By 2023, the market for AI hardware and software would cross $500 billion dollars. And, by 2030, it would be close to $5 trillion USD. A big lobby of business leaders, however, don’t want to let out one thing leak out from their walled gardens, because of the issues we are going to discuss here, if they reached the customers, the AI companies risk losing a lot of money, their reputation and above all, their ability to promote trust and reliance among AI users.
A lot of ongoing work today in Applied AI courses focuses on three important things. These are Bias, Fairness, and Ethics in applied AI and machine learning complying with the newly formalized regulations controlling Artificial Intelligence governance. In this article, we will objectively evaluate the role of fairness in AI models and why data scientists are still languishing light years away from the ultimate objective of achieving a bias free AI model.
Why “fairness” is the talk of the AI town?
The talk in the Artificial Intelligence quarters for many years has ignored one very important aspect of fairness and bias in machine learning models. There are way too many loopholes in the Machine Learning algorithms that are arising due to the imminent biases.
AI bias measurement and analytics, therefore, is a raging cause for the failure of many AI projects around the world, and a majority of the academic work that is happening in this area focuses on finer nuances of inculcating fairness. In 2020, Denmark became the first country in the world to officially legalize the inclusion of data ethics and fairness in AI research, to bring in and motivate Applied AI engineers and analysts to embrace fairness.
In the language of machine learning algorithms, a fairness application is an extension of Applied AI models that can deliver a 100% or close to cent percent impartiality in sensitive data sets. When fairness is the ultimate goal, data scientists prefer to work with AI that can handle the predictions and outcomes in a fair, unbiased, and truly transparent manner. Fairness in AI could mean the obvious understanding of how different factors affect the outcomes and predictions in a sensitive data science project. For example, let us take this simple sentence – “European women from the Caucasian race between the age group of 25 and 50 years earning at least $100 dollars per week prefer to use more sunscreen during the summer compared to younger women in Europe, and women from Asia, and North America.”
If we break down the sentence, we are comparing the behavior of European women in a certain age group and economic standing with those belonging to other groups from the same population and others. Bias would suggest that women (gender / sex), age, financial health, and region would certainly affect the outcome of the data set when applied to a modern AI and machine learning algorithms. A fair treatment would empower data scientists to arrive at a conclusion that irrespective of where women come from, or what race and ethnicity they belong to, sunscreen usage data would still generate a fair result if treated on a balanced plane.
Another simple example is, fill in the blank – “This is a _____________.”
Now a biased and unfair Machine learning algorithm will answer it on the basis of input information and what it is computed with. “This is a CAT” will be equivalent to an answer that states “This is a SHEEP.”But, it’s not equivalent to something like this “This is a BEAUTIFUL PICTURE.” Or “This is NOT ALLOWED.”
That’s where data scientists have to characteristically define levels of “sensitivity” in designing their ML models for a fair trial. The lesser the influence of sensitivity on the ML algorithms, the better the AI models would perform in the fairness trial.
AI Fairness: Can anyone achieve this Holy Grail?
9 out of 10 data scientists would staunchly agree that achieving fairness in AI is next to impossible, and accepting this makes their work so much easier in the current context of data governance, ethical engagements, and bias removal. AI engineers and data analysts want to build new machine learning models that score very high on the fairness scale, powered by 100% optimization, accountability, interpretability, and above all these, data privacy and user safety. But, bringing all these under one roof and still guaranteeing that the Machine Learning model will deliver outcomes is beyond possibility! While AI companies lookout for AI solutions to reduce the workloads of human workers, the same AI makers are relying on some kind of human intelligence to appropriately design the next gen machine learning models free from biases. Is that possible? Some companies like Google, Baidu, and Microsoft think, yes, it’s possible. But, how is it possible?
To explain this, AI modelers use a human centred fairness measurement approach that evaluates the various factors influencing the biases that exist or arise from algorithms. While the objective is to correct these biases, a large number of ML clusters are so complex that data engineers find it extremely difficult to even identify these biases in the first place. So, delivering an unbiased and fair ML model gets beyond the reach of AI practitioners.
Where can AI models score on fairness?
In AI development, we can no longer ignore the role of fairness especially when biases are heavily criticized.
AI fairness is measured at different stages of development. These are during the processing of data, data analytics, and finally during the post-processing stages. It’s as simple as tossing a coin to find if the data set is biased towards a preference or predilection or it’s fair to an extent where no further equitable processing is required to establish that the AI model is working as per the objectives. This is the basis of all the development we are seeing in the field of facial recognition, natural language processing (NLP), automation, and computer vision.