1
00:00:00,160 --> 00:00:05,200
He started getting sick last summer, in
North Carolina...

2
00:00:05,320 --> 00:00:08,320
and none of it made any sense to him.

3
00:00:08,520 --> 00:00:13,600
Not one bad week. Again, and then again,
across 4

4
00:00:13,760 --> 00:00:16,720
months. He was losing weight.

5
00:00:16,880 --> 00:00:20,480
People who knew him said he looked
unrecognizable.

6
00:00:20,520 --> 00:00:25,920
And on the worst day of it he spent 72
hours on the couch, heart going

7
00:00:26,080 --> 00:00:31,520
wrong, pain through his abdomen, confusion
so deep he couldn't put a text message

8
00:00:31,600 --> 00:00:35,040
together. He worked in cybersecurity.

9
00:00:35,120 --> 00:00:40,093
Which is to say he spent his professional
life looking for the thing that was off...

10
00:00:40,133 --> 00:00:43,280
and he couldn't find it in his own house.

11
00:00:43,360 --> 00:00:47,973
So he started wondering about his drinks,
because one of them had tasted bitter.

12
00:00:48,027 --> 00:00:49,960
And he did the thing you would do.

13
00:00:50,000 --> 00:00:54,480
He poured it into a clear cup and had a
proper look at it, and he filmed himself

14
00:00:54,560 --> 00:00:58,440
doing it on his phone. He stared into that
cup.

15
00:00:58,480 --> 00:01:02,800
He found nothing wrong with it, he threw
the can away, and he carried on being

16
00:01:02,880 --> 00:01:08,080
sick. Months later, the police watched
that same video...

17
00:01:08,160 --> 00:01:13,280
and they saw white powdery residue,
clumping at the bottom of the cup, as it

18
00:01:13,360 --> 00:01:17,440
poured. Same footage. Same man.

19
00:01:17,920 --> 00:01:23,120
He was looking straight at the thing that
was poisoning him and he couldn't see it,

20
00:01:23,240 --> 00:01:27,200
and somebody who knew what to look for saw
it immediately.

21
00:01:27,680 --> 00:01:30,560
But the police weren't working from that
video.

22
00:01:30,680 --> 00:01:32,560
They had something better.

23
00:01:32,640 --> 00:01:35,360
They were reading his wife's chat logs.

24
00:01:35,920 --> 00:01:39,440
Everyone keeps asking whether the chatbot
is to blame.

25
00:01:39,520 --> 00:01:43,680
But what about what it was supposed to do
when the conversation turned?

26
00:01:43,760 --> 00:01:47,440
What is it supposed to notice, and who
decided that?

27
00:01:47,520 --> 00:01:50,400
We are not talking about somebody tricking
it.

28
00:01:50,880 --> 00:01:55,440
Everything after this comes out of the
lawsuit he has since filed against OpenAI

29
00:01:55,480 --> 00:01:58,640
and its chief executive. None of it is
proven.

30
00:01:59,200 --> 00:02:04,240
A grand jury has charged her, she has not
been convicted of anything, and her case

31
00:02:04,320 --> 00:02:08,080
has not been heard. It starts with her
being frightened.

32
00:02:08,200 --> 00:02:12,400
She tells the chatbot she thinks her
husband might be trying to kill her, and

33
00:02:12,480 --> 00:02:14,640
she needs to know what to watch for.

34
00:02:14,760 --> 00:02:17,120
Which is a reasonable thing to be afraid
of.

35
00:02:17,240 --> 00:02:19,520
And a reasonable thing to ask.

36
00:02:19,640 --> 00:02:24,000
So it answers. And because it answers, she
asks again.

37
00:02:24,320 --> 00:02:29,200
There are at least 25 of these
conversations, across months, and the

38
00:02:29,280 --> 00:02:31,920
interesting thing is not any one of them.

39
00:02:31,960 --> 00:02:37,040
It's the drift. Because the questions stop
being about what to watch for.

40
00:02:37,120 --> 00:02:41,760
She starts asking whether a particular
drug would show up on an autopsy report.

41
00:02:41,920 --> 00:02:46,320
Whether anti-nausea medicine would make a
death come faster, or slower.

42
00:02:46,480 --> 00:02:50,400
Whether something could be injected
instead of swallowed and still work.

43
00:02:50,480 --> 00:02:54,080
Those are not the questions of somebody
protecting herself.

44
00:02:54,200 --> 00:02:59,280
And somewhere in there, in a single line,
she stops pretending they are.

45
00:02:59,360 --> 00:03:04,760
She asks how much would be, quote, pretty
much instantly fatal by injection...

46
00:03:04,800 --> 00:03:08,720
to me OR HIM. Him in capital letters.

47
00:03:09,120 --> 00:03:11,200
I'm not going to tell you what it
answered.

48
00:03:11,600 --> 00:03:16,000
The answers were specific, and the
specifics are the one part of this story

49
00:03:16,080 --> 00:03:17,640
that doesn't need repeating.

50
00:03:17,840 --> 00:03:22,200
But there's a detail I can't get past, and
it isn't an answer at all.

51
00:03:22,400 --> 00:03:26,400
It's a title. The model names its own
conversations.

52
00:03:26,640 --> 00:03:29,840
It reads what's in them and writes a label
at the top.

53
00:03:30,080 --> 00:03:35,680
One of these is called why a particular
energy drink would be, quote, more stealth

54
00:03:35,760 --> 00:03:38,080
for drugging. It wrote that.

55
00:03:38,560 --> 00:03:44,080
It read the conversation, worked out
exactly what the conversation was, gave it

56
00:03:44,120 --> 00:03:47,200
a name... and then carried on having it.

57
00:03:47,600 --> 00:03:51,520
And when she asked whether she could erase
all of this afterwards, it told her she

58
00:03:51,680 --> 00:03:56,320
could. Which, the lawsuit points out, is
not how the product's data retention

59
00:03:56,400 --> 00:04:01,120
works. Think about it. The system was
given a reason, right at the start.

60
00:04:01,280 --> 00:04:04,520
The reason was good enough for everything
that came after it.

61
00:04:04,560 --> 00:04:08,440
Now. I wanted to know how unusual this
was.

62
00:04:08,720 --> 00:04:13,120
Whether this was one strange conversation
that went somewhere nobody could have

63
00:04:13,400 --> 00:04:18,560
predicted. It isn't. That lawsuit was
filed inside a 2 day stretch in

64
00:04:18,640 --> 00:04:21,360
October, alongside 12 others.

65
00:04:21,600 --> 00:04:24,400
8 of the 12 came from state attorneys
general.

66
00:04:24,640 --> 00:04:30,320
Arizona, Arkansas, Idaho, Iowa, Kansas,
Montana,

67
00:04:30,400 --> 00:04:36,080
Nebraska, West Virginia. 8 county
courthouses, the same lawyers,

68
00:04:36,240 --> 00:04:38,480
long stretches of identical wording.

69
00:04:38,880 --> 00:04:40,720
The private cases are families.

70
00:04:41,200 --> 00:04:44,720
One was brought by the parents of a 15
year old boy who died.

71
00:04:45,280 --> 00:04:49,520
13 lawsuits. 2 days. ONE product.

72
00:04:49,840 --> 00:04:54,960
So there's a question sitting underneath
all 13, and I haven't heard anybody ask it

73
00:04:55,040 --> 00:05:00,960
out loud. What is a chatbot required to do
when somebody walks it, slowly and

74
00:05:01,040 --> 00:05:04,480
politely, toward harming a named human
being?

75
00:05:05,120 --> 00:05:09,360
Simple. I could not find a published
answer to that anywhere.

76
00:05:09,600 --> 00:05:12,520
And that absence is the entire problem.

77
00:05:12,560 --> 00:05:15,200
You can see it in what these filings are
demanding.

78
00:05:15,680 --> 00:05:20,240
Detect and interrupt conversations where
somebody appears to be planning violence

79
00:05:20,320 --> 00:05:22,440
against an identifiable person.

80
00:05:22,720 --> 00:05:25,680
Check ages. Delete children's data.

81
00:05:26,160 --> 00:05:30,000
Stop building models that improve
themselves until somebody can show they

82
00:05:30,080 --> 00:05:33,120
are safe. None of that is a claim for
money.

83
00:05:33,360 --> 00:05:35,520
Those are product requirements.

84
00:05:35,760 --> 00:05:40,320
Which means the safety specification for
the chatbot sitting on your phone is being

85
00:05:40,400 --> 00:05:45,760
drafted right now, in Maricopa County and
Faulkner County and Ada County, by

86
00:05:45,840 --> 00:05:48,320
plaintiffs' lawyers and state attorneys
general.

87
00:05:48,640 --> 00:05:52,640
And it's being drafted there because I
could not find anybody who wrote one

88
00:05:52,720 --> 00:05:58,240
anywhere else first. OpenAI says safety is
a top priority, that it banned the

89
00:05:58,320 --> 00:06:02,560
user, and that it gave information to law
enforcement for their investigation.

90
00:06:02,800 --> 00:06:07,440
Not one of these 13 cases has been
decided, and the company hasn't answered

91
00:06:07,520 --> 00:06:08,920
any of them in court yet.

92
00:06:08,960 --> 00:06:11,680
He is well now. The marriage is over.

93
00:06:11,920 --> 00:06:14,120
The chat logs are exhibits.

94
00:06:14,480 --> 00:06:20,240
The residue was sitting in that video the
entire time he was ill, and the only

95
00:06:20,400 --> 00:06:22,400
thing that ever changed...

96
00:06:22,480 --> 00:06:25,760
was that somebody finally knew to go and
look for it.

97
00:06:26,720 --> 00:06:29,840
I'm Guillermo, and this is my AI clone.