WEBVTT
Kind: captions
Language: en

1
00:00:00.109 --> 00:00:03.270
Back when online video was a different shape,

2
00:00:03.270 --> 00:00:05.630
searching for videos on YouTube relied on…

3
00:00:05.630 --> 00:00:07.570
well, honesty, really.

4
00:00:07.570 --> 00:00:08.590
Creators would put up a video,

5
00:00:08.590 --> 00:00:10.269
they would tell YouTube what was in it,

6
00:00:10.269 --> 00:00:13.200
and YouTube would look at the title and the
description and the tags

7
00:00:13.200 --> 00:00:14.450
and take it from there.

8
00:00:14.450 --> 00:00:16.930
And immediately, people tried to cheat.

9
00:00:16.930 --> 00:00:18.020
It turned into an arms race.

10
00:00:18.020 --> 00:00:21.430
Keyword stuffing, tag spam,
the reply girls of 2012.

11
00:00:21.430 --> 00:00:24.930
If there was any way to guess
the YouTube algorithm's priorities,

12
00:00:24.930 --> 00:00:26.550
then spammers exploited it.

13
00:00:26.550 --> 00:00:29.240
And desperate video creators changed what
they were making

14
00:00:29.240 --> 00:00:31.620
to fit what they thought YouTube wanted,

15
00:00:31.620 --> 00:00:35.710
based on rumour, speculation,
and extrapolating way too much

16
00:00:35.710 --> 00:00:38.050
from way too little data.

17
00:00:38.050 --> 00:00:41.400
So recently, YouTube has just kept quiet.

18
00:00:41.400 --> 00:00:43.720
Their parent company, Google,
learned this lesson back when

19
00:00:43.720 --> 00:00:45.989
online video was a RealPlayer window

20
00:00:45.989 --> 00:00:49.470
and connecting to the internet sounded like
the screeching of a robot cat:

21
00:00:49.470 --> 00:00:54.220
if Google ever gave a hint about how to boost
web sites up their search results,

22
00:00:54.220 --> 00:00:57.420
then there would be a rush of spammers trying
to exploit that knowledge.

23
00:00:57.420 --> 00:00:58.540
So the advice was always:

24
00:00:58.540 --> 00:01:00.190
Just make good things.

25
00:01:00.190 --> 00:01:01.990
We’ll figure it out.

26
00:01:01.990 --> 00:01:04.680
But now, on YouTube, it's not that the folks
who control the algorithm

27
00:01:04.680 --> 00:01:06.690
won't tell people how it works.

28
00:01:06.690 --> 00:01:08.380
It's that they can't.

29
00:01:08.380 --> 00:01:12.090
And here's the evidence: a paper written by
YouTube engineers,

30
00:01:12.090 --> 00:01:17.150
explaining they're using Google's research
into machine-learning to recommend videos.

31
00:01:17.150 --> 00:01:21.060
That is the same kind of software that creates
those weird Deep Dream images,

32
00:01:21.060 --> 00:01:23.520
that makes their text to speech sound so realistic,

33
00:01:23.520 --> 00:01:27.790
and that beat the world's best Go player at
his own game.

34
00:01:27.790 --> 00:01:29.360
And I know I'm simplifying a bit here,

35
00:01:29.360 --> 00:01:31.799
but machine learning,
the way Google does it,

36
00:01:31.799 --> 00:01:34.170
is basically a black box.

37
00:01:34.170 --> 00:01:35.650
You give a neural network some input,

38
00:01:35.650 --> 00:01:37.520
like the game board in Go.

39
00:01:37.520 --> 00:01:40.470
And it gives you outputs:
moves it thinks might work.

40
00:01:40.470 --> 00:01:41.700
Those outputs are tested,

41
00:01:41.700 --> 00:01:43.280
and the results go back into the box,

42
00:01:43.280 --> 00:01:45.750
and then you repeat that process
a billion or so times,

43
00:01:45.750 --> 00:01:47.810
and it starts to get really good.

44
00:01:47.810 --> 00:01:50.980
But no-one can look inside that black box
and see how it works:

45
00:01:50.980 --> 00:01:54.141
it’s designed by a computer, for a computer.

46
00:01:54.141 --> 00:01:56.460
And neural networks are great
if you're playing a game

47
00:01:56.460 --> 00:01:58.600
with an obvious scoring and points system.

48
00:01:58.600 --> 00:02:00.930
You win, or you lose.

49
00:02:00.930 --> 00:02:04.220
But training that black box on YouTube videos
is a bit messier.

50
00:02:04.220 --> 00:02:06.540
It’s not just that human behaviour
is unpredictable and complicated:

51
00:02:06.540 --> 00:02:10.459
it’s trying to work out what counts as winning
in the first place.

52
00:02:10.459 --> 00:02:14.170
If YouTube tells the algorithm
"show videos that people like",

53
00:02:14.170 --> 00:02:16.500
then it'll kill any channel
which talks about politics,

54
00:02:16.500 --> 00:02:18.569
where people hit dislike if they disagree.

55
00:02:18.569 --> 00:02:21.300
And it'll silence anyone who has
a small but vocal group

56
00:02:21.300 --> 00:02:23.120
loudly disagreeing with them.

57
00:02:23.120 --> 00:02:26.700
If YouTube tells the algorithm,
"okay, show videos that people share",

58
00:02:26.700 --> 00:02:30.569
then videos about private things like medical
issues or sex education vanish,

59
00:02:30.569 --> 00:02:34.680
and folks who have a
small, loyal, but quiet fanbase

60
00:02:34.680 --> 00:02:36.050
disappear into their own little world.

61
00:02:36.050 --> 00:02:37.730
And YouTube creators, of course,

62
00:02:37.730 --> 00:02:41.090
would love the algorithm to recommend
only their own videos...

63
00:02:41.090 --> 00:02:45.129
even when the rest of the world doesn’t
actually want to watch them.

64
00:02:45.129 --> 00:02:47.630
So YouTube started out,
according to the paper,

65
00:02:47.630 --> 00:02:49.660
giving their algorithm
the reasonable goal of

66
00:02:49.660 --> 00:02:52.170
"increase watch time".

67
00:02:52.170 --> 00:02:54.099
But that has a few problems.

68
00:02:54.099 --> 00:02:57.670
Because there's no way for a computer
to determine quality, or truth.

111
00:02:57.670 --> 00:02:59.379
At least, not yet.

70
00:02:59.379 --> 00:03:01.030
The system doesn't understand context,

71
00:03:01.030 --> 00:03:03.910
it can't tell the difference between
actual, reliable information

72
00:03:03.910 --> 00:03:06.790
and unhinged, paranoid conspiracy clickbait.

73
00:03:06.790 --> 00:03:09.569
Although, admittedly,
neither can a lot of people,

74
00:03:09.569 --> 00:03:12.330
which is why these videos
are getting a lot of traction.

75
00:03:12.330 --> 00:03:14.120
And it can't tell the difference

76
00:03:14.120 --> 00:03:16.580
between videos that are suitable for children,

77
00:03:16.580 --> 00:03:18.800
made with education in mind,

78
00:03:18.800 --> 00:03:22.349
and creepy, trademark-infringing
unofficial efforts.

79
00:03:22.349 --> 00:03:26.640
It just knows what kids click on
and what they watch.

80
00:03:26.640 --> 00:03:30.989
So, sure, the algorithm might increase watch
time, in the short term…

81
00:03:30.989 --> 00:03:33.739
but a lot of the videos it recommends are
going to be questionable at best

82
00:03:33.740 --> 00:03:35.880
and actively harmful at worst.

83
00:03:35.880 --> 00:03:37.220
And they’re going to be the sort of thing

84
00:03:37.230 --> 00:03:39.520
that advertisers get really nervous about.

85
00:03:39.520 --> 00:03:41.800
Remember, the algorithm is a black box.

86
00:03:41.800 --> 00:03:43.180
No-one knows what it’s doing.

87
00:03:43.190 --> 00:03:46.459
All YouTube can do is change the feedback
it’s getting,

88
00:03:46.459 --> 00:03:50.810
change the signals that say
“this is good” or “this is bad”.

89
00:03:50.810 --> 00:03:54.319
If YouTube wanted a human to watch and categorise
every video being uploaded

90
00:03:54.319 --> 00:03:57.299
as “safe” or “unsafe” in real time,

91
00:03:57.299 --> 00:04:01.129
they would need about 100,000 employees
working shifts round the clock.

92
00:04:01.129 --> 00:04:03.150
Plus, that would expose them to legal issues:

93
00:04:03.150 --> 00:04:05.060
in most of the countries
where YouTube has an office,

94
00:04:05.060 --> 00:04:07.060
if you let an algorithm do the filtering

95
00:04:07.060 --> 00:04:09.519
and then manually step in when you get a complaint,

96
00:04:09.519 --> 00:04:10.900
you're legally fine.

97
00:04:10.900 --> 00:04:13.300
But if you approve everything
with a human in the loop,

98
00:04:13.300 --> 00:04:14.510
you are a publisher,

99
00:04:14.510 --> 00:04:18.609
and you're opening yourself up to
some very expensive lawsuits.

100
00:04:18.609 --> 00:04:21.069
The ideal algorithm, the ideal black box,

101
00:04:21.069 --> 00:04:22.349
from YouTube's point of view,

102
00:04:22.349 --> 00:04:26.059
would be one with a goal of "increase ad revenue",

103
00:04:26.059 --> 00:04:28.020
and which thought about the long-term,

104
00:04:28.020 --> 00:04:29.409
which knew about social issues,

105
00:04:29.409 --> 00:04:32.289
and potential advertiser boycotts,
and financial strategies,

106
00:04:32.289 --> 00:04:34.870
and public perception,
and what’s suitable for kids,

107
00:04:34.870 --> 00:04:36.940
and... and about truth.

108
00:04:36.940 --> 00:04:39.409
At which point,
what you have is something

109
00:04:39.409 --> 00:04:42.699
that can do the job of YouTube's
senior management team.

110
00:04:42.699 --> 00:02:57.670
And artificial intelligence
hasn't gotten that good.

111
00:02:57.670 --> 00:02:59.379
At least, not yet.