Ethical AI and ML fairness: A business priority for enterprises 

By Hasit Trivedi, CTO Digital Technologies & Global Head – AI, Tech Mahindra

With great potential comes great accountability and Artificial Intelligence (AI) is no different. The advancement of AI is creating new openings to enhance the lives of people around the world, bringing far-reaching changes to how we interact, do our work, understand our world, and pretty much everything we do. AI can bring in sustainable and equitable economic growth and tackle global challenges at a scale that has proven difficult to tackle till recently, helping businesses to reach a greater height quickly and efficiently. However, if not designed and governed properly, AI can go awry, it can increase societal bias, damage customer relationships for businesses, and erode finely balanced institutions that we have built over our history.

Undoubtedly, AI and Machine Learning (ML) solutions are transforming several industries’ operational, functional, and strategic landscapes. In the enterprise context, AI systems have the potential to improve productivity and open up new avenues of revenue by building insights from data that define business and customer interaction patterns. On the flip side, perceived unfair treatment by the AI and ML model can lead to trust issues, brand value erosion, talent retention, and acquisition, and missed business opportunities—all directly impacting the top line.

Decoding ethics, bias, and fairness

As organisations are scaling up their usage of AI to obtain higher business outcomes, concerns around AI ethics are also becoming significant.

While AI can be a powerful tool to help companies make decisions, it also has the potential to be harmful if not carefully managed. AI is more about how decisions are made. If we want our employees and customers to feel safe and valued, then we need to make sure that the systems we use for decision-making have been thoughtfully designed from the ground up.

The goal of responsible AI is to design and deploy AI with good intentions to empower employees and businesses and fairly impact customers and society. This allows companies to generate trust in their products while also scaling their processes with confidence.

AI ethics is associated with a variety of subjects like privacy and surveillance, behaviour manipulation, accountability of autonomous systems, robotic rights, etc. Organisations should begin with a deep understanding of the ethical problems that they are trying to solve to implement, scale, and maintain effective AI ethical risk mitigation strategies. The first step should consist of learning how to talk about it in concrete and actionable ways.

An example of this could be a person being asked to pay a higher premium based on the predictions made by the model which took into account some of the attributes such as gender and race for making the predictions. Another instance is if an application for a credit card or bank loan does not get approved due to the educational background or race of an applicant, though he/she ticked all criteria which are required to be approved.

There are myriad of such examples. Thus, organisations must choose the right data sampling strategy based on predictable discrimination issues in the ML model. Proper fairness metrics must be implemented at the testing phase to weed out fairness-related issues before the ML model is live. Moreover, a live model has to be continuously assessed for fairness as a part of the overall performance monitoring.

Building an AI Ethics Framework from a people, process, and technology standpoint

Ethical AI and ML fairness are still very new and evolving concepts. Given the importance of ethics in AI, companies are now creating an ethics officer role tasked with ensuring compliance at all levels – people, processes, and technology. 

At the people level, awareness of AI ethical practices, the possibility of injecting bias through the data science lifecycle, and the potential business impact of bias and fairness are necessary. Establishing courses to train data engineers, data scientists, ML modelers, and operations personnel are essential.

Ethical practices at the process level include establishing an end-to-end Data Science Lifecycle and proper governance. This should encompass information architecture that considers the necessary bias-related checks on the input data and AI fairness algorithmic design principles. The data and delivery process should include clear guidelines for diversity requirements in data, choosing fairness measures, coverage of model user profiles, identification of advantaged and disadvantaged groups, and data sampling to check bias and fairness. Further, establishing end-to-end traceability and accountability is an essential foundation for auditing, tracing, identifying, and fixing issues as they occur. Operational practices would include processes to flag bias-related issues and steps for manual intervention.

From a technology standpoint, ethical practices include Responsible First approach while designing and developing AI work package. This means the use of right tools and techniques to detect bias and fairness in data to detect bias and fairness in models, to monitor AI performance on an ongoing basis, to explain models, and to mask data wherever applicable. Besides, a structured and comprehensive Responsible AI maturity assessment and human-in-loop approach in critical AI-led inferences and actions are also essential.

Way forward for ethical AI

An ethical approach to AI is an absolute imperative. Knowing that an AI system will do whatever it is assigned to without considering its effects or legality, we must have strict regulations to prevent AI from generating adverse outcomes. In this direction, organisations must make their algorithms more transparent, introduce model milestones that make it possible to understand and correct the output at each stage and study the diversity of biases that occur to eliminate them. Fortunately, the government, private organisations, and institutions are already categorising and identifying safe ways to deploy AI into their daily operations and secure sustainable growth.