Most of us are convinced that we can dissociate humans from machines, but is it really the case?
This study dives into this question using AI-generated profiles on dating apps. Would you swipe right for an AI? Let's find out.
Watch the video
References
This AI Reads Your Brain to Generate Personally Attractive Faces, https://youtu.be/-ZibKcoW-7k
GPT-2 Colab Notebook, https://colab.research.google.com/dri...
[1] Sandra Bryant et al., "We Asked Artificial Intelligence to Create Dating Profiles. Would You Swipe Right?", (2021), UNSW Sydney blog, https://studyonline.unsw.edu.au/blog/....
[2] M. Spape, K. Davis, L. Kangassalo, N. Ravaja, Z. Sovijarvi-Spape and T. Ruotsalo, "Brain-computer interface for generating personally attractive images," in IEEE Transactions on Affective Computing, doi: 10.1109/TAFFC.2021.3059043.
[3] Tero Karras and Samuli Laine and Miika Aittala and Janne Hellsten and Jaakko Lehtinen and Timo Aila, "Analyzing and Improving the Image Quality of StyleGAN", (2019), published at CVPR, by NVIDIA.
[4] Tero Karras and Samuli Laine and Miika Aittala and Janne Hellsten and Jaakko Lehtinen and Timo Aila, "Training Generative Adversarial Networks with Limited Data", (2020), published at NeurIPS, by NVIDIA.
[5] Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya, "Language Models are Unsupervised Multitask Learners", (2019), by OpenAI.
[6] Tom B. Brown et al., "Language Models are Few-Shot Learners", (2020), arXiv, by OpenAI.
Chapters
0:00 Hey! If you see this, let me know if you liked the longer format of this video!
1:06 A super-intelligent machine... ft. Homo Deus
2:57 The study
3:47 Key findings
4:19 What is an AI-made-up person?
5:13 How to create a human-looking face using StyleGAN2
6:53 How to create an identity using GPT-2
8:42 The results
9:46 Don't forget to subscribe. You'll learn a lot of cool stuff, I promise!
1:06 A super-intelligent machine... ft. Homo Deus
2:57 The study
3:47 Key findings
4:19 What is an AI-made-up person?
5:13 How to create a human-looking face using StyleGAN2
6:53 How to create an identity using GPT-2
8:42 The results
9:46 Don't forget to subscribe. You'll learn a lot of cool stuff, I promise!
Video Transcript
00:00
would you swipe right on an ai profile
00:02
most of us are convinced that we can
00:04
dissociate
00:05
humans from machines but is it really
00:08
the case
00:09
this is what the study reveals using ai
00:11
meta people on dating apps
00:14
as we discussed in the last video it's
00:15
now possible to train an artificial
00:17
intelligence to muddle what beauty is to
00:20
us
00:20
within a couple years the famous saying
00:23
beauty is in the eye of the beholder
00:25
may have to change to beauty is in the
00:27
ai of the beholder
00:29
indeed beauty is something rather
00:31
complex and subjective
00:33
in a recent study researchers showed
00:35
that artificial intelligence analyzing
00:37
your tastes and interests could generate
00:40
personally attractive faces i think you
00:43
will all agree that when you look at
00:44
someone and find the person attractive
00:46
you cannot really explain
00:48
why there may be many reasons involved
00:51
in this decision plus some of these
00:53
reasons may not even be directly related
00:56
to the physique of the person
00:58
it seems like these reasons may all be
01:00
electrical signals
01:01
and chemical mixtures in your brain
01:03
making you feel this way
01:05
if that's the case this data can be
01:07
processed and understood much more
01:09
effectively by a powerful computation
01:11
machine
01:12
then by ourselves of course this will
01:14
only be possible
01:16
considering a super intelligent machine
01:18
with access to all our personal data in
01:20
real time
01:22
such a machine will know even more about
01:24
you than yourself
01:25
able to explain its decision process
01:27
better than you are
01:28
give the exact reasons why this person
01:31
is attractive to you
01:32
and find your perfect match it will use
01:34
data that you don't even have
01:36
access to since it already knows the
01:38
other person
01:39
and won't be affected by temporary bias
01:42
caused by your feelings
01:44
such a google powered machine could
01:46
exist in the future
01:47
as uval noah harari will suggest in his
01:50
amazing book
01:51
homodes this thought probably sounds
01:54
extremely weird or even scary at first
01:57
but you have to consider that many
01:59
things will have to change in our
02:00
society to make this happen
02:02
and we are not even talking about the
02:04
trust issue arising
02:06
with such a technology that we first
02:08
need to face
02:09
anyways is this even a good thing do we
02:12
want an ai to tell us
02:13
what to like or even tell us that we are
02:16
wrong and show us what we really like
02:19
i don't know any of my relatives that
02:21
will enjoy following a super intelligent
02:23
machine
02:23
even if it was proven to know us better
02:25
than we do
02:27
for example such a machine could be
02:29
analyzing you constantly
02:31
including when you are dating someone
02:33
and it will be providing you with your
02:35
true
02:35
feelings regarding your date thus using
02:38
probabilities
02:39
statistics and a lot of data it could
02:42
judge
02:43
your two previous dates and come up with
02:45
the best match for you
02:47
now knowing that it's most certainly the
02:49
best decision for you considering all
02:51
the elements taking into account
02:53
the question is will you follow it
02:55
blindly
02:56
in the meantime researchers are
02:58
analyzing how human-like artificial
03:00
intelligences can be
03:02
most of us are convinced that we can
03:04
dissociate an
03:05
actual human from a machine but is it
03:08
truly the case
03:09
this is what this study reveals using ai
03:12
made up people
03:13
on dating apps indeed it's just like the
03:16
turing test
03:17
but on dating apps sandra bryant and her
03:19
team at unsw sydney
03:22
generated made-up people using
03:24
artificial intelligence to measure
03:25
precisely
03:26
how much of what you see on a person's
03:29
dating profile do you consider real
03:32
before we dive into the crunchy details
03:34
of this exciting research
03:36
let's watch a little overview of what
03:38
they found as it may peak your interest
03:40
as it did for me here are two key
03:43
findings of this study on 1008 singles
03:46
surveyed
03:47
they found that 62 percent of single
03:49
people will swipe right on an ai
03:52
generated dating profile
03:53
and 43 of these respondents said these
03:57
profiles looked real
03:58
this means that approximately half the
04:00
people cannot distinguish an
04:02
ai from an actual human based on a
04:05
person's looked and short description
04:07
of course this survey only represents a
04:10
small fraction of the population
04:11
being only 1000 single people using
04:14
dating apps
04:15
which does not correctly represent the
04:17
overall population
04:19
in their experiment an ai made up person
04:22
looked like this
04:23
well not exactly that will pretty much
04:25
give it away it more of looked like this
04:28
as you can see amy looks like a real
04:30
person and i'm not only talking about
04:32
her appearance
04:33
but even her presence seem real as they
04:36
mentioned
04:37
everything from their faces to their
04:39
words and identities were completely
04:41
constructed by artificial intelligence
04:44
indeed amy has an utterly fake identity
04:48
using a short and original description
04:50
and genuinely answering generic
04:52
questions asked by the dating app
04:54
it's even using humor and imagination
04:57
two things we will not suspect a machine
04:59
to have
04:59
this time soon it's pretty astonishing
05:02
in order to create such humanized fake
05:04
profiles
05:05
they had to achieve two main objectives
05:08
create a real looking profile pictures
05:11
and create a new identity for the farmer
05:14
the team used this person does not
05:17
exist.com
05:18
a website that generates not existing
05:21
people
05:22
randomly when you access or refresh it
05:24
the system uses a method
05:26
called stylegantu published in late
05:29
2019.
05:30
stargand2 and its most recent variation
05:33
sargent ii
05:34
ada are state-of-the-art generative
05:36
adversarial networks or gun
05:38
a gun is a neural network architecture
05:41
able to generate a completely artificial
05:43
image
05:43
following the style of the image it was
05:46
trained on thus replicating
05:48
what it has seen during its training to
05:50
generate new images
05:51
that don't even exist like human faces
05:54
in this case
05:56
here the style of the image can be
05:58
interpreted differently
05:59
in fact it can be adjusted to not only
06:01
replicate the overall style of the image
06:04
which in this case are pictures of human
06:06
faces but also to learn the style of
06:08
local features such as a nose eyes
06:10
hair etc which can also be controlled or
06:13
even randomized
06:14
during the generation process this is
06:17
possible because of the gann
06:18
architecture's power
06:20
where it learns by visually comparing
06:22
the results between the generated images
06:25
and the images that was supposed to
06:27
generate
06:28
such supposed or expected images are the
06:31
ones found in the training data
06:33
before entering into the second part of
06:35
these fake profiles responsible for
06:37
creating the fake identity i would like
06:39
to refer the most technical of you to a
06:41
previous video explaining guns
06:44
this 5-minute video will help you
06:46
understand how it works
06:48
with more technical details using a
06:50
great cartoonify example
06:52
the profiles were then made alive using
06:55
gpt2
06:56
a text generating model developed by
06:58
openai in 2019
07:00
you've undoubtedly heard of his big
07:02
brother gpdt
07:03
well the advantage of using gpt2 is that
07:06
it is open access and quite powerful
07:09
indeed you can try it yourself without
07:11
using a powerful computer or graphic
07:13
card
07:14
just by running the few lines in this
07:16
google collab notebook
07:18
which is linked in the description and
07:20
generate text for
07:22
any task you have data to train on even
07:25
better
07:25
if you don't have any data you can use
07:28
pre-trained models which are models that
07:30
have already been trained
07:31
that you can upload and run right away
07:34
to explain it simply
07:35
gpt 2 works by looking at hundreds of
07:38
thousands of text examples of your
07:39
choice
07:40
to generate by itself a brand new text
07:43
following again
07:44
the same style as the training data in
07:47
this case the training data
07:48
is text and the style is a dating
07:50
profile type of text
07:52
what is fun about this model is that it
07:54
is very intuitive
07:56
we can see that gpt2 finds its
07:58
inspiration by looking at
08:00
all the other dating profiles to find
08:03
winning
08:03
patterns and then tries to replicate
08:05
them in a brand
08:06
new dating profile that never existed
08:09
before
08:10
remember this study was about using
08:12
these fake profiles on dating apps to
08:14
study
08:15
how real life singles swiped on the fake
08:18
images
08:18
as well as their overall perceptions of
08:21
ai generated matches
08:23
before getting into this next section
08:25
here's a video i made about gpd2's big
08:27
brother gpt3
08:29
for the most curious among you
08:31
showcasing some impressive applications
08:33
and explaining how it works
08:35
since gpt3 is pretty much an upgraded
08:37
version of its predecessor
08:39
it is highly similar it just has a much
08:42
bigger memory meet michael a
08:45
non-existing adventurous 29 year old
08:47
oakland programmer
08:49
for michael 44 of respondents said that
08:52
he looked
08:52
real and 82 said that it was highly
08:55
attractive
08:56
with a subjective rating of more than 7
08:59
out of 10.
09:00
with this in mind 73.6 percent of people
09:04
said that they will swipe right seeing
09:06
this profile
09:07
similarly for amy 86 percent of
09:09
respondents
09:10
said she was highly attractive with
09:12
again a subjective rating of more than 7
09:15
out of 10.
09:16
69.8 said they will swipe right
09:19
similarly 44 said this profile looked
09:22
real
09:22
they showed the statistical results for
09:24
several other profiles in their study
09:26
which i invite you to read if you want
09:28
to see more fake profiles
09:30
like these you won't be disappointed the
09:33
results are very similar for all
09:35
profiles they tested in their
09:36
experiments
09:37
concluding that around half of their
09:39
respondents cannot distinguish a fake
09:41
from an actual person
09:43
moreover this half finds them highly
09:45
attractive
09:47
of course online dating is not the real
09:49
world and i am pretty confident the 44
09:52
of respondents wouldn't be duped when
09:54
seeing the person
09:55
on their first date but still the study
09:58
shows that
09:59
as they say the person you are talking
10:01
to may not always be
10:02
who they say they are to which i would
10:04
like to add that it might not even be a
10:06
person
10:07
at all when it's too good to be true it
10:10
probably is
10:11
to end on a more positive note the
10:12
advancement of artificial intelligence
10:14
is incredible
10:16
here they used previous state-of-the-art
10:18
techniques meaning that the fake
10:20
profiles could look even better now
10:22
not to talk about how these generative
10:25
models will perform in a couple of years
10:27
feel free to check out the study linked
10:29
in the comments and let me know what you
10:31
think about these fake profiles
10:33
do they look real to you i will gladly
10:36
discuss it with you in the comments
10:38
thank you for watching
10:47
[Music]