WEBVTT
Kind: captions
Language: en

1
00:00:00.120 --> 00:00:03.330
I was going to tell a science fiction story
about mass blackmail.

2
00:00:03.330 --> 00:00:04.750
It had a protagonist and a narrative,

3
00:00:04.750 --> 00:00:06.350
and a beginning, middle and end,

4
00:00:06.350 --> 00:00:07.569
and a warning about the future.

5
00:00:07.569 --> 00:00:09.510
And it was going to start with a fictional app,

6
00:00:09.510 --> 00:00:12.660
an app that let you swap your face
into movie scenes.

7
00:00:12.660 --> 00:00:14.509
So, take a clip from a movie,

8
00:00:14.509 --> 00:00:17.060
emote a bit into your smartphone's camera,

9
00:00:17.060 --> 00:00:18.330
and it would put you in the scene.

10
00:00:18.330 --> 00:00:20.390
I actually can't believe
no-one's built that yet.

11
00:00:20.390 --> 00:00:25.150
I was going to explore how that technology
could be used and misused and abused...

12
00:00:25.150 --> 00:00:26.680
And then the news happened,

13
00:00:26.680 --> 00:00:30.290
and I realised the future had arrived a lot
faster than I thought it would.

14
00:00:30.290 --> 00:00:35.969
Just so you know, I'm not going to show anything
adult-rated or ethically questionable in this video.

15
00:00:35.969 --> 00:00:41.510
All the faceswapping that I've been working
on has been entirely clothed and consensual.

16
00:00:41.510 --> 00:00:47.339
So when I saw that story, I had to find out:
is this a real bit of code?

17
00:00:47.339 --> 00:00:51.800
Has Hollywood-grade faceswapping become
available to the world that quickly?

18
00:00:51.800 --> 00:00:53.201
So I went and found the software itself.

19
00:00:53.201 --> 00:00:56.879
And it turns out the raw code for this has
been around for a couple of months;

20
00:00:56.879 --> 00:01:03.760
and it had been developed and used pretty much
exclusively for creating non-consensual pornography.

21
00:01:03.760 --> 00:01:06.260
Taking the face of someone who hasn't consented

22
00:01:06.260 --> 00:01:10.830
and putting it on the body of someone else
who hasn't consented.

23
00:01:10.830 --> 00:01:14.810
Last week, someone decided to package up this
complicated bit of code,

24
00:01:14.810 --> 00:01:19.760
this whole massive project that required large
amounts of technical skill and knowledge,

25
00:01:19.760 --> 00:01:24.920
and turn it into a Windows app that anyone
with a bit of time and a bit of tech skills could use.

26
00:01:24.920 --> 00:01:26.250
The system requirements were:

27
00:01:26.250 --> 00:01:31.130
a Windows 10 PC, a powerful Nvidia graphics card,
moderate tech skills

28
00:01:31.130 --> 00:01:32.791
and poor or no sense of empathy.

29
00:01:32.791 --> 00:01:37.200
And if you want to find someone with a fast
computer and no empathy,

30
00:01:37.200 --> 00:01:38.200
you of course go to Reddit.

31
00:01:38.200 --> 00:01:42.450
I found the software there, downloaded it,
and realised I wasn't going to run

32
00:01:42.450 --> 00:01:49.140
2.6 gigabytes of unknown code, designed to
create unethical porn, on my own computer.

33
00:01:49.140 --> 00:01:52.680
So I dug out an old machine, wiped it clean,
and ran the code.

34
00:01:52.680 --> 00:01:58.290
And it gave me... just a box, really, with
unfriendly error messages and no help files.

35
00:01:58.290 --> 00:02:03.210
It reminds me a lot of the hacked-together
Windows stuff I used to code many years ago,

36
00:02:03.210 --> 00:02:06.030
only mine was for creating steps for
Dance Dance Revolution games.

37
00:02:06.030 --> 00:02:08.550
So you know, bit of a different audience there.

38
00:02:08.550 --> 00:02:12.640
Probably only a slightly different audience,
actually, but...

39
00:02:12.640 --> 00:02:16.459
To make the code work, you have to assemble
images of the source and the target.

40
00:02:16.459 --> 00:02:20.020
So I asked my friend Matt Parker if I could
use his videos as an example,

41
00:02:20.020 --> 00:02:21.810
and he was fine with it.

42
00:02:21.810 --> 00:02:26.269
So I extracted 500 frames from
his YouTube channel, and 500 from mine.

43
00:02:26.269 --> 00:02:28.700
The first thing the software does is called
extraction:

44
00:02:28.700 --> 00:02:30.860
get all the faces from those images,

45
00:02:30.860 --> 00:02:34.120
and put the eyes, nose and mouth
in the same place for each.

46
00:02:34.120 --> 00:02:37.060
And you might think,
"Tom, did you make a creepy GIF out of this?"

47
00:02:37.060 --> 00:02:39.920
And the answer is yes, of course I did.

48
00:02:39.920 --> 00:02:42.140
Two, actually.

49
00:02:42.140 --> 00:02:44.700
I had slightly more hair then.

50
00:02:44.700 --> 00:02:46.849
Mind you, so did Matt.

51
00:02:46.849 --> 00:02:50.140
After it had finished, the computer was still
kind-of churning away,

52
00:02:50.140 --> 00:02:51.420
working on something in the background

53
00:02:51.420 --> 00:02:52.909
and I was a little bit worried about what
it was doing,

54
00:02:52.909 --> 00:02:54.670
so I had a look in the Task Manager.

55
00:02:54.670 --> 00:02:59.299
It turns out Windows 10 was trying to update,
so I turned that off.

56
00:02:59.299 --> 00:03:00.499
And it still updated.

57
00:03:00.499 --> 00:03:02.219
And rebooted itself.

58
00:03:02.219 --> 00:03:03.219
Twice.

59
00:03:03.219 --> 00:03:05.219
Anyway, I got it training.

60
00:03:05.219 --> 00:03:09.440
And as it trains, it shows you a load of faces
that look like some sort of 'wall of ham'.

61
00:03:09.440 --> 00:03:13.580
The right one of each triple are what it's
adjusted the face into, what you're looking for.

62
00:03:13.580 --> 00:03:15.000
After 10 minutes, they were rough;

63
00:03:15.010 --> 00:03:17.260
after an hour, they were starting to become
recognisable;

64
00:03:17.260 --> 00:03:19.090
but then it all slowed down.

65
00:03:19.090 --> 00:03:23.780
After 24 hours of whirring fans and a nearly-overheating
graphics card, it was still blurred;

66
00:03:23.780 --> 00:03:27.180
but after 48 hours, I thought that some of
it might be good enough to try.

67
00:03:27.180 --> 00:03:32.720
So here's the question: with 48 hours of training,
on a graphics card from a couple of years ago,

68
00:03:32.720 --> 00:03:36.820
can you convincingly faceswap Matt Parker
into this video?

69
00:03:38.460 --> 00:03:39.460
No.

70
00:03:39.460 --> 00:03:40.819
No, not really.

71
00:03:40.819 --> 00:03:42.769
Not at all.

72
00:03:42.769 --> 00:03:46.180
The other question is, can you convincingly
faceswap the other way around?

73
00:03:46.180 --> 00:03:48.440
And the answer is still no.

74
00:03:50.840 --> 00:03:54.860
So I was trying to figure out: how can I demonstrate
to an audience on YouTube

75
00:03:54.860 --> 00:03:57.319
that the code is a real thing?

76
00:03:57.319 --> 00:04:00.010
And then this news article came along.

77
00:04:00.010 --> 00:04:04.460
I don't know what the copyright situation
is on what I'm about to show you, but...

78
00:04:07.249 --> 00:04:12.780
To try and make sure that this counts as criticism
and review: that faceswap is perfect.

79
00:04:12.790 --> 00:04:15.360
The lighting matches.
The expression matches.

80
00:04:15.360 --> 00:04:21.819
That looks like a Hollywood CGI team spent an
enormous amount of time and money changing a face,

81
00:04:21.819 --> 00:04:24.800
instead of one person taking a couple of days
on their own computer.

82
00:04:24.800 --> 00:04:27.100
The user in question said this:

83
00:04:28.920 --> 00:04:31.590
Try explaining that sentence
to someone 20 years ago.

84
00:04:31.590 --> 00:04:35.800
Actually, never mind copyright,
is that ethically okay?

85
00:04:35.800 --> 00:04:40.140
Neither of the people involved in that clip,
Amy Adams or Nicolas Cage, have consented to it.

86
00:04:40.140 --> 00:04:44.120
I think it's just about OK, but then everyone
draws the moral line of what's acceptable

87
00:04:44.120 --> 00:04:47.530
just slightly below
what they're actually doing.

88
00:04:47.530 --> 00:04:51.160
Maybe even faceswapping people
like that isn't OK.

89
00:04:51.160 --> 00:04:53.780
Making clips like that
is only going to get easier.

90
00:04:53.780 --> 00:04:57.060
Yes, it requires days of training now, but
it won't for long.

91
00:04:57.060 --> 00:05:02.550
Someone's going to come along and package up
an algorithm in a way that doesn't seem questionable:

92
00:05:02.550 --> 00:05:07.660
it'll be a fun app about putting your face
in movies rather than the current horrible use.

93
00:05:07.660 --> 00:05:09.560
That's just a question of marketing.

94
00:05:09.560 --> 00:05:15.600
As of last week, you can no longer trust a video
of someone unless you also trust its source.

95
00:05:15.600 --> 00:05:18.700
If you wanted to fake a certain
alleged Donald Trump tape,

96
00:05:18.710 --> 00:05:23.490
then all you need is a body double, a few
days of compute time, and no conscience.

97
00:05:23.490 --> 00:05:26.990
The story I was going to tell
was about algorithmic blackmail.

98
00:05:26.990 --> 00:05:31.710
The idea that someone could write a script
to pull out selfies from Instagram accounts,

99
00:05:31.710 --> 00:05:36.940
and train up an algorithm to create plausible
fake porn, as if from a hidden or hacked webcam,

100
00:05:36.940 --> 00:05:40.560
and then send out blackmail threats automatically.

101
00:05:40.560 --> 00:05:46.390
And even if only a few people fall for it,
that could still be enough to make a decent income,

102
00:05:46.390 --> 00:05:49.990
particularly if the sender's in a developing
country with a lower cost of living.

103
00:05:49.990 --> 00:05:55.100
All they have to do is cause emotional trauma
to a huge number of people.

104
00:05:55.100 --> 00:05:57.660
But they're just people on the internet, right?

105
00:05:57.660 --> 00:05:59.180
They don't count(!)

106
00:06:00.140 --> 00:06:02.710
We're not ready for this.
As society, as humanity.

107
00:06:02.710 --> 00:06:06.400
Normally when I tell a story like this I try
to end on an upbeat note:

108
00:06:06.400 --> 00:06:10.349
the people in the sci-fi tale that I had to ditch,
they sort-of ended up okay in the end,

109
00:06:10.349 --> 00:06:13.570
because it's fiction and you can
make sure that happens.

110
00:06:13.570 --> 00:06:15.879
But this isn't science fiction any more.

111
00:06:15.879 --> 00:06:18.120
I don't have an upbeat note to end on.

112
00:06:18.120 --> 00:06:20.700
All I have is a vague sense of future shock,

113
00:06:20.700 --> 00:06:23.860
and 500 distorted images of Matt Parker.

114
00:06:31.880 --> 00:06:34.600
That would've been a lot easier without the
dogs barking in the background.