We will take the simple Image recognition: the robot who recognizes dogs experience to illustrate what is bias.
With more than 400.000 images analyzed, I am very disappointed because the chihuahua has higher results than the Labrador.
InceptionV3 engine is trained to recognize all objects from environment equally.
It's a generic model dedicated to image recognition.
The result could have been totally different.
My favorite dog is the Labrador
I could have "trained" inceptionV3 to recognize labrador very well and have a lower level of recognition on a chihuahua.
Chihuahua would be counted within the irrelevant images. Labrador could have reached the top of the list.
I could proudly announce the following conclusion: "Artificial Intelligence decided that the Labrador has the best place under the hashtag #dogs"
If I do that, I introduce bias in the analysis: my preference for Labradors.
In one hand, we could say that this result is "false". In the other hand, this experience enforces my conviction and position.
Then I could put an end to the debate by saying: "Artificial intelligence does not make mistake, it never lies."
You have only access to the result without any way to verify.
Basically, you trust it. It seems to be rock solid.
There is obviously a problem of transparency.
Indeed, it does not lie. It's just a partial view on things.
Not within the categorizing rules themselves but in the choice of data and the way of teaching artificial intelligence.
If we transpose the problem in the real life: Keep calm. It does already exist
Experts begin to show that AI's engines could have some racist and sexist behaviors.
Article from the Telegraph: Robots, sexist and racist?
The Science (Semantics derived automatically from language corpora contain human-like biases) notes that these biases exist in AI's engines. They conclude that is logical to find them there.
These engines analyze large sets of data which embed implicit traces of moral and historical bias.
The existence of these biases can be diminished or accentuated. It's directly linked to the creator of the group of creators of a system or a product.
For example, if we give the software development of a color recognition application to a person with color blindness. The engine should logically mess with colors.
Consciously or not, when we educate AI from our data, we also integrate into it some weird parts of our behavior in variable proportions.
That's why DeepMind/Google begins seriously to work on the ethical part.
That's also why this topic should be widely shared and understood.