7962
Facebook’s Gender Bias
Produced by RTS
VO
00:00:01,772 --> 00:00:26,036
-Finding job offers has never been easier
on the internet. In a few clicks you have hundreds of
instant and personalized
ads. Thanks to their algorithms, Facebook and Google are
widely used in
the recruitment world. Their belief is that machines are more
efficient than humans.
but people are already speaking out.
ISABELLE COLLET
00:00:26,061 --> 00:00:37,990
-We're in a tech myth: where we think that because it is an
algorithm, it is objective.
Ads are objective: not at all! Since mostly men and women develop
software and algorithms
they're not objective.
VO
00:00:38,373 --> 00:01:46,506
The proof in images. -Look at this test. A man and a woman,
will do an identical job search within 5km. Neither specified
their gender, experience, or age
but the results have nothing to do with one another. For
her, secretary, babysitter,
and part-time work. For him, professor, team leader, mostly
full time. Quite a sexist result, and we are not
at the end of our surprises. Our experience started in
Berlin a few weeks ago, near the touristic Alexanderplatz.
Nicolas Kayser-Bril works for an NGO monitoring
algorithm. For us, he posed as a recruiter and posted real job offers on Google
and Facebook: lawyer, nurse, driver. He wants to check
who really gets these ads.
NICOLAS KAYSER BRIL
00:01:46,951 --> 00:02:26,746
-Some advertisements have only been targeting men like the one for the drivers.
Others were aimed specifically at women, like nursing jobs,
teachers. Discrimination takes place
before anyone clicks on the advertisement. Optimization is
done according to unknown criteria
even before the advertising is shown.
VO
00:02:28,629 --> 00:02:42,381
-Facebook's algorithm was never made public. It's impossible to understand how
Facebook's algorithm makes its choices. In Berlin, the team has taken
the experiment one step further.
NICOLAS KAYSER-BRIL
00:02:42,997 --> 00:03:09,545
-We changed the photos and we saw that Facebook uses the
pictures from the ads to
discriminate based on stereotypes. If we do a commercial for
drivers with a picture of cosmetics,
the ad will only be shown to women. If you specify that you
want to female drivers
and put a picture of heavy trucks, it will still be shown
only to men.
VO
00:03:10,864 --> 00:03:42,567
This raises questions such as hiring discrimination is
prohibited by law in Switzerland.
So are algorithms sexist or
neutral? Facebook and Google didn't reply
to our interview requests. Far from Silicon Valley, Isabelle
Collet is the pet peeve of the GAFAM.
This computer scientist is tracking gender discrimination created
by digital technology
She followed the case of Amazon The giant failed to follow through
with recruitment algorithms.
ISABELLE COLLET
00:03:44,379 --> 00:04:20,357-Amazon wanted to automate
its recruitment. The idea was that the top five résumés
would be hired directly. There was a
majority of men. They figured that the algorithm
was sexist, but no. Amazon HR was sexist. The algorithm
replicated
what we used to do: for a promotion or a good salary, it was
better to be a man.
It wasn't an algorithm, they needed
to re-evaluate how they managed human resources.
VO
00:04:21,638 --> 00:04:53,261
-What about Switzerland? We're not
at the stage of selecting CVs
or recruiting through algorithms, This leading company
of temporary workforce in Switzerland, gets 200'000
applications per year.
They're thinking about artificial
intelligence, but they are also suspicious.
Gregory Papin brought AI into to
the process, to speed up the experience of the candidates.
GREGORY PAPIN
00:04:53,910 -> 00:05:18,293
-When a person is applying, we just ask them for a CV and a
robot
performs semantic analyses to recognize his coordinates. AI
helps fill out a form
but the candidate still has the lead on that document. It
would be difficult
to trust AI 100% in the processing of these applications.
VO
00:05:20,017 --> 00:05:32,410
-At the group's headquarters, we're
thinking about the next step.
Cross-referencing CVs and offers to offer targeted jobs to
the candidates. But the ethical issues
are not far away.
GREGORY PAPIN
00:05:33,261 --> 00:05:54,998
-We'd have an ethical problem to process candidates automatically.
We need
our consultants' point of view. Our position now is not to
exclude new technologies,
but you can't see everything from
an artificial intelligence viewpoint.
We will always need humans in the processing of recruitment.
VO
00:05:56,260 --> 00:06:14,510
A few kilometres away others are already much further into
AI In Martigny, in this
well-known
institute research, we're working
on an algorithm designed to analyze job interviews.
These men are the masterminds of this project.
JEAN-MARC ODOBEZ
00:06:15,400 --> 00:06:18,951
-We have examples of job interviews where our algorithms can
learn
to recognize how emotions are expressed.
VO
00:06:23,858 --> 00:06:33,607
-Fear, joy, surprise: the aim is to help recruiters find out
when a candidate
mimics his enthusiasm or if his speech is not consistent
with his emotions.
The slightest movement of the muscles of the face is
detected.
JEAN MARC ODOBEZ
00:06:44,101 --> 00:07:16,287
-For example, what we need to identify on the face is the
raised eyebrows, characteristic of surprise
Open your mouth too, but it's rare
when we talk. Eyebrows are important to watch out.
For sadness, the lips will fall down.
For joy, it's the other way around.
For anger, it's the eyebrow twitch.
VO
00:07:17,580 --> 00:07:20,780
-The purpose of the machine is to be impartial. But many
gender or age biases
may exist.
JEAN MARC ODOBEZ
00:07:28,199 --> 00:07:59,622
-An elderly person will have wrinkles that could be interpreted
as expressions, even though the
no one expresses that emotion. Age must be provided so the
data takes it into account.
When you do not have a CV, you cannot do it without an interview
It can be a tool, but
it must remain only a tool. Behind them, humans make
decisions.
VO
00:08:01,002 --> 00:08:05,556
Correcting the algorithms and neutralize the gender bias, It is possible
In the high security biometric room, is one of Sebastien
Marcel's duties. In this lab,
they train the algorithms to spot the real from the fake, especially
thanks to
these silicon masks. To correct gender bias, you have to get to the root of it
and the ways teaching of data.
SEBASTIEN MARCEL
00:08:34,061 --> 00:08:37,569
The reason why some of these algorithms are biased is
because they've been trained
on biased databases. We're trying
to correct the data, matching the number of men and women
for example, but that's not always
possible. Systems are often already created.
We no longer have access to the original data.
VO
00:09:04,654 --> 00:09:27,183
-The second option is to break the codes, or
change the people who write them. Luanna Braga
studies code and algorithms at the University of Applied Sciences
in Geneva. In 2020, she's still an
exception in a highly masculine universe.
LUANA BRAGA
00:09:28,813 --> 00:09:36,935
-The first day, I counted, there were four of us out of 65. In
my class,
we're two girls out of 14. I'm used to it, but I wish I could have more women by my
side.
VO
00:09:58,179 --> 00:10:01,535
-That night, creating a game to learn writing command lines.
Luanna's holding on because she thinks
that more diversity could change the way algorithms govern society.
LUANA BRAGA
00:10:14,090 --> 00:10:58,911
-We have a percentage of 90% white and European men and they're bound to make
things programmed for them. I do not think it's voluntary. The ones that are here
do as they think. To me, it's
strategy. I wanted to see the areas where I could stand out. In other areas,
it's very difficult. Now it's the other way around.
VO
00:10:59,978 --> 00:11:04,218
-In the meantime, it's up to the recruiters
to be extremely careful. Charlène Kurer
and her associate
are aware of this. They need to recruit assistants for
wellness centre.
CHARLENE KURER
00:11:20,778 --> 00:11:25,538
-There will be seats
where we can lie down.
VO
00:11:25,686 --> 00:11:27,726
-To find the perfect assistant she chose to place her ad only
on Facebook.
CHARLENE KURER
00:11:31,667 --> 00:11:33,867
-What worked well is the localisation. I'm
looking for someone dynamic,
available 10 to 40% of the time. We received 40
applications.
VO
00:11:48,307 --> 00:11:50,827
A non-gender specified ad
and a neutral image,
here, we've paid attention to
detail
hoping that the algorithm
does not bias this recruitment.
CHARLENE KURER
00:11:57,991 --> 00:12:00,751
-We'd like men like women, it would be a problem that
the algorithm chooses for us criteria that we didn't set.
VO
00:12:16,273 --> 00:12:39,047
-The efforts paid off for the recruiters, as their ad
attracted
as many men as women. Facebook and Google still haven't answered us
and have not announced a corrigendum of their algorithm.
Our questions have nonetheless reached them
since Google has removed the majority of
our ads
even though they were paid for and still valid.