loading...
The

BIAISES
of
Artificial Intelligence






by Biljana Petreska, 2022

*More precisely, a subfield of artificial intelligence called machine learning.
Start

BIAS noun. A systematic distortion.



Select and click on 2 balls
to find out who is behind them...


Next...
Thanks !

Now, imagine that
YOU are the prediction algorithm.

Your goal is to find out if
behind all these balls there are
more indigos clouds or
more green clouds.


Click on 8 other balls.

Done !

indigo clouds


You have chosen the indigo clouds. However, there are more green clouds on the screen.

Try again...
Today,
the prediction algorithm
is you !

So, according to your own observations,
there are more :



green clouds
Artificial intelligence algorithms learn
by observing data.

After observing a lot of data,
AI algorithms can predict
whether it will shine or it will rain,
whether you could take the bus or the train,
whether you should watch TV series A ou B instead...
I see
Artificial intelligence algorithms
do not think, they just count.

They observe lots of data,
and then make reasonable predictions
based on all the data they have observed.
I understand
Here's an example.
In order to complete your sentences, an algorithm
observes a large number of sentences
and counts their words.

The AI algorithm will then suggest
the most frequent words
found in sentences similar to yours.


Let's continue our observation.
Click on 10 more balls.

Done !

green clouds


You have chosen the green clouds. However, there are more indigo clouds on the screen.

Try again...
The more observations there are,
the more we can hope to
appproach the truth.

Now there are more :

indigo clouds

Qu'est-ce que ces observations
vous permettent de conclure ?

Quelle prédiction un algorithme ferait-il :
y a-t-il plus de nuages indigos
ou plus de nuages verts qui se cachent
derrière tous les points d'interrogation ?

Cliquez sur le bouton pour le savoir.

Show all
Actually, there are exactly the same number
of indigo clouds and green clouds.

But, depending on the data we observe,
we can make mistakes.


You have just been a witness to
the DATA BIAS. Oh ?
As it is impossible to observe everything,
the AI algorithm makes decisions
based on data that is
an imperfect representation of reality.

In general,
the more data the algorithm observes,
the better its predictions.

 Yet still...
Artificial intelligence algorithms
are getting more and more powerful !

But it is not because they are
getting more intelligent.
It is because they are observing
more and more data :
thanks to BIG DATA. Another bias

an indigo clover


If you observe well, in general, the green clouds prefer the green plants.

Try again...
We are fortunate to have access to more data.
Each cloud's social media profile
shows its favourite plant species.

You are invited to dinner by a green cloud.
Based on your observations, would this green cloud
prefer an indigo clover or a green plant ?



 a green plant
Aouch !

This green cloud is one of a
minority of green clouds
that hate green plants.

This is the SOCIETAL BIAS :
the data reflects all the
biases and prejudices of our society.
 Really ?
Just imagine
that the green clouds are all female.

So you would give the females clouds
a green plant
just because they are female ?

You could be accused of being sexist !
I didn't know...
Imagine that the color of the clouds
is in fact a skin color.
You have decided someone's personal taste
based on their skin color.

You could be accused of being racist !

In short, you are
a biased and a discriminatory algorithm.
 It's not me !
I agree,
it is not your reasoning that is biased,
but the data you have observed.


But the damage is done !

 How so ?
That's why Facebook's algorithm
promoted sexist job ads
by reproducing social stereotypes :
male truck drivers, female nurses.

That's why Google's hate speech detection algorithm
is racist :
tweets written by African Americans
are more likely to be identified as toxic. Oh no !
Open AI and Microsoft's algorithm
may generate sexist and racist texts
because it learns from data published on the web,
which is itself often sexist and racist.


While all the algorithm does is count,
like yourself for the green plant. What can I do ?
Pour contrer le BIAIS des données,
et le racisme,
les chercheurs font attention
à récolter des données
qui incluent différentes minorités.

Pour contrer le BIAIS sociétal,
et le sexisme
les chercheurs créent des fausses données
pour compenser les différences.
Que puis-je faire ?

It's impossible to eliminate all biases !

It is necessary to be critical of the results
that algorithms give to us,
and to support the creation of laws that
protect us from dangerous drifts.
 Dangerous ?
Would you agree to let an algorithm decide :

• your admission to university
• how much you can borrow
• whether or not you go to prison
• about your organ donation

based on your gender or skin color ?


 Another danger.

Social media suggest that we connect with people
who have similar preferences and opinions.

They enclose us in a filter bubble :
a view of the world that hides many
preferences and opinions that
are different from ours.
 Oops

Imagine that the clouds-plants
are afraid of global warming,
while the clouds-cloves believe in
the Earth's ability to repair itself.

The clouds-plants do not know
the arguments of the clouds-cloves.
The clouds-cloves do not understand
why the clouds-plants are so alarmed.
 Oh !





Two disconnected filter bubbles... It's a shame

yes


You said yes. But what do you know about the coral cloud ?

We saw with the data bias that the algorithm's predictions are incorrect when too little data has been observed.

If the algorithm has never observed any data, it can only make an uninformed guess.

Try again...
Finally, let's meet
a new coral cloud for the first time !

Can an AI algorithm predict
whether this cloud likes green plants
or prefers indigo clovers ?

 no
This is correct !
If the algorithm has never observed a piece of data,
it can only guess.

Looking at their shape or color,
what prediction would you make ?

These new coral clouds prefer...


 I have an idea
The coral cloud has the same shape as the green cloud:
the algorithm could predict that the coral cloud
prefers green plants.

But every decision has consequences...

If the social media suggests
green plants and never indigo clovers,
the coral cloud will end up preferring green plants.
 It's influenced ?
A sort of self-fulfilling prophecy.

The coral cloud will never know
that an algorithm has manipulated its opinion
and will sincerely believe that it likes green plants.

Under the social influence of other clouds-plants,
it will be reassured in its choice. And me too ?
Two take-home messages.

With little data,
the most intelligent algorithm in the world
will make bad decisions,
this is the data bias.

But even with enough data
the algorithm can make bad decisions,
that is the societal bias.
 What now ?




Let's cultivate our diversity and
burst our filter bubbles.


I would like to know more... Thank you !
created by

BILJANA PETRESKA VON RITTER
I would like
to know more...


inspired by the sublime game

"Crowds" by NICKY CASE


the music is

"Moonshine" by KETSA


thanks to

HEP VAUD MODULO BETA-TESTERS