1
00:00:04,050 --> 00:00:07,560
- [Jon] AI ethics isn't about
imaginary sci-fi scenarios

2
00:00:07,560 --> 00:00:09,000
like killer robots.

3
00:00:09,000 --> 00:00:11,790
It's about recognizing
real present dangers

4
00:00:11,790 --> 00:00:13,650
of generative AI.

5
00:00:13,650 --> 00:00:14,640
Now, there's a lot of good stuff

6
00:00:14,640 --> 00:00:16,440
that could come from generative AI,

7
00:00:16,440 --> 00:00:20,670
from ways to diagnose cancers,
to discover new drugs,

8
00:00:20,670 --> 00:00:23,220
to find solutions for climate change,

9
00:00:23,220 --> 00:00:24,450
but it's easy to forget

10
00:00:24,450 --> 00:00:26,760
amidst all this potentially good news

11
00:00:26,760 --> 00:00:30,963
about the many potential
harms of this new technology.

12
00:00:32,460 --> 00:00:35,880
The IMPACT RISK acronym offers a mnemonic

13
00:00:35,880 --> 00:00:38,100
so you or your students

14
00:00:38,100 --> 00:00:40,920
can help take these
various harms into account.

15
00:00:40,920 --> 00:00:41,940
My name is Jon Ippolito,

16
00:00:41,940 --> 00:00:44,223
and I'm gonna go through
these one at a time.

17
00:00:45,900 --> 00:00:48,660
I is for Infowar.

18
00:00:48,660 --> 00:00:50,160
AI-powered information

19
00:00:50,160 --> 00:00:52,410
that threatens to undermine democracy,

20
00:00:52,410 --> 00:00:54,810
that includes deep fakes
of high profile figures

21
00:00:54,810 --> 00:00:59,810
like President Biden here
heard in a faked robocall

22
00:01:00,000 --> 00:01:03,600
as part of the 2024
presidential primaries.

23
00:01:03,600 --> 00:01:04,620
- [Joe] Voting this Tuesday

24
00:01:04,620 --> 00:01:06,510
only enables the Republicans in...

25
00:01:06,510 --> 00:01:08,640
- [Jon] It also includes
personalized attacks

26
00:01:08,640 --> 00:01:12,240
on local figures like this faked audio

27
00:01:12,240 --> 00:01:14,640
from a high school principal in Maryland.

28
00:01:14,640 --> 00:01:17,250
- [Principal] "You know, I
seriously don't understand

29
00:01:17,250 --> 00:01:18,810
why I have to constantly put up

30
00:01:18,810 --> 00:01:20,460
with these dumbasses here every day."

31
00:01:20,460 --> 00:01:22,590
- [Jon] The ease of
generating customized code

32
00:01:22,590 --> 00:01:25,350
and text has also made
cyber attacks cheaper

33
00:01:25,350 --> 00:01:26,940
and faster to deploy.

34
00:01:26,940 --> 00:01:29,190
Whether that means writing custom malware

35
00:01:29,190 --> 00:01:31,593
or writing spearfishing emails.

36
00:01:34,110 --> 00:01:36,630
M is for Monopoly.

37
00:01:36,630 --> 00:01:39,660
This one flies under the radar
of a lot of people's mind

38
00:01:39,660 --> 00:01:41,610
when they think about AI harms.

39
00:01:41,610 --> 00:01:43,740
But a handful of Silicon Valley titans

40
00:01:43,740 --> 00:01:45,930
occupy the generative AI market

41
00:01:45,930 --> 00:01:47,700
with Google vying for dominance

42
00:01:47,700 --> 00:01:50,400
against the OpenAI-Microsoft partnership,

43
00:01:50,400 --> 00:01:53,460
Meta prevailing when it
comes to open-source models.

44
00:01:53,460 --> 00:01:55,800
This concentration of
power is even more evident

45
00:01:55,800 --> 00:01:58,470
in the AI chip market
where Nvidia's dominance

46
00:01:58,470 --> 00:02:00,990
leaves competitors scrambling for crumbs.

47
00:02:00,990 --> 00:02:02,970
The massive amount of computation required

48
00:02:02,970 --> 00:02:05,850
to compete in this market
has led to brittle monopolies

49
00:02:05,850 --> 00:02:08,733
that can stifle
entrepreneurship and innovation.

50
00:02:09,780 --> 00:02:12,930
P is for Plagiarism and Privacy.

51
00:02:12,930 --> 00:02:16,170
Artists, musicians, and
authors are suing AI companies

52
00:02:16,170 --> 00:02:18,540
for training their models
on copyrighted works,

53
00:02:18,540 --> 00:02:20,400
sometimes churning out knockoffs

54
00:02:20,400 --> 00:02:22,290
that resemble their own work.

55
00:02:22,290 --> 00:02:23,490
On the privacy front,

56
00:02:23,490 --> 00:02:24,930
facial recognition software

57
00:02:24,930 --> 00:02:27,780
can trawl the internet for your face,

58
00:02:27,780 --> 00:02:30,390
and teens in middle school
can create deepfake nudes

59
00:02:30,390 --> 00:02:31,830
of their classmates,

60
00:02:31,830 --> 00:02:34,323
turning privacy into a relic of the past.

61
00:02:35,550 --> 00:02:37,950
A is for Automated Labor.

62
00:02:37,950 --> 00:02:41,490
While few jobs may be
completely replaced by robots,

63
00:02:41,490 --> 00:02:43,950
companies looking to cut costs
can maintain productivity

64
00:02:43,950 --> 00:02:46,860
by relying on senior
employees armed with AI.

65
00:02:46,860 --> 00:02:48,000
That means fewer openings

66
00:02:48,000 --> 00:02:49,830
for junior lawyers and copywriters,

67
00:02:49,830 --> 00:02:51,600
and fewer freelance opportunities

68
00:02:51,600 --> 00:02:53,820
for designers and composers.

69
00:02:53,820 --> 00:02:55,800
Meanwhile, high stakes decisions,

70
00:02:55,800 --> 00:02:57,750
whether on Wall Street
or on the battlefield,

71
00:02:57,750 --> 00:03:00,030
are being outsourced to algorithms.

72
00:03:00,030 --> 00:03:02,430
Without sufficient human oversight,

73
00:03:02,430 --> 00:03:03,840
sudden shifts in job markets

74
00:03:03,840 --> 00:03:06,570
due to automation could
destabilize economies

75
00:03:06,570 --> 00:03:07,983
faster than they can adapt.

76
00:03:10,590 --> 00:03:12,780
C is for Climate Impact.

77
00:03:12,780 --> 00:03:16,230
By 2027, the AI sector could
eat as much energy per year

78
00:03:16,230 --> 00:03:17,580
as the Netherlands.

79
00:03:17,580 --> 00:03:21,150
Protestors from Oregon to Uruguay
are targeting data centers

80
00:03:21,150 --> 00:03:24,180
that suck up drinking water
from local reservoirs.

81
00:03:24,180 --> 00:03:25,830
Fevers to win the race to roll out

82
00:03:25,830 --> 00:03:27,960
ever bigger generative AI models,

83
00:03:27,960 --> 00:03:30,300
Microsoft and Google
have fallen well behind

84
00:03:30,300 --> 00:03:32,100
in their carbon reduction goals,

85
00:03:32,100 --> 00:03:34,743
accelerating the threat of climate change.

86
00:03:36,300 --> 00:03:39,240
T is for Tainted Data.

87
00:03:39,240 --> 00:03:43,410
The ghosts of human bias
haunt our AI systems.

88
00:03:43,410 --> 00:03:45,420
Resume screeners
discriminate against women

89
00:03:45,420 --> 00:03:47,130
and non-white applicants.

90
00:03:47,130 --> 00:03:50,250
Facial recognition systems
struggle with darker skin tones,

91
00:03:50,250 --> 00:03:52,500
potentially resulting
in a self-driving car

92
00:03:52,500 --> 00:03:55,170
failing to identify a Black pedestrian.

93
00:03:55,170 --> 00:03:58,290
Even tools meant to
identify AI-generated essays

94
00:03:58,290 --> 00:04:01,740
can incorrectly flag
non-natives English speakers.

95
00:04:01,740 --> 00:04:03,120
These aren't just glitches.

96
00:04:03,120 --> 00:04:06,303
They're mirrors reflecting
society's inequities.

97
00:04:08,400 --> 00:04:09,840
Now let's do RISK.

98
00:04:09,840 --> 00:04:12,873
Starting with R, Reality Distortion.

99
00:04:13,950 --> 00:04:16,530
AI's outputs are often
disconnected from reality

100
00:04:16,530 --> 00:04:18,750
with potentially dangerous consequences.

101
00:04:18,750 --> 00:04:21,480
Bing's Sydney Chatbot told
a New York Times reporter

102
00:04:21,480 --> 00:04:23,160
to leave his wife for her,

103
00:04:23,160 --> 00:04:26,193
while ChatGPT accused a law
professor of sexual harassment.

104
00:04:27,120 --> 00:04:28,800
Unscrupulous users can now generate

105
00:04:28,800 --> 00:04:31,920
low quality news articles
several per minute

106
00:04:31,920 --> 00:04:34,770
that could be hard to distinguish
from human-written posts.

107
00:04:34,770 --> 00:04:36,600
These confident sounding fabrications

108
00:04:36,600 --> 00:04:40,290
or hallucinations as they're
sometimes misleadingly called

109
00:04:40,290 --> 00:04:43,140
don't have to be deliberate
disinformation to cause harm,

110
00:04:43,140 --> 00:04:45,480
that makes them different from infowar.

111
00:04:45,480 --> 00:04:48,330
Persistent exposure to
AI-generated distortions

112
00:04:48,330 --> 00:04:50,370
may erode trust inauthentic sources

113
00:04:50,370 --> 00:04:51,960
of information over time,

114
00:04:51,960 --> 00:04:54,993
or simply bury them in a
mountain of digital debris.

115
00:04:57,570 --> 00:04:59,160
I for Injustice.

116
00:04:59,160 --> 00:05:01,830
While Silicon Valley
reaps the rewards of AI,

117
00:05:01,830 --> 00:05:03,750
the human cost is often hidden.

118
00:05:03,750 --> 00:05:07,230
Kenyans making a dollar an
hour suffering emotional trauma

119
00:05:07,230 --> 00:05:10,680
while flagging detailed
descriptions of murder, rape,

120
00:05:10,680 --> 00:05:14,010
and child sex abuse generated by chatbots.

121
00:05:14,010 --> 00:05:15,240
Meanwhile, in the Congo,

122
00:05:15,240 --> 00:05:17,910
children as young as
seven are mining cobalt

123
00:05:17,910 --> 00:05:19,743
for our AI-powered devices.

124
00:05:20,640 --> 00:05:22,710
In government, powerful AI companies

125
00:05:22,710 --> 00:05:25,050
might influence
regulations in their favor,

126
00:05:25,050 --> 00:05:27,480
perpetuating inequality in the market.

127
00:05:27,480 --> 00:05:29,550
In education, the subscription costs

128
00:05:29,550 --> 00:05:31,320
of better large language models

129
00:05:31,320 --> 00:05:33,753
raise concerns about a new digital divide.

130
00:05:36,570 --> 00:05:38,520
S is for Stereotyping.

131
00:05:38,520 --> 00:05:40,770
Generative AI often devolves to cliches

132
00:05:40,770 --> 00:05:42,180
when answering questions.

133
00:05:42,180 --> 00:05:44,280
Ask an image generator for a doctor,

134
00:05:44,280 --> 00:05:47,640
you're likely to get a middle
aged white guy in a lab coat.

135
00:05:47,640 --> 00:05:49,440
Request an African village,

136
00:05:49,440 --> 00:05:51,903
and it might depict mud huts and sunsets.

137
00:05:52,950 --> 00:05:54,900
These outputs aren't simply the result

138
00:05:54,900 --> 00:05:57,540
of tainted data described earlier.

139
00:05:57,540 --> 00:05:59,700
They operate at a deeper level,

140
00:05:59,700 --> 00:06:01,740
reflecting the probabilistic nature

141
00:06:01,740 --> 00:06:02,850
of large language models,

142
00:06:02,850 --> 00:06:06,360
which return the average
result from a prompt.

143
00:06:06,360 --> 00:06:09,570
Nevertheless, by surfacing
stereotypes rather than outliers,

144
00:06:09,570 --> 00:06:12,330
AI might favor dominant
cultures and languages,

145
00:06:12,330 --> 00:06:14,973
leading to the erosion
of cultural diversity.

146
00:06:17,190 --> 00:06:20,730
Finally, K is for Knockoff Experiences.

147
00:06:20,730 --> 00:06:22,050
In classrooms across the world,

148
00:06:22,050 --> 00:06:24,300
teachers grapple with
AI-generated homework

149
00:06:24,300 --> 00:06:26,820
as students outsource
their learning to chatbots.

150
00:06:26,820 --> 00:06:28,890
In the music world, services like Suno

151
00:06:28,890 --> 00:06:30,870
allow everyone to create songs

152
00:06:30,870 --> 00:06:32,220
without touching an instrument,

153
00:06:32,220 --> 00:06:34,050
potentially flooding streaming platforms

154
00:06:34,050 --> 00:06:36,270
with AI-generated tunes.

155
00:06:36,270 --> 00:06:38,220
These shortcuts threaten to replace

156
00:06:38,220 --> 00:06:41,730
genuine skill development
and creative expression

157
00:06:41,730 --> 00:06:44,613
with shallow imitations that
undermine personal growth.

158
00:06:45,870 --> 00:06:48,600
Beyond the individual, human
relationships may devolve

159
00:06:48,600 --> 00:06:50,820
as AI generates thank you notes

160
00:06:50,820 --> 00:06:53,820
and adults turn to AI
partners for companionship

161
00:06:53,820 --> 00:06:56,073
instead of their flesh and blood peers.

162
00:07:00,270 --> 00:07:03,450
By keeping the IMPACT
RISK framework in mind,

163
00:07:03,450 --> 00:07:06,810
we can navigate the AI
revolution more thoughtfully,

164
00:07:06,810 --> 00:07:09,690
examining whether AI can
be a force for progress

165
00:07:09,690 --> 00:07:13,743
rather than just a Pandora's
box of unintended consequences.