WEBVTT

1
00:00:01.830 --> 00:00:03.770
I am not Morgan Freeman.

2
00:00:03.830 --> 00:00:05.970
And what you see is not real.

3
00:00:06.470 --> 00:00:07.740
What is real?

4
00:00:07.750 --> 00:00:09.020
And what's fake?

5
00:00:09.030 --> 00:00:11.660
Working that out
is increasingly important.

6
00:00:11.670 --> 00:00:16.630
Deepfakes, manipulated videos, photos
and voices are becoming more common.

7
00:00:18.360 --> 00:00:20.390
AI has ushered in a 
whole new level of fakes

8
00:00:20.390 --> 00:00:23.380
and it's becoming more difficult 
to tell what's true and what's false.

9
00:00:25.332 --> 00:00:30.240
And AI also makes it much easier
to create fake images and videos.

10
00:00:31.990 --> 00:00:36.630
We're confronted by a flood of images
every day and more and more of them

11
00:00:36.640 --> 00:00:38.440
have been manipulated.

12
00:00:41.840 --> 00:00:45.840
Like any tool, AI can be used 
and also abused.

13
00:00:45.960 --> 00:00:49.680
We look at these potentially 
abusive uses and try to counteract them.

14
00:00:51.560 --> 00:00:53.860
To detect AI-generated fakes,

15
00:00:53.860 --> 00:00:58.027
researchers at the Fraunhofer Institute 
for Applied  and Integrated Security

16
00:00:58.027 --> 00:01:00.553
 employ another AI technology:

17
00:01:00.553 --> 00:01:02.153
 super machine learning.

18
00:01:02.200 --> 00:01:05.393
By studying many examples of 
audio and video tracks,

19
00:01:05.393 --> 00:01:09.760
it can learn to detect patterns
 that allow it to flag up faked content.

20
00:01:10.410 --> 00:01:12.980
This kind of AI support is crucial.

21
00:01:13.000 --> 00:01:16.640
Because before long it will be
effectively impossible for us humans

22
00:01:16.640 --> 00:01:19.420
to recognize deepfakes.

23
00:01:19.440 --> 00:01:23.200
And not every deepfake is as harmless
as this one,

24
00:01:23.200 --> 00:01:24.956
in which former German Chancellor 
Angela Merkel

25
00:01:24.956 --> 00:01:27.822
appears to recite some comic verse.

26
00:01:30.633 --> 00:01:33.400
Deepfake detection is
a race against time.

27
00:01:33.400 --> 00:01:36.819
And that's mainly because the
aggressors are upping their game.

28
00:01:36.830 --> 00:01:40.280
AI models don't get better on their own, 
but because humans put effort

29
00:01:40.280 --> 00:01:41.660
into generating them.

30
00:01:41.840 --> 00:01:45.720
So we have to find better ways
to uncover and detect fakes

31
00:01:45.720 --> 00:01:47.300
and push further development.

32
00:01:49.400 --> 00:01:52.840
Similar technology is also
widely used in advertising.

33
00:01:52.920 --> 00:01:56.200
Our willingness to believe the
unbelievable is something

34
00:01:56.200 --> 00:01:58.600
Claudia Bussjaeger has observed 
for many years.

35
00:02:01.100 --> 00:02:03.030
People want to believe 
what they see.

36
00:02:03.040 --> 00:02:05.750
And what they don't want to see,
they also don't want to believe.

37
00:02:05.760 --> 00:02:06.630
They're in a bubble.

38
00:02:08.507 --> 00:02:12.720
Bussjaeger runs the first
German-based platform for AI artists.

39
00:02:15.587 --> 00:02:18.387
Before that, she worked 
in advertising,

40
00:02:18.387 --> 00:02:22.419
an industry which often 
stretches the frontiers of truth.

41
00:02:26.370 --> 00:02:29.190
People have been faking it 
in advertising for ages.

42
00:02:29.200 --> 00:02:33.240
There are no ads that haven't been
edited or had parts swapped out.

43
00:02:33.600 --> 00:02:35.230
But nobody questions that.

44
00:02:35.240 --> 00:02:36.700
Nobody ever has.

45
00:02:40.480 --> 00:02:44.000
Images can have power,
we've known that for a long time.

46
00:02:44.040 --> 00:02:48.200
Just like faked texts, when
manipulated images can be used

47
00:02:48.200 --> 00:02:49.870
to spread misinformation.

48
00:02:49.880 --> 00:02:52.419
Everywhere, from art to politics.

49
00:02:52.760 --> 00:02:57.040
Soviet dictator Joseph Stalin
notoriously had those he fell out with

50
00:02:57.040 --> 00:02:59.030
erased from photographs.

51
00:02:59.040 --> 00:03:02.680
Such manipulation is problematic 
for a number of reasons.

52
00:03:05.560 --> 00:03:09.327
The public has to be able 
to rely on the information it gets

53
00:03:09.327 --> 00:03:13.139
because we all need to find out
about the world we live in.

54
00:03:14.760 --> 00:03:16.770
That includes journalistic information,

55
00:03:16.770 --> 00:03:18.780
which plays a role in that process.

56
00:03:19.040 --> 00:03:22.360
And if we can no longer trust 
theinformation that we're given,

57
00:03:22.360 --> 00:03:23.900
we have a fundamental problem.

58
00:03:25.720 --> 00:03:29.280
That's actually quite a frightening
development when you think about it.

59
00:03:31.680 --> 00:03:35.720
We can create things that are
completely, or almost completely,

60
00:03:35.720 --> 00:03:40.400
detached from reality, and we can
manipulate them in a targeted way,

61
00:03:40.400 --> 00:03:42.350
for our own gains.

62
00:03:42.360 --> 00:03:44.340
And that can be dangerous.

63
00:03:47.160 --> 00:03:50.153
Because if everything 
can theoretically be faked,

64
00:03:50.153 --> 00:03:52.270
how do we know 
what hasn't been?

65
00:03:52.280 --> 00:03:55.540
Images stop being a
reliable source of evidence.

66
00:03:55.640 --> 00:03:58.967
And yet at the same time, 
these developments bring new opportunities

67
00:03:58.967 --> 00:04:00.610
and new roles.

68
00:04:00.620 --> 00:04:04.473
AI artists, for example, 
use the new technology to create art:

69
00:04:05.667 --> 00:04:08.577
I think there's no stopping it.
Even if you say:

70
00:04:08.577 --> 00:04:10.513
"AI is not for me!"

71
00:04:10.513 --> 00:04:12.680
Sooner or later it will be.

72
00:04:12.680 --> 00:04:15.420
It's always been like that, 
throughout history.

73
00:04:15.720 --> 00:04:18.760
When a new technology arrives, 
others disappear.

74
00:04:19.200 --> 00:04:22.960
We no longer have landlines, 
and one day we won't have cars

75
00:04:22.960 --> 00:04:24.550
that run on fossil fuels.

76
00:04:24.560 --> 00:04:26.220
That's just how it is.

77
00:04:26.240 --> 00:04:30.320
So we need to ask: how can
I use AI in a positive way?

78
00:04:30.510 --> 00:04:33.290
So that it's not
twisted and used destructively?

79
00:04:36.480 --> 00:04:39.520
Like in the grandparent
phone scam, for example.

80
00:04:39.520 --> 00:04:43.560
Just a few words of recorded speech
are now enough to clone a voice,

81
00:04:43.570 --> 00:04:46.920
which can then be used by
scammers to confuse elderly relatives

82
00:04:46.920 --> 00:04:48.180
into giving them money.

83
00:04:48.610 --> 00:04:50.890
A more positive use of
the same technology

84
00:04:50.890 --> 00:04:52.779
is being developed by Google.

85
00:04:56.200 --> 00:04:59.186
It enables speech-impaired people 
to communicate

86
00:04:59.186 --> 00:05:01.300
using  their own voices again.

87
00:05:04.550 --> 00:05:09.750
It uses the same technology that
can be misused to make deepfakes.

88
00:05:09.750 --> 00:05:13.950
And that shows, that as a technology,
AI itself is morally neutral.

89
00:05:13.950 --> 00:05:16.570
It's down to us humans 
and how we use it.

90
00:05:18.000 --> 00:05:20.760
That's key when it comes to
protecting what's true

91
00:05:20.760 --> 00:05:23.250
and also to protecting credibility.

92
00:05:23.279 --> 00:05:27.390
Media companies use fact checks and
multiple-source checks as safeguards

93
00:05:27.400 --> 00:05:30.720
in what is still largely 
an unregulated zone.

94
00:05:34.560 --> 00:05:38.240
Currently we are in the
Wild West as far as AI is concerned.

95
00:05:38.560 --> 00:05:42.130
We're seeing all kinds of different
players jumping on the bandwagon,

96
00:05:42.140 --> 00:05:45.300
like golddiggers who want
to use it for their own ends.

97
00:05:46.760 --> 00:05:50.640
The legal framework
is only being set up now.

98
00:05:50.640 --> 00:05:55.100
And as is often the case, 
the law is slower than advances,

99
00:05:55.100 --> 00:05:57.540
it's behind technological developments.

100
00:05:59.040 --> 00:06:01.500
So politicians need to get into gear.

101
00:06:01.800 --> 00:06:04.640
They have to step up 
to make sure that we don't have

102
00:06:04.640 --> 00:06:06.300
these problems in future.

103
00:06:11.279 --> 00:06:15.000
Some countries have now signed
a legally binding treaty aimed

104
00:06:15.000 --> 00:06:17.150
at regulating the use of AI.

105
00:06:17.160 --> 00:06:19.892
But until and if it 
finally comes into effect,

106
00:06:19.892 --> 00:06:22.630
responsibility remains with companies.

107
00:06:22.640 --> 00:06:26.720
Claudia Bussjaeger has drawn up
ethical guidelines for her platform.

108
00:06:29.920 --> 00:06:32.720
Clearly, we don't take people 
and disparage them.

109
00:06:33.400 --> 00:06:36.540
We don't imitate artists
and pretend to be them.

110
00:06:37.000 --> 00:06:40.680
We have very clear ethics 
when it comes to dealing with property

111
00:06:40.680 --> 00:06:43.680
belonging to people
 and artists and celebrities.

112
00:06:46.753 --> 00:06:51.400
You can basically say that it's about
staying real in an unreal world.

113
00:06:52.567 --> 00:06:54.560
I think that's really important.

114
00:06:58.707 --> 00:07:02.520
There are also a number of
newly emerging apps and platforms

115
00:07:02.520 --> 00:07:05.600
which focus on revealing 
AI-generated fakes.

116
00:07:05.920 --> 00:07:09.087
The Fraunhofer Institute 
for Applied and Integrated Security

117
00:07:09.087 --> 00:07:15.160
 offers the "Deepfake Total" website, 
which can be used by anyone free of charge.

118
00:07:18.680 --> 00:07:22.120
Dealing with deepfakes is
partly about media literacy.

119
00:07:22.120 --> 00:07:24.433
That means questioning 
what you see online,

120
00:07:24.433 --> 00:07:26.630
and not just taking it 
at face value.

121
00:07:26.640 --> 00:07:30.600
It's also about using
technology to uncover deepfakes.

122
00:07:30.600 --> 00:07:32.413
And last but not least,

123
00:07:32.413 --> 00:07:35.750
it's about implementing 
protective verification methods.

124
00:07:35.760 --> 00:07:38.320
You can protect websites 
with  digital signatures,

125
00:07:38.320 --> 00:07:40.750
and you can do the same 
with media content.

126
00:07:40.760 --> 00:07:43.180
If we use these 
three building blocks,

127
00:07:43.180 --> 00:07:45.660
I think we'll be 
well-positioned as societies.

128
00:07:47.840 --> 00:07:51.840
We've become used to living in
a world flooded with information.

129
00:07:51.840 --> 00:07:55.960
Now we also need 
to develop a critical eye,

130
00:07:55.960 --> 00:07:59.720
to avoid falling for increasingly 
deceptive deepfakes in future.