,040
DEVELOPED, THEY ALL HAVE THESE

1484
00:56:51,107 --> 00:56:51,808
WHACKY-LOOKING ARCHITECTURES,

1485
00:56:51,875 --> 00:56:55,145
LIKE THEY ARE NOT WHEELS, THEY

1486
00:56:55,211 --> 00:56:58,314
ARE LIKE CROSSES, COULD CLIMB

1487
00:56:58,381 --> 00:56:59,849
THE STAIRS, CATERPILLAR WHEELS

1488
00:56:59,916 --> 00:57:00,717
ON THEM.

1489
00:57:00,784 --> 00:57:02,786
AND WHEN YOU ASK SOMEBODY IN A

1490
00:57:02,852 --> 00:57:06,055
WHEELCHAIR, OH, DO YOU WANT THIS

1491
00:57:06,122 --> 00:57:06,923
STAIR-CLIMBING WHEELCHAIR, LOOK,

1492
00:57:06,990 --> 00:57:08,158
I INVENTED THIS, THEY ARE LIKE,

1493
00:57:08,224 --> 00:57:10,994
NO, I DON'T WANT THAT.

1494
00:57:11,060 --> 00:57:11,861
THAT LOOKS TERRIFYING.

1495
00:57:11,928 --> 00:57:16,699
I JUST WANT A RAMP.

1496
00:57:16,766 --> 00:57:18,334
YOU REALIZE YOU DON'T NEED TO

1497
00:57:18,401 --> 00:57:20,670
OVERENGINEER THE SOLUTION.

1498
00:57:20,737 --> 00:57:22,372
AND ALSO YOU SHOULD HAVE JUST

1499
00:57:22,438 --> 00:57:24,574
ASKED THE PERSON USING THE

1500
00:57:24,641 --> 00:57:25,341
WHEELCHAIR, LIKE LISTEN, WHAT'S

1501
00:57:25,408 --> 00:57:29,045
GOING TO MAKE YOUR LIFE EASIER?

1502
00:57:29,112 --> 00:57:31,114
INSTEAD OF ASSUMING THAT THIS

1503
00:57:31,181 --> 00:57:34,450
THING THAT YOU CAME UP WITH IS

1504
00:57:34,517 --> 00:57:37,787
GOING TO HELP.

1505
00:57:37,854 --> 00:57:40,657
SO, THE IDEA IS THAT WHEN WE'RE

1506
00:57:40,723 --> 00:57:43,526
DESIGNING TECHNOLOGIES WE NEED

1507
00:57:43,593 --> 00:57:45,061
TO CONSULT WITH AFFECTED

1508
00:57:45,128 --> 00:57:46,362
COMMUNITIES AND BUILD OUT OF THE

1509
00:57:46,429 --> 00:57:47,997
NEEDS OF THE AFFECTED

1510
00:57:48,064 --> 00:57:50,066
COMMUNITIES AS OPPOSED TO THIS

1511
00:57:50,133 --> 00:57:52,168
KIND OF TOP-DOWN SILICON VALLEY

1512
00:57:52,235 --> 00:58:00,043
APPROACH OF, OH, I AM THE

1513
00:58:00,109 --> 00:58:01,644
DESIGNER, I KNOW EVERYTHING.

1514
00:58:01,711 --> 00:58:03,546
>> I KNOW WE'RE AT TIME.

1515
00:58:03,613 --> 00:58:05,315
I'M HAILEY, A POSTDOC.

1516
00:58:05,381 --> 00:58:06,549
I DO METABOLOMICS RESEARCH

1517
00:58:06,616 --> 00:58:08,017
WHETHER WE TAKE A LOT OF

1518
00:58:08,084 --> 00:58:10,353
CHEMICALS AND TRY TO FIND OUT

1519
00:58:10,420 --> 00:58:11,788
HOW THEY CONTRIBUTE TO A

1520
00:58:11,855 --> 00:58:12,021
DISEASE.

1521
00:58:12,088 --> 00:58:14,090
WE'RE LOOKING AT A LOT OF

1522
00:58:14,157 --> 00:58:16,826
VARIABLES, TRYING TO FIND MORE

1523
00:58:16,893 --> 00:58:19,128
GRANULARITY BEYONDS WHAT WE

1524
00:58:19,195 --> 00:58:20,330
ALREADY KNOW.

1525
00:58:20,396 --> 00:58:21,531
I SAW THAT REFLECTED HERE WHEN

1526
00:58:21,598 --> 00:58:23,066
THINKING ABOUT STUFF LIKE RACE,

1527
00:58:23,132 --> 00:58:25,602
LIKE YOU SAID, IT CAN BE OVERLY

1528
00:58:25,668 --> 00:58:30,440
BROAD OR WITH DISABILITIES,

1529
00:58:30,506 --> 00:58:31,975
MAYBE STUFF WE DON'T UNDERSTAND.

1530
00:58:32,041 --> 00:58:34,477
TO TAKE IT TO THE BIGGER

1531
00:58:34,544 --> 00:58:36,312
PICTURE, NOTICING A PATTERN IN

1532
00:58:36,379 --> 00:58:38,414
YOUR TALK, WHAT WE NEED TO

1533
00:58:38,481 --> 00:58:41,251
EVALUATE ARE INPUTS AND OUTPUTS,

1534
00:58:41,317 --> 00:58:43,019
HOW WE INTERPRET OUR DATA.

1535
00:58:43,086 --> 00:58:45,054
I THINK YOU GAVE TWO GOOD

1536
00:58:45,121 --> 00:58:46,055
EXAMPLES BUT THEY INTRODUCED

1537
00:58:46,122 --> 00:58:48,491
QUESTIONS FOR ME BECAUSE EVEN

1538
00:58:48,558 --> 00:58:51,094
WHEN TALKING ABOUT THAT FOUR

1539
00:58:51,160 --> 00:58:52,428
DIMENSIONAL EVALUATION OF A.I.,

1540
00:58:52,495 --> 00:58:54,731
THE QUESTION THAT IT BROUGHT UP

1541
00:58:54,797 --> 00:58:56,900
IS THESE ARE STILL -- HUMANS

1542
00:58:56,966 --> 00:58:58,768
CREATE THE A.I., HUMANS ARE ALSO

1543
00:58:58,835 --> 00:59:01,170
EVALUATING THE A.I., MADE ME

1544
00:59:01,237 --> 00:59:02,605
WONDER HOW WOULD OUTPUTS CHANGE

1545
00:59:02,672 --> 00:59:04,340
IF THERE WERE DIFFERENT PEOPLE

1546
00:59:04,407 --> 00:59:06,209
DOING THE EVALUATION, HOW WOULD

1547
00:59:06,276 --> 00:59:07,410
OUTPUT CHANGE IF INSTEAD OF

1548
00:59:07,477 --> 00:59:09,145
REPORTING ON A MAJORITY VOTE WE

1549
00:59:09,212 --> 00:59:10,480
USED DIFFERENT KIND OF METRIC.

1550
00:59:10,546 --> 00:59:12,548
SO THESE ARE ALL THE QUESTIONS

1551
00:59:12,615 --> 00:59:15,618
THAT I'M WONDERING IS THAT WHAT

1552
00:59:15,685 --> 00:59:20,089
GOES INTO YOUR ALGORITHMIC

1553
00:59:20,156 --> 00:59:21,090
ACCOUNTABILITY TOOL?

1554
00:59:21,157 --> 00:59:23,393
AND IT SEEMS LIKE YOU'RE

1555
00:59:23,459 --> 00:59:25,728
PROPOSING A HUMAN SOLUTION TO A

1556
00:59:25,795 --> 00:59:27,530
MACHINE PROBLEM, WHICH I THINK

1557
00:59:27,597 --> 00:59:28,531
IS ABSOLUTELY APPROPRIATE

1558
00:59:28,598 --> 00:59:30,233
BECAUSE LIKE YOU SAID, COMPUTERS

1559
00:59:30,300 --> 00:59:33,803
ARE ONLY GOING TO ADDRESS SO

1560
00:59:33,870 --> 00:59:34,570
MANY PROBLEMS.

1561
00:59:34,637 --> 00:59:36,339
SO, IS IT APPROPRIATE FOR THESE

1562
00:59:36,406 --> 00:59:38,441
TOOLS TO CONTINUE BEING

1563
00:59:38,508 --> 00:59:38,708
COMPUTERS?

1564
00:59:38,775 --> 00:59:40,944
AND ALSO THERE ARE A LOT OF

1565
00:59:41,010 --> 00:59:41,878
ELEMENTS TO THIS QUESTION,

1566
00:59:41,945 --> 00:59:43,579
BECAUSE I'M THINKING BOTH ABOUT

1567
00:59:43,646 --> 00:59:44,981
THE INPUTS AND OUTPUTS AT THE

1568
00:59:45,048 --> 00:59:45,982
SAME TIME.

1569
00:59:46,049 --> 00:59:48,017
SO TWO PARTS THEN.

1570
00:59:48,084 --> 00:59:51,220
ONE PART RELATED TO THE RACE OR

1571
00:59:51,287 --> 00:59:52,655
OTHER VARIABLES THAT ARE MAYBE

1572
00:59:52,722 --> 00:59:55,191
TOO LARGE TO ACCOUNT FOR.

1573
00:59:55,258 --> 00:59:57,160
WHAT HAVE YOU SEEN THAT MAYBE

1574
00:59:57,226 --> 00:59:58,227
HAS BEEN HIGHLIGHTED OR

1575
00:59:58,294 --> 01:00:00,396
SOMETHING THAT COULD BE

1576
01:00:00,463 --> 01:00:02,432
EFFECTIVE FOR BREAKING THESE

1577
01:00:02,498 --> 01:00:03,499
DOWN INTO MORE GRANULAR CHUNKS

1578
01:00:03,566 --> 01:00:05,134
LIKE WHAT WE DO WITH

1579
01:00:05,201 --> 01:00:08,004
METABOLOMICS TO TRY TO

1580
01:00:08,071 --> 01:00:08,805
UNDERSTAND MORE GRANULAR

1581
01:00:08,871 --> 01:00:11,808
CHEMISTRY, AND ALSO WHEN IT

1582
01:00:11,874 --> 01:00:13,376
COMES TO THE OUTPUTS, YOU'RE

1583
01:00:13,443 --> 01:00:14,877
EVALUATING THESE THINGS, WHAT

1584
01:00:14,944 --> 01:00:15,878
ARE THE KINDS OF REPRODUCIBLE

1585
01:00:15,945 --> 01:00:17,180
QUESTIONS THAT WE CAN ASK ABOUT

1586
01:00:17,246 --> 01:00:20,049
HOW THE DATA IS BEING

1587
01:00:20,116 --> 01:00:20,783
INTERPRETED?

1588
01:00:20,850 --> 01:00:21,851
LIKE FOR EXAMPLE THEY CAN BE

1589
01:00:21,918 --> 01:00:23,553
LIKE WHO IS IN THE ROOM, BECAUSE

1590
01:00:23,619 --> 01:00:24,854
THAT WAS THE FIRST QUESTION THAT

1591
01:00:24,921 --> 01:00:28,691
CAME UP FOR ME WITH THE A.I.

1592
01:00:28,758 --> 01:00:29,225
EVALUATION.

1593
01:00:29,292 --> 01:00:30,093
>> YEAH.

1594
01:00:30,159 --> 01:00:31,194
>> SO, YEAH.

1595
01:00:31,260 --> 01:00:32,929
>> YES, THAT'S A FANTASTIC

1596
01:00:32,996 --> 01:00:33,196
QUESTION.

1597
01:00:33,262 --> 01:00:34,564
I THINK YOU BASICALLY

1598
01:00:34,630 --> 01:00:35,631
ENCAPSULATED MY ENTIRE TALK IN

1599
01:00:35,698 --> 01:00:37,200
ONE QUESTION.

1600
01:00:37,266 --> 01:00:38,634
THANK YOU FOR THAT.

1601
01:00:38,701 --> 01:00:40,937
I WOULD SAY THAT THE THING THAT

1602
01:00:41,004 --> 01:00:42,672
HELPS ME IS SOMETHING FROM

1603
01:00:42,739 --> 01:00:45,875
SCIENCE AND SOMETHING FROM

1604
01:00:45,942 --> 01:00:46,175
LITERATURE.

1605
01:00:46,242 --> 01:00:49,145
SO, FROM LITERATURE, I THINK A

1606
01:00:49,212 --> 01:00:51,014
LOT ABOUT HUBRIS, RIGHT?

1607
01:00:51,080 --> 01:00:54,417
YOU KNOW, I TOOK A LOT OF

1608
01:00:54,484 --> 01:00:55,518
SHAKESPEARE CLASSES IN COLLEGE.

1609
01:00:55,585 --> 01:00:59,789
YOU KNOW, WHAT YOU LEARN FROM

1610
01:00:59,856 --> 01:01:03,292
SHAKESPEARE IS DON'T GET TOO FAR

1611
01:01:03,359 --> 01:01:04,927
OUT OR YOU'RE GOING TO GET

1612
01:01:04,994 --> 01:01:05,795
KNOCKED DOWN.

1613
01:01:05,862 --> 01:01:09,032
I TRY NOT TO ADD HUBRIS, TRY NOT

1614
01:01:09,098 --> 01:01:11,167
TO SOLVE ALL OF THE PROBLEMS OF

1615
01:01:11,234 --> 01:01:13,469
THE UNIVERSE WITH ONE ALGORITHM.

1616
01:01:13,536 --> 01:01:15,571
AND SO THAT I THINK IS A REALLY

1617
01:01:15,638 --> 01:01:16,639
IMPORTANT APPROACH BECAUSE ONE

1618
01:01:16,706 --> 01:01:19,509
OF THE THINGS WE SEE COMING OUT

1619
01:01:19,575 --> 01:01:21,044
OF SILICON VALLEY, OH, THIS A.I.

1620
01:01:21,110 --> 01:01:28,151
IS GOING TO SOLVE EVERYTHING,

1621
01:01:28,217 --> 01:01:29,118
USEFUL FOR ALL (INDISCERNIBLE).

1622
01:01:29,185 --> 01:01:30,019
MAYBE IT'S NOT.

1623
01:01:30,086 --> 01:01:32,955
THE OTHER THING, I TRY AND

1624
01:01:33,022 --> 01:01:34,524
APPROACH ALGORITHMIC QUESTIONS

1625
01:01:34,590 --> 01:01:36,392
MORE LIKE A SCIENTIST THAN LIKE

1626
01:01:36,459 --> 01:01:39,562
A COMPUTER SCIENTIST.

1627
01:01:39,629 --> 01:01:42,065
SO I LIKE TO TAKE A REALLY SMALL

1628
01:01:42,131 --> 01:01:44,400
PIECE, I LIKE TO GET REALLY

1629
01:01:44,467 --> 01:01:47,637
GRANULAR AND LOOK AT A REALLY

1630
01:01:47,703 --> 01:01:50,073
FINITE CASE IN ORDER TO

1631
01:01:50,139 --> 01:01:51,574
ILLUMINATE THE LARGER UNIVERSE

1632
01:01:51,641 --> 01:01:55,411
AS OPPOSED TO TRYING TO WRITE

1633
01:01:55,478 --> 01:01:57,280
ABOUT EVERYTHING THROUGHOUT

1634
01:01:57,346 --> 01:01:57,647
TIME.

1635
01:01:57,713 --> 01:01:59,115
BECAUSE THAT'S HOW SCIENTIFIC

1636
01:01:59,182 --> 01:01:59,882
KNOWLEDGE EVOLVES, RIGHT?

1637
01:01:59,949 --> 01:02:03,052
AND SO I THINK THAT'S HOW OUR

1638
01:02:03,119 --> 01:02:05,121
KNOWLEDGE IN ALGORITHMIC SPACES

1639
01:02:05,188 --> 01:02:06,456
NEEDS TO EVOLVE TOO.

1640
01:02:06,522 --> 01:02:07,356
>> THANK YOU.

1641
01:02:07,423 --> 01:02:07,657
>> GREAT.

1642
01:02:07,723 --> 01:02:09,192
THANK YOU SO MUCH, EVERYBODY.

1643
01:02:09,258 --> 01:02:10,593
THIS HAS BEEN REALLY WONDERFUL.

1644
01:02:10,660 --> 01:02:10,960
>> THANK YOU.

1645
01:02:11,027 --> 01:02:12,361
THANK YOU SO MUCH.

1646
01:02:12,428 --> 01:02:15,698
A LOT OF COMMENTS COMING IN,

1647
01:02:15,765 --> 01:02:17,266
JUST THANKING YOU FOR A GREAT,

1648
01:02:17,333 --> 01:02:17,934
GREAT TALK.

1649
01:02:18,000 --> 01:02:20,670
PLEASE, ONE MORE ROUND OF

1650
01:02:20,736 --> 01:02:20,937
APPLAUSE.

1651
01:02:21,003
THANK YOU SO MUCH.