Artificial intelligence or AI technology — what remained as a fantasy that existed just in books & movies is a reality today. Applications of AI have exponentially grown and have penetrated almost every industry. It’s just a matter of time for AI-enabled machines to replace humans in all the mundane jobs. Soon, AI bots would assist humans in executing tasks much efficiently.
As AI is transforming many sectors, and influencing our day-to-day life, one question raises.
Can we trust AI?
One of the many benefits offered by AI solutions is that it helps humans in decision-making. But, what if AI’s decisions are biased? Recent studies about algorithmic bias have pointed out such flaws in AI systems. These flaws termed algorithm bias could result in biased decision-making. But, how does it happen?
The root cause of bias in AI is human bias, which is an extensive issue and an ongoing research topic in psychology for years. A person can be biased and unaware of it. Unfortunately, the datasets, databases we have— reflect the bias of the person who curated it. When such datasets are used for training AI models, their decisions reflect the bias that was hiding in the dataset. Let’s see some examples;
In 2016 Microsoft launched an AI-powered conversational chatbot on Twitter that was designed to communicate with users via tweets and direct messages. However, within a few hours of its launch, it began responding with highly offensive and racist messages. The chatbot was trained using anonymous public data and had an internal learning function built-in. A group of people was able to manipulate the system with a concerted attack to inject racial prejudice into it. Some users were successful in flooding the bot with misogynistic, racist, and anti-semitic words.
In another example, Amazon worked for years on developing an AI model to automate the hiring process. However, they had to discontinue the automation tools as they exhibited bias. The AI-powered solution was supposed to sort through a stack of resumes and identify the best applicants. For that, the company fed the machine a decade’s worth of resumes from people who applied for various positions. But it was, later on, discovered that the algorithm restricted the applicants from women’s colleges as it was only favoring the words which were commonly found on men’s resumes.
Today, as more companies and organizations are planning to deploy AI systems across their operations, it is essential and an urgent priority for the stakeholders to be aware of the potential risks associated with it and how to mitigate them.
In a nutshell, AI algorithms will learn the patterns of the data that is used to train it. In other words, AI models will have a bias that reflects the creators’ or data’s bias. Implying that cognitive prejudices are at the root of both modern AI and data biases.
Several approaches can be enforced to prevent the biasing constraints on AI models. The first approach consists of pre-processing the data to maintain quality accuracy while reducing the relationship between outcomes and sensitive characteristics. That is, to the best capability, remove the evident bias from the dataset.
The second approach consists of post-processing techniques in which models’ predictions are made to satisfy the fairness constraints. The fairness constraints would oversee the outcomes and can help to detect the bias. The third approach imposes fairness constraints on the optimization process to minimize the system’s ability to predict the sensitive characteristic.
Researchers and data scientists are working on designing and evaluating new approaches to achieve a more effective result. Here are some suggestive steps to reduce bias in AI
Trying to solve too many scenarios also results in an unmanageable number of labels spread through an unmanageable number of groups. Identify challenges and model algorithms specific to the challenge.
For a single data point, there are usually several valid opinions or labels. Your model will be more flexible if you collect those opinions and account for legitimate, often subjective, disagreements.
There are classes and labels in both academic and commercial datasets that can introduce bias into your algorithms. You’re less likely to be surprised by objectionable labels if you understand and own your data. Check to see if your data accurately reflects the diversity of your end-users.
Recognize that the end-users may not be identical to you or your team. Empathize with others. Recognize your end users’ diverse backgrounds, perspectives, and demographics. Avoid AI bias by anticipating how people that aren’t like you would communicate with your technology and the issues that might occur as a result.
The more diverse your views are, the larger the pool of human annotators. This will significantly minimize bias, both during the initial launch and when the models are retrained. One choice is to use a global crowd of annotators, who can not only provide a variety of viewpoints, but also help a wide range of languages, dialects, and geographically specific languages.
Models are seldom static throughout their lives. A popular, but serious, error is deploying the model without allowing end-users to provide input on how the model is performing in the real world. Opening a discussion and feedback platform will help to ensure that the model maintains optimum success levels for all.
You should revisit the model regularly, not only based on customer reviews, but also by having independent people check it for updates, edge cases, instances of bias that you might have overlooked, and so on. Make sure you receive feedback from your model and provide feedback of your own to enhance its consistency, iterating toward greater accuracy at all times.
Businesses and policymakers can make an effort to subsidize these algorithm biases. They just have to look into things more carefully and here are some suggestions.
To conclude, as we progress in identifying the bias points in AI, we should re-organize the standards to determine the fairness of human decisions and input as many datasets into the system as possible to reduce the biasing.