Understanding biases in the AI world

In the past few months we had seen quite a bit of chatter happening on the artificial intelligence world especially after the launch of ChatGPT. The world seems to revolve around the idea that AI systems will completely take over technology and the way how we do our day to day work. While this is not a remote possibility it is equally important to understand that AI is built by humans and hence will include biases that humans could bring in their day to day work. The intent of this article is to discuss the various biases that is possible in the AI systems and how they could impact the outcome of AI driven systems.

pic courtesy — Wikimedia Commons

Before we dive into how biases impact AI systems let us try to understand some of the real world biases that exists within us. Some of this will certainly impact the way how we think, how we act, and how we go about doing our day to day activities at work, at home, and in the social circle. The descriptions used below is an (over)simplified version, but it gives a fair understanding to appreciate the context of this article.

1. Anchoring — The intuition to rely on the on the first piece of information to make decisions. For example, the price of the first car that you see when you are looking to buy a new car will heavily influence the price range of what you would be willing to spend for the car. Or the value of the first proposal that you review during a proposal review will set a benchmark for the rest of the proposals
2. Attribution bias — This normally happens when people try to discover explanations behind actions (own and others). For example when people tend to look at others actions as a function of their character/upbringing/linguistic or religious background etc. whereas for their own actions is attributed to the unique situation or ‘need of the hour’.
3. Confirmation bias — This is the tendency to search for data points that are supporting your hypothesis by giving undue importance to the data points that supports your hypothesis while down playing the data points that are not supporting your hypothesis. For example, when you are evaluating different software tools for a certain functionality, if you have used one of them in your prior job it’s quite possible that you might end up buying the software you are familiar with.

There are many other types of biases that exists but for the relevance of this article let’s limit the list to these 3 biases to understand how they play a role in building and running artificial intelligence systems.

How does these biases impact AI?

• The real world biases is reflected in the data that used to trained the AI models creating a data bias
• The data bias is exposed by the algorithm, further complicated by the real world bias of the person who built (programmed) the algorithm resulting in algorithmic bias
• The outcome of the AI systems, if acted upon without due diligence will result in business bias.
• And the business bias will in turn impact the real world bias (either amplifying or creating new force fields)

As you can see, this becomes a vicious circle if left unchecked. Let’s try to understand these biases with relatable examples

1. Data Bias — Any AI system before it starts operating on its own needs to train using datasets. The quality of data that is included in this data set is a big factor on how the algorithm learns how to act and perform the respective actions that it is expected to perform.
   For example Consider you are training an algorithm to study the medical data of humans over the last 25 years to predict possible interdependencies on lifestyle with multiple terminal illnesses where the training data set included data spread of 80% of Caucasians, 10% of Blacks and maybe 5% each of Latinos and Asians. The algorithm tends to assume that this data set is a fair representation of the world in terms of race, food habits, environment and lifestyles. Imagine someone doing medical interventions in the far-east based on the recommendations provided by this algorithm which is trained on a skewed data set.
2. Algorithmic Bias — These are systematic and repeatable system errors that can create unfair outcomes as a result. It doesn’t necessarily need to have malicious intent when developed, but can be an outcome on how the algorithm was coded, data collected, selected and used to train the algorithm. This type of bias is commonly seen with search results and recommendation engines used by social media platforms.
   The recent experience with countries, governments and societies becoming more sectarian and divisive can be attributed to the lack of commitment from some of the major social media platforms to curtail the negative effect of algorithmic bias in their autonomous engines. Another example — The relative inability of facial recognition technology to accurately identify darker-skinned faces has been linked to multiple wrongful arrests of black men, an issue stemming from imbalanced datasets
3. Business Bias — What happens to organizations (who make data driven decisions) when they lack awareness about data bias and algorithmic bias. It sure results in skewed business practices and business models that has the potential to land organizations to face legal sanctions, penalties and sometimes even jail-term for the leaders who are involved. We hear about the deposition of various business leaders leading big-tech companies in the recent past where they made decisions that could be classified as — not necessarily illegal, but unethical. And the legislation is catching up with the advancement of such technologies.
   The infamous data scandal involving Cambridge Analytica and Facebook is a classic example of business bias converted to an opportunity.

The end user experience from the actions coming out such biased decision making will result in normalization of such outliers influencing the real lives thereby cementing their position into the real world biases spinning off the world in to a trajectory that none of us can forecast or imagine — which is a scary thought.

What can be done?

The first step to solving any problem is to acknowledge that there is one.

• Acknowledge that there will be bias in technology.
• Acknowledge that it is humans developing these algorithms and their thought process will influence the algorithms. Even for the computer self-developed algorithms there is some human who built the algorithm to build those algorithms.
• Acknowledge that computers do not have a brain for itself (yet). So the outcome is purely based on what you feed the algorithms through data. If you feed good data you get reliable outcomes. If you feed bad data it is garbage in garbage out.
• Data Governance is an important aspect that is a ‘lost art’ in many organizations. That needs to be revived.
• The current possibilities that are available using technology in the artificial intelligence space is mindboggling. Learn, unlearn, and relearn.
• And never forget to have fun!