[ Archived Post ] Bias in AI

Please note that this post is for my own educational purpose.

Most algorithms are related to Deep learning → and many of the bias of human beings will carry over.

Bias → can happen in any stages of AI process → this is a very important question to tackle and create solutions. → fairness is a key question to answer.

The Data → itself can contain more bias data → such as light skin faces vs dark skin faces → or → amazon → remove females.

Preparing the data → feature engineering can also play a role in the creation of bias. This is a key is to tackle as well.

Which feature does AI choose → a hard question to answer.

Why is this hard to fix?

We might not even know that bias exists → we might not even know that a problem exists.

Not a lot of pre-trained model → does not care about the bias values → and this pre-trained model can carry over.

People who create ML → does not care about the social problems → just implement and beat the benchmark. (different social aspects have to come into play.).

We are addressing this problem as well → detect and fight back the bias process is critical.

A computer program used for bail and sentencing decisions was labeled biased against blacks. It’s actually not that clear.

An algorithm is a bias against towards black people → this is actually used in the court system.

A race is not used → so is this fair?

Since we do not care about the race → this seems pretty fair?

Black people are more likely to get a high score. (since some metadata does contribute).

However, there is some limit on how much an algorithm can be less bias.

Some other solutions do exist → but this is not an algorithm solution.

Machine Bias

Some girls went into trouble and charged with theft. So many people have done shady things and software → thought black people would again do criminal activity.