WEBVTT
Kind: captions
Language: en

1
00:00:00.000 --> 00:00:02.080
[Simulated voice]
- "Hi, I'm not Tom Scott."

2
00:00:02.080 --> 00:00:03.440
- But I am. Hello!

3
00:00:03.440 --> 00:00:05.040
This is the real me,
not simulated.

4
00:00:05.040 --> 00:00:07.280
This week's guest video
is from Jordan Harrod

5
00:00:07.280 --> 00:00:09.280
who is probably the
smartest person I know.

6
00:00:09.280 --> 00:00:12.320
Jordan is working on a
joint Harvard-MIT doctorate

7
00:00:12.320 --> 00:00:14.240
in medical engineering
and medical physics,

8
00:00:14.240 --> 00:00:16.320
is researching using AI in neuroscience,

9
00:00:16.320 --> 00:00:19.440
and also has a channel with deep dives
into AI and algorithms.

10
00:00:19.440 --> 00:00:21.440
So if there was anyone
I was going to trust

11
00:00:21.440 --> 00:00:24.400
with making a simulated
version of me, it was Jordan.

12
00:00:24.400 --> 00:00:26.560
But I want to take a moment
to manage expectations,

13
00:00:26.560 --> 00:00:29.840
this isn't a face replacement on
a body double. That's mainstream now.

14
00:00:29.840 --> 00:00:31.760
Hollywood's doing it,
anyone can do it.

15
00:00:31.760 --> 00:00:35.360
This is trying to create a cheaper
complete replacement for me,

16
00:00:35.360 --> 00:00:38.640
an automatic digital puppet
on a budget of just $100.

17
00:00:39.200 --> 00:00:42.160
Oh, and one more thing.
Just because I trust Jordan to do this,

18
00:00:42.160 --> 00:00:44.800
this is not consent for anyone else.

19
00:00:44.800 --> 00:00:46.640
Please: don't try this at home!

20
00:00:46.640 --> 00:00:49.840
- "Anytime you see or hear
Tom Scott in this video,

21
00:00:49.840 --> 00:00:52.560
"what you're actually seeing
is a fake version of Tom Scott

22
00:00:52.560 --> 00:00:54.560
"that was generated using machine learning."

23
00:00:54.560 --> 00:00:55.280
- Why?

24
00:00:55.280 --> 00:00:59.360
Well, some of you are getting better and better
at creating Tom Scott-style videos,

25
00:00:59.360 --> 00:01:03.680
and it seemed inevitable that someone
might try to use machine learning next,

26
00:01:03.680 --> 00:01:08.480
given the amount of free training data Tom
has posted on the internet over the years.

27
00:01:08.480 --> 00:01:11.920
Given that, I decided to get
ahead of you and do it myself.

28
00:01:11.920 --> 00:01:17.200
In fact, you're probably already familiar
with this type of video. It's called a deepfake,

29
00:01:17.200 --> 00:01:18.400
and contrary to what you might think,

30
00:01:18.400 --> 00:01:20.640
it was relatively straightforward
to make this video.

31
00:01:20.640 --> 00:01:23.200
Which is why the whole
world was pretty concerned

32
00:01:23.200 --> 00:01:25.440
about how deepfakes might affect anything

33
00:01:25.440 --> 00:01:28.640
from election results to
general misinformation.

34
00:01:28.640 --> 00:01:30.160
But we're getting a
little ahead of ourselves.

35
00:01:30.160 --> 00:01:32.640
- "After all, what is a deep fake anyway?"

36
00:01:32.640 --> 00:01:35.120
- Well, deepfakes are a
type of synthetic media

37
00:01:35.120 --> 00:01:38.240
or media created using
indirect acquisition methods.

38
00:01:38.240 --> 00:01:40.160
For example, the video that
you're watching right now

39
00:01:40.160 --> 00:01:43.200
is not synthetic media
because it was acquired

40
00:01:43.200 --> 00:01:46.640
through direct means.
By filming me making this video.

41
00:01:46.640 --> 00:01:49.680
- "On the other hand, this clip
of Tom is synthetic media

42
00:01:49.680 --> 00:01:52.080
"because I took video of Tom
and used machine learning

43
00:01:52.080 --> 00:01:54.240
"to create a new one with
things he's never said."

44
00:01:54.240 --> 00:01:56.640
- Now, some types of
synthetic media generation

45
00:01:56.640 --> 00:01:57.920
are easier than others.

46
00:01:57.920 --> 00:02:00.640
For example, I had originally
planned to make this video

47
00:02:00.640 --> 00:02:03.280
by generating a script
using machine learning,

48
00:02:03.280 --> 00:02:05.600
based on past scripts from Tom.

49
00:02:05.600 --> 00:02:07.920
Unfortunately, or rather, fortunately,

50
00:02:07.920 --> 00:02:12.560
there's a difference between generating text
that sounds human and generating text

51
00:02:12.560 --> 00:02:14.880
that actually makes sense,
or is even correct.

52
00:02:14.880 --> 00:02:18.080
- "These robots are made of a lot of
different kinds of sensors,

53
00:02:18.080 --> 00:02:20.000
"and they’re all built to be really precise.

54
00:02:20.800 --> 00:02:24.000
"So that means that you can’t program it by hand.

55
00:02:24.000 --> 00:02:25.520
"And this is a really big problem,

56
00:02:25.520 --> 00:02:29.840
"because the most advanced robots are made
out of metal that can’t be moved precisely.

57
00:02:29.840 --> 00:02:32.880
"You can’t even drive a robot 
arm with a mouse and keyboard."

58
00:02:32.880 --> 00:02:37.360
- The generated text was riddled with
factual errors and nonsensical phrases.

59
00:02:37.360 --> 00:02:39.600
So in the interest of you
actually getting something

60
00:02:39.600 --> 00:02:41.840
out of this video and not
spreading misinformation...

61
00:02:41.840 --> 00:02:45.520
- "I decided to stick with synthetically putting
my own words into Tom's mouth."

62
00:02:45.520 --> 00:02:48.960
- For those with a decent technical background
or willingness to learn,

63
00:02:48.960 --> 00:02:51.600
generating things like deepfakes
are fairly straightforward.

64
00:02:51.600 --> 00:02:55.040
And as to not incentivize
you to do so in this video,

65
00:02:55.040 --> 00:02:58.240
I won't be sharing specific resources
that you can use to do this.

66
00:02:58.240 --> 00:03:01.440
Although I will credit the authors of the
tools that I used in the description.

67
00:03:01.440 --> 00:03:03.760
In most cases, synthetic
media can be generated

68
00:03:03.760 --> 00:03:08.000
using models called Generative
Adversarial Networks or GANs,

69
00:03:08.000 --> 00:03:10.240
which pit two models against each other.

70
00:03:10.240 --> 00:03:14.400
On one side, we have the generator model
which wants to create synthetic data

71
00:03:14.400 --> 00:03:17.520
that is essentially
indistinguishable from real data.

72
00:03:17.520 --> 00:03:19.840
And on the other side, we
have the discriminator model

73
00:03:19.840 --> 00:03:22.640
which wants to accurately
predict whether any given piece

74
00:03:22.640 --> 00:03:25.440
of data is synthesised
from the generator model,

75
00:03:25.440 --> 00:03:27.360
or whether it's a real piece of data.

76
00:03:27.360 --> 00:03:31.280
As the two models go back and forth,
the generator learns to create data

77
00:03:31.280 --> 00:03:34.480
that is increasingly realistic
and the discriminator learns

78
00:03:34.480 --> 00:03:38.720
to find increasingly subtle differences between
that synthetic data and the real data.

79
00:03:38.720 --> 00:03:41.360
- And once you decide that
your generator is good enough,

80
00:03:41.360 --> 00:03:44.480
you can go on to use it to
create synthetic data as needed

81
00:03:44.480 --> 00:03:47.280
typically in the form of
audio, video or images.

82
00:03:47.280 --> 00:03:50.480
- Now, just a few years ago,
generators weren't actually that good

83
00:03:50.480 --> 00:03:53.440
at creating realistic data to
the point where you could look

84
00:03:53.440 --> 00:03:57.280
at synthetic media and pretty easily tell
that something was up.

85
00:03:57.280 --> 00:03:59.600
However, with advances in
machine learning techniques

86
00:03:59.600 --> 00:04:02.800
and increased access
to computing resources,

87
00:04:02.800 --> 00:04:05.040
deepfake generation systems
have gotten to the point where

88
00:04:05.040 --> 00:04:07.920
they can create synthetic
media that you may not

89
00:04:07.920 --> 00:04:10.160
realise is synthetic when you see it.

90
00:04:10.160 --> 00:04:12.880
Understandably with the advent
of synthetic media generation

91
00:04:12.880 --> 00:04:16.320
came a lot of concern
about how synthetic media

92
00:04:16.320 --> 00:04:20.720
might be used to spread misinformation
and literal fake news.

93
00:04:20.720 --> 00:04:22.720
Interestingly, however,
we haven't actually seen

94
00:04:22.720 --> 00:04:25.600
that many examples of
synthetic media being used

95
00:04:25.600 --> 00:04:30.640
to influence things like elections,
international politics or public figures.

96
00:04:30.640 --> 00:04:32.000
- "At least, as far as we know."

97
00:04:32.000 --> 00:04:34.240
- And this is probably
because it turns out

98
00:04:34.240 --> 00:04:37.360
that there are just easier ways
of spreading misinformation.

99
00:04:37.360 --> 00:04:39.920
For example, most of the
cases of claimed interference

100
00:04:39.920 --> 00:04:43.680
in the 2016 and 2020 US
presidential elections came

101
00:04:43.680 --> 00:04:47.840
from things like fake social media accounts,
emails from fake groups

102
00:04:47.840 --> 00:04:51.120
and leaked information
from national political organisations.

103
00:04:51.120 --> 00:04:52.800
This isn't to say that
we shouldn't be worried

104
00:04:52.800 --> 00:04:54.800
about how deepfakes might
be used going forward,

105
00:04:54.800 --> 00:04:57.520
especially considering that
there is a long history

106
00:04:57.520 --> 00:05:01.600
of deepfake generation being
used to create sexual images

107
00:05:01.600 --> 00:05:04.320
and videos without the
consent of the people in them.

108
00:05:04.320 --> 00:05:06.080
However, it is to say
that at least right now,

109
00:05:06.080 --> 00:05:09.120
it seems like the work required
to generate synthetic media

110
00:05:09.120 --> 00:05:12.480
is a little bit more than groups who
are spreading misinformation

111
00:05:12.480 --> 00:05:14.560
nationally and globally are willing to put in.

112
00:05:14.560 --> 00:05:16.480
However, this doesn't mean
that we shouldn't look for ways

113
00:05:16.480 --> 00:05:18.400
to identify synthetic media.

114
00:05:18.400 --> 00:05:20.640
In fact, recent research
has shown that the methods

115
00:05:20.640 --> 00:05:22.960
that we use to generate
synthetic media actually

116
00:05:22.960 --> 00:05:25.520
leave fingerprints on the media itself,

117
00:05:25.520 --> 00:05:27.920
which can later be
identified using algorithms.

118
00:05:27.920 --> 00:05:31.360
Even better, it turns out that we
as living, breathing human beings,

119
00:05:31.360 --> 00:05:33.520
leave our own fingerprints in video data

120
00:05:33.520 --> 00:05:37.040
that can't currently be replicated
via synthetic media generation.

121
00:05:37.040 --> 00:05:40.000
And I did a whole video on this, so if it
sounds interesting and you want to learn more,

122
00:05:40.000 --> 00:05:41.440
I will leave a link in the description.

123
00:05:41.440 --> 00:05:46.000
It's also important to note that there are some
proposed positive applications of deepfakes.

124
00:05:46.000 --> 00:05:50.560
In particular, there's a lot of interest in using
deepfakes to dub media more realistically,

125
00:05:50.560 --> 00:05:52.800
so that people from other countries,

126
00:05:52.800 --> 00:05:55.200
who speak other languages
or have other cultures,

127
00:05:55.200 --> 00:05:59.120
can experience things that they might not
have access to otherwise.

128
00:05:59.120 --> 00:06:01.120
So while we keep trying to find new ways

129
00:06:01.120 --> 00:06:04.000
to identify deepfakes in
the wild, we can also look

130
00:06:04.000 --> 00:06:07.200
towards positive applications
of them in our everyday lives.

131
00:06:07.200 --> 00:06:11.120
Unfortunately, however, the nature
of Generative Adversarial Networks means

132
00:06:11.120 --> 00:06:13.840
that deepfake detection
and deepfake generation

133
00:06:13.840 --> 00:06:16.160
will always be a bit of
a cat-and-mouse game.

134
00:06:16.160 --> 00:06:18.960
After all, as we develop better
deepfake detection systems,

135
00:06:18.960 --> 00:06:21.280
deepfake generators
also improve with them,

136
00:06:21.280 --> 00:06:23.760
learning to create that
more realistic media.

137
00:06:23.760 --> 00:06:25.120
There are also other proposed methods

138
00:06:25.120 --> 00:06:27.920
of identifying deepfakes,
such as embedding signals

139
00:06:27.920 --> 00:06:30.240
in anything that could
be used as training data

140
00:06:30.240 --> 00:06:32.880
that would be preserved
through the synthesis process,

141
00:06:32.880 --> 00:06:36.560
similar to the fingerprints that synthetic
media generation techniques already leave.

142
00:06:36.560 --> 00:06:38.480
And who knows, maybe we
will all start having

143
00:06:38.480 --> 00:06:41.520
to wear necklaces or
pins that emit signals

144
00:06:41.520 --> 00:06:45.120
that automatically poison any sort of
recording that someone might be taking.

145
00:06:45.120 --> 00:06:47.600
So while you might not be
able to spot synthetic media

146
00:06:47.600 --> 00:06:50.080
with your eyes and ears going forward,

147
00:06:50.080 --> 00:06:53.200
websites and other platforms
that host media may be able

148
00:06:53.200 --> 00:06:55.840
to deploy tools that may
still be able to let you know

149
00:06:55.840 --> 00:06:57.360
that what you're watching isn't real.

150
00:06:57.360 --> 00:07:00.000
- "After all, I've told you
that I'm not Tom Scott."

151
00:07:00.000 --> 00:07:01.840
- Who's to say that I'm Jordan either?

152
00:07:01.840 --> 00:07:03.680
- Go check out Jordan's channel!

153
00:07:03.680 --> 00:07:05.200
She doesn't do clickbait videos,

154
00:07:05.200 --> 00:07:08.960
she just explains concepts
about AI really well.

155
00:07:08.960 --> 00:07:11.920
You know, when she's not doing
potentially lifesaving medical research.

156
00:07:11.920 --> 00:07:13.520
I am not doing enough with my life.

157
00:07:13.520 --> 00:07:16.720
Try her video about AI detecting
COVID from your cough.

158
00:07:16.720 --> 00:07:18.720
Next week, on the final guest video,

159
00:07:18.720 --> 00:07:22.240
a little bit of radioactivity.