Gender Bias and Inequality in AI

Introduction

Artificial intelligence has transformed the today’s world. Artificial intelligence and machine learning have completely integrated into our day-to-day computing needs. Artificial intelligence is everywhere from smart speakers like Amazon Echo Dot to booking taxis using Uber, Ola etc. The goal of Artificial intelligence is to simulate human learning. Now comes the big question, if artificial intelligence tries to simulate human learning is there a chance some human associated biases like gender bias can also be learned by the artificial intelligence?

Sadly, answer to the above question is yes. Woman have been discriminated in society for ages in past. In past they were subjected to cruel social hierarchy system where men were placed above them. They were denied social services and even were discriminated in workplaces. Though now things have changed for better. There comes another question that if things have changed for better then

Why a modern concept like artificial intelligence and machine learning are prone to gender bias?

One of the main reasons that can be attributed to gender bias in Artificial intelligence is data.

You may be asking data? How is it even related to the Siri that I am using on my IPhone?

Well, let me give you short description of how artificial intelligence and machine learning work. In simple terms you give artificial intelligence model data, model gives you predictions or the desired output. To elaborate it more, you have a dataset which is divided into three parts: test, training and validation. You train the model on training data. Model here is nothing but a mathematical formulation or in terms of computer science, an algorithm. After training you test and validate the model using testing and validation data respectively. If you are satisfied with results, deploy the model.

To cut the story short, data is core essence of artificial intelligence and machine learning.

Now time to address the elephant in the room, how is data responsible for gender bias

Well as mentioned above artificial intelligence makes inference based upon the data given during training, so if training data is somewhat biased toward a particular gender that’s when problem comes. This bias can be quite harmful as artificial intelligence models are also used in sectors such as banking, healthcare etc. Biases in these domains towards a particular gender can have severe consequences.

How does this data get biased you may ask?

Well, it’s quite simple. Data is collected by humans. Now that is all fine but then again what if that person is biased or has some latent bias i.e. person lacks self-awareness to know that person has some preconceived notion about the society. Then this type of data will obviously be biased and results will be biased.

Now what if we somehow get a non-biased data, data that is true representative of observations involved in the study / problem considered. Then this problem will be solved right? well hold your horses, there is still another hurdle to cross, AI industry itself.

AI industry, is largely dominated by men. How much dominated it is? You speak.

Here are some statistics that give us some idea about gender diversity in AI industry. WIRED and Element AI a Montreal research firm in 2018 conducted a study and estimated that only 12% of leading machine learning scientists were women. Another study conducted by Nesta which is a United Kingdom based foundation estimated that in 2019 only 13.83% of research papers related to field of AI were authored by women

Ok so how does it relate to our original problem of AI bias?

See to build an AI model to predict something we not only need data to train the model but we also need the model itself. Model is just a mathematical formulation or simply an algorithm. The models are made by leading researchers in this field. Models usually find patterns within the data and use this information for prediction. Moreover, depending upon model and use case, model can decide that certain features can have high importance as compared to others. Well, it seems all fine, then where is the problem? the problem is that even if we have completely representative unbiased data, there is still a chance that we get a biased algorithm that can produce biased results. Even if we make data and algorithm both unbiased there is still another major issue which is that, as data is supposed to be representation of the society, so data can also represent prejudices of our society.

So, there are various multifaceted problems. To show how severe these biases can be, here are some real-life examples: -

Amazon, one of biggest company in 2015 employed and Ai based algorithm for hiring employees. It was soon found out that this algorithm was sexists to woman. The reason being that algorithm used hiring data from past 10 years and since maximum of employees hired were men. So, algorithm preferred men over women for hiring purposes.

Another major company Google also faced the same issue, few years back it was found that there was less probability of high paid job advertisement shown to women as compared to men on google. The study was conducted by a group of scientists from Carnegie Mellon University (CMU).

Besides these companies there are more instances of this gender discriminatory behavior conducted by algorithm. Though the above-mentioned instances of amazon and google were rectified by the company but the point still stands that there are these types of discriminatory instances exhibited by AI.

Credit Industry

This behavior is quite frequent and can be seen in credit industry. Earlier to determine credit score factors such as martial status and gender were heavily used. Since then, the approach has changed. This seems promising then where is the problem? Well problem lies in the fact that AI uses traditional data for prediction of output. So as the traditional data is itself biased, these biased properties will be reflected in the predictions made by AI itself. This same behavior was observed for quite some time in Apple Card.

The case of Apple Card

Apple card originally was released in 2019. This card is a credit card that was created by Apple and was issued by Goldman Sachs. After its initial release in 2019, it was observed that card was gender discriminatory in nature i.e., there were several instances observed where credit limit of women was lower as compared to men despite the fact that credit score of women was equal or more than men. Later on it was investigated and in 2021 was approved

The above examples just highlight the fact that if mainstream big companies with their seamless resources are getting biased results, then clearly something is wrong with AI industry.

Now you might be asking what to do about this? What are the solutions to this problem?

Well, first obvious solution is to increase the diversity in AI industry. More the women working in industry more there is a chance that such biased algorithm will be detected early and rectified before it comes to mainstream market. Moreover, women in AI developments can test the models and give their feedbacks regarding such issues.

Second solution can be to improve data gathering process. One way to do is to send the data set to a third party where this data can be checked to see if there is a bias or not.

Moreover, AI teams should work more towards explainability of model. Instead of black box approach, we can try to make model more explainable. This allows to see what features are considered more potent for a problem, from which we can infer if there is a potential of bias or not. More investment in research and development of algorithms and models is one of the ways to identify and reduce potential biases.

Final thoughts

In conclusion AI is a very versatile and has a lot of application from health care to banking, Ai has slowly integrated into our day-to-day life. But like all other inventions it has some demerits. One of them being gender bias. The reason why it becomes targeted with something that is a human concept is can be attributed to fact that AI is supposed to simulate human learning, so data used in learning can have biased or model used in learning can also be biased and so on. This gender bias at first may sound harmless but we have seen from real life examples how harmful can it be from effecting credit limit to job opportunities, biases have long lasting consequences. Some possible solutions were also listed from diversification of AI industry to investing more in research and developments of algorithms. The goal of all these solutions is mainly to identify and rectify the biases, that may otherwise be neglected.

Coming to real world, industries are making changes to make AI more inclusive but there still is a long way to go and we must remind ourselves the along the way to push for more inclusive AI