AI Bias & Fairness

Business of AI Jul 5, 2021

In our previous blogs, you have seen the power of AI and its enormous applications. But like they say with great power comes great responsibility.

One of the biggest threats to the applications of AI is its vulnerability to falling prey to developing AI bias.

AI bias is the phenomenon where algorithms develop systematic prejudice due to erroneous assumptions in the machine learning process.

Real-World Examples of AI Bias

Take for example the results of a gender detection model developed by companies like Microsoft and IBM.

Gender and Racial Bias in Gender Classifier by Microsoft, IBM, and Face++

As it is evident from the above figure, most of the AI models had a hard time classifying darker-skinned people!

Now that we have seen a real-world example of where the AI Models seem to be highly biased, let us try to understand why this could have happened.

Root Causes of AI Bias

There are many stages in the pipeline of building an AI Model, where bias can stem from. The most prominent of them are as follows:

  1. Data Bias: Artificial Intelligent models are only as good as their data. If there exist problems in the data set, like class imbalance or under-representation of certain groups, then you can most certainly expect your model to biased.

    In the example of the gender detection model, the COCO data set was used, which broadly contains data samples from Western Europe and Canada. Hence it was not trained on an appropriate number of dark-skinned people, resulting in bias.
  2. Model Bias: Lack of interpretability and performance metrics can also propagate prejudice in the model. Therefore it is important to tune your model on metrics other than accuracy, and on all sub-classes as well.
  3. Evaluation Bias: When you evaluate a model on the bulk of data, instead of all subgroups present, you are making your model more susceptible to developing bias.
  4. Interpretation Bias: This primarily results from human bias in judging the results of you model. A good thing to remember while interpreting results is, correlation does not mean causation.


It is extremely important that as AI practitioners, one understands the risks of AI bias and takes precautionary measures to build fairer algorithms. As our society becomes more autonomous and we delegate more tasks to machines, mitigating bias become extremely crucial.

We don't want our models to propagate racial, gender, cultural biases in our society. When mitigating biases becomes our primary concern, it means we are well on the road to fairer AI.


Nachiketa Hebbar

Computer Vision Engineer at Awiros| YouTuber|

Great! You've successfully subscribed.
Great! Next, complete checkout for full access.
Welcome back! You've successfully signed in.
Success! Your account is fully activated, you now have access to all content.