1
00:00:00,040 --> 00:00:03,600
The paper we're diving into 
today is Accountability AI 

2
00:00:03,600 --> 00:00:06,960
decisions. 
Signing for future liability 

3
00:00:07,080 --> 00:00:10,360
When is endangered ends by Marco
ABBA. 

4
00:00:11,240 --> 00:00:14,840
OK, the title is a mouth, but 
it's pretty timely. 

5
00:00:15,440 --> 00:00:20,160
AI is everywhere in healthcare 
and well, things aren't as safe 

6
00:00:20,160 --> 00:00:23,920
or as fair they should be. 
Yeah, exactly. 

7
00:00:23,920 --> 00:00:27,320
I reread the title too and 
thought, whoa, signing for 

8
00:00:27,320 --> 00:00:30,200
future liability? 
That's a whole new level. 

9
00:00:30,760 --> 00:00:34,680
So Marco Saba asks this pretty 
uncomfortable question. 

10
00:00:35,120 --> 00:00:38,600
When AI systems alignment, 
meaning the way it's behavior is

11
00:00:38,600 --> 00:00:42,960
adjusted, messes up, gives out 
dangerous advice, or refuses to 

12
00:00:42,960 --> 00:00:46,560
warn you about a deadly 
interaction, who's actually for 

13
00:00:46,560 --> 00:00:50,280
the harm, Right. 
And that's the core problem the 

14
00:00:50,280 --> 00:00:54,320
paper goes after. 
There's this gap when something 

15
00:00:54,320 --> 00:00:58,040
awful happens because of an AI 
alignment decision, like say 

16
00:00:58,120 --> 00:01:01,240
it's over al and stays silent 
instead of flagging a drug 

17
00:01:01,240 --> 00:01:03,760
problem. 
Or it's under aligned and does 

18
00:01:03,760 --> 00:01:07,520
recklessly helpful. 
No one really the price least of

19
00:01:07,520 --> 00:01:10,680
all the actual human decision 
makers behind the alignment 

20
00:01:10,680 --> 00:01:13,000
protocols. 
And they are here. 

21
00:01:13,440 --> 00:01:16,480
The paper shares some real world
case, like the story about 

22
00:01:16,480 --> 00:01:20,400
Sarah, a diabetic who literally 
suffered organ damage because an

23
00:01:20,480 --> 00:01:23,080
AI tool withheld a warning about
her meds. 

24
00:01:23,640 --> 00:01:27,320
Or the mental health chat bot 
that, after persistent requests,

25
00:01:27,480 --> 00:01:30,800
gave a teenager instructions on 
lethal drug combos. 

26
00:01:31,320 --> 00:01:33,520
That wild. 
It's not science fiction's 

27
00:01:33,520 --> 00:01:37,040
happening now. 
Yeah, and it's not just rare 

28
00:01:37,040 --> 00:01:40,680
freak events. 
Saba's numbers staggering 

29
00:01:41,000 --> 00:01:46,160
alignments gone wrong add up to 
.2% higher mortality rate in 

30
00:01:46,160 --> 00:01:50,160
underserved communities. 
That's thousands of disability 

31
00:01:50,160 --> 00:01:53,960
adjusted life years lost every 
year just in the US. 

32
00:01:54,600 --> 00:01:57,640
Honestly it made me keep 
checking if my own uses. 

33
00:01:57,640 --> 00:02:02,440
Any AI bots I should be wary of?
The fundamental insight is that 

34
00:02:02,440 --> 00:02:06,120
current legal frameworks, the 
so-called corporate veil, and 

35
00:02:06,120 --> 00:02:09,600
the box defense make it 
basically impossible to pin 

36
00:02:09,600 --> 00:02:11,720
accountability on any one 
person. 

37
00:02:12,240 --> 00:02:16,200
It's always the company, never 
the execs, never the 

38
00:02:16,200 --> 00:02:18,440
individually. 
It's like when you're on the 

39
00:02:18,440 --> 00:02:21,120
phone with some cable company 
and nobody can help you. 

40
00:02:21,320 --> 00:02:24,600
But, you know, lives are at 
stake, yeah. 

41
00:02:24,600 --> 00:02:28,440
And the proposed solution is 
interesting, a sort of tiered 

42
00:02:28,440 --> 00:02:31,840
sign system for big risky 
alignment decisions. 

43
00:02:32,440 --> 00:02:36,440
The model comes from places, 
aviation or pharma, where 

44
00:02:36,440 --> 00:02:40,440
someone an actual person to put 
their name to high stakes 

45
00:02:40,440 --> 00:02:43,960
processes. 
So if the AI alignment ends up 

46
00:02:43,960 --> 00:02:47,840
killing or harming people, you 
can trace it back and people can

47
00:02:47,840 --> 00:02:51,320
be held criminally liable. 
I like that. 

48
00:02:51,400 --> 00:02:56,040
I mean like I love the idea of 
visible accountability, but it 

49
00:02:56,040 --> 00:02:59,960
also, I know it's daunting. 
There's old fear that if you 

50
00:02:59,960 --> 00:03:03,160
start putting that much onto 
individuals, nobody would want 

51
00:03:03,160 --> 00:03:05,800
the OR maybe it'd slow down 
progress. 

52
00:03:06,360 --> 00:03:09,440
Then again, as points out, it 
hasn't hurt innovation in 

53
00:03:09,440 --> 00:03:13,560
aviation or pharma, but I 
wonder, does anyone actually 

54
00:03:13,560 --> 00:03:17,480
check all those sign off forms 
or is it just a bureaucratic 

55
00:03:17,480 --> 00:03:19,960
ritual? 
Sorry, that's a tangent. 

56
00:03:20,760 --> 00:03:25,120
No, I get it, that kind of 
skepticism is, but the argument 

57
00:03:25,120 --> 00:03:29,560
in the paper is without personal
people will always play it safe 

58
00:03:29,560 --> 00:03:33,080
for themselves, not the patient 
that saps. 

59
00:03:33,080 --> 00:03:37,120
And honestly, it can kill. 
The paper suggests making 

60
00:03:37,120 --> 00:03:40,720
everything edible. 
Public registries, whistleblower

61
00:03:40,720 --> 00:03:45,800
protections, a full system so 
alignment isn't just some tweak 

62
00:03:45,800 --> 00:03:49,520
but a line item. 
The biggest take away from me is

63
00:03:49,520 --> 00:03:52,720
that these alignment decisions 
aren't neutral, they're value 

64
00:03:52,720 --> 00:03:55,280
choices. 
Just like in law or medicine, 

65
00:03:55,840 --> 00:03:59,920
somebody's judgment about better
safe than sorry versus don't 

66
00:03:59,920 --> 00:04:03,400
hide important info has direct 
consequences. 

67
00:04:04,520 --> 00:04:08,720
No AI is perfectly safe, but the
point is should know who what 

68
00:04:09,440 --> 00:04:12,440
exactly. 
And for us, probably anyone 

69
00:04:12,440 --> 00:04:16,000
listening, it's this reminder. 
Even though research like 

70
00:04:16,000 --> 00:04:19,480
doesn't give final answers, it 
points out urgent gaps and 

71
00:04:19,480 --> 00:04:22,440
practical paths. 
It's not just a theoretical 

72
00:04:22,440 --> 00:04:24,560
debate. 
It's about demanding that 

73
00:04:24,560 --> 00:04:28,280
healthy algorithms come with 
real human responsibility 

74
00:04:28,280 --> 00:04:30,120
attached. 
Yeah. 

75
00:04:30,280 --> 00:04:34,560
So bottom line, if you trust N 
with your health, wouldn't you 

76
00:04:34,560 --> 00:04:37,200
too? 
There a person somewhere is 

77
00:04:37,200 --> 00:04:41,400
actually answerable? 
No would me too. 

78
00:04:41,880 --> 00:04:45,760
All right, let's hope that 
signing for Future Liability 

79
00:04:46,000 --> 00:04:49,400
isn't just a catchy title, but a
real change on the way.