AI and Algorithmic Bias

AI written on Chip

The readers must know what exactly is the term algorithm to go ahead with the article. An algorithm is a specific set of instructions or steps used to solve a given problem. While codes are commonly used to express algorithms, not all algorithms are written in codes. This can be made easy by understanding how a particular dish or dessert is prepared using the perfect blend of the right ingredients. e.g. Biryani. In the case of Biryani, the biryani recipe is the algorithm that we need. The ingredients of biryani will form the input to the recipe, similar to the instructions given to the computer to receive a result as output. Now, after the correct mix of ingredients is ready, a bunch of steps need to be followed following the recipe and the dish is prepared. The result is a mouth-watering biryani. The algorithms are made exactly like that, just that the algorithms may comprise steps that are themselves algorithms and may be used as individual algorithmic units.

Surprisingly, algorithms for search engines work in a very simple manner. The moment something is entered on the search bar, the algorithm is set in motion. It starts operating on the words entered on the search bar in the particular search engine, by first breaking down the search query, very similar to the process of digestion. This is done to look for and identify the keywords, check for spellings and cross-referencing with synonyms. This algorithm allows the search engine to derive meanings. The next algorithm does the searching by comparing the query against an enormous database containing information about hundreds of billions of web pages. This database is itself created by another set of algorithms, called crawlers which search the web constantly looking for new and updated pages to index and rank them in a gigantic database.

This results in a big list of outcomes which are far too many to be shown on the screen. The order in which the results are shown is done by ranking them by in the order of their expertise, authoritativeness and trustworthiness. One way of doing this is by seeing how many other authoritative sites have linked to this particular result. This step ranks the site on how user-friendly it is. Finally, the last step compares the result with the previous web user history and personal details to ensure that the content is best suited to you.

To simplify, algorithms are a series of instructions that are followed step by step to do something useful or solve a problem. Incredibly, search engine algorithms do all this in a fraction of a seconds. Algorithms are incredibly useful tools for dealing with large data sets but are also behind a lot of things that we encounter as we go about our daily lives. When we type a text onto the phones, algorithms are used to convey it to the person you are sending it to. When we buy something online, algorithms are used to make it harder for cyber criminals to intercept our credit or debit card information. Furthermore, they can be used for medical imaging technologies.

Algorithms serve a lot of purposes in our daily lives. They are used to predict the path of natural disasters like storms, and cyclones and in developing new drugs to cure diseases. Algorithms are learning about human beings, their tendencies, habits, likes and dislikes only to name a few. They are increasingly used in ways that shape our lives, but while algorithms bring many benefits, they can also raise concerns. One of the most important challenges with algorithms is

bias. To understand precisely what exactly is algorithm bias we need to revisit search engine algorithms. When we talk about algorithms that are responsible for quality control, and authoritativeness and trustworthiness, who measures this and how is it quantified?

What is algorithm bias? How can an algorithm get biased? A biased algorithm reflects legally or ethically problematic dimensions based on an attribute such as gender, race, age, sexual orientation, religion and so on. The bias in the algorithm can come from data. It can also come from the people who design or use algorithms. The internet is full of data, on our browsing history, purchasing habits of people, employment history, data of people with whom we communicate online, social media posts and much more, but such data cannot accurately describe people or the biases, prejudices and stereotypes prevalent in the society. This can be attributed to the fact that not everyone in the world has equal access to opportunities such as education, financial services and jobs. In addition, data can be used and interpreted in biased ways. Data which is used to train facial recognition algorithms are trained by engineers, by showing them images of lots of faces. In general, the more face images, an algorithm can analyze when it is being trained, the better it will become at performing facial recognition, but a problem occurs when the database of training images is not diverse enough. If most of the images used to train the facial recognition algorithm are people of lighter skin, the algorithm will perform poorly when analyzing the faces of people with comparatively darker skin. This can occur if the engineers who choose the training data don't sufficiently expose the algorithm to the different types of faces that the algorithm will encounter. There have been numerous instances of African Americans who were falsely arrested because the facial recognition algorithm mistook them for someone else.

How can an algorithm be measured to be free of bias? It is indeed a difficult question because bias in an algorithm is not purely a mathematical concept or particular code. It is also socially constructed, a decision that one person might consider fair, might be considered biased by someone else. Although, not a mathematical concept, math can help in understanding the bias in algorithms and mitigate it further. But this can take place only when the term bias is defined concerning algorithms and at the same time, there is no such perfect definition of bias. Depending on the situation, goals and the perspectives of the people designing and using the algorithms, there may not be a single right way to measure bias.

Since algorithms in AI are becoming increasingly fundamental to our daily lives bringing a lot of benefits, the bias in the system can result in problematic outcomes. Addressing AI bias is part of the broader goal of promoting ethical AI to ensure that this technology can be used in ethical and inclusive ways. The legislative bodies, civil societies various organizations of researchers, and companies are working to develop frameworks for responsible use of AI. Many of these efforts share common goals like fairness, transparency, diversity, inclusion, data privacy etc. The data used by AI algorithms must equally reflect the community's diversity. AI tools generate an enormous amount of data, that data often includes personal information like credit history, purchasing habits and medical history. AI designers need to ensure that their systems fully protect the privacy of the people whose data is being utilized. Whether we realize it or not, in the

Coming years, AI will often be at the crux of many key potentially, life-altering turning points. AI controls and plays a role in the information that we receive online, the people we meet and whether an application for a loan or a job is successful. AI is no less than an extension of the people who create it and the biases that they have. It can easily reflect human biases. The underlying source of bias is people. AI can help in removing these biases by giving a powerful set of tools to analyze vast amounts of data, they can help in identifying biases that might not have even been recognized. It is, therefore important to remember that no matter how advanced AI might get, it still isn't human, if it is used in the right ways, it can bring iimmense benefits.