Exploring Explainability in the Realm of Responsible AI

Time of India

08-05-2025

Relishing delicious delicacies is something that most people often look forward to - be it visiting a bakery and tasting creamy pastries or enjoying one's favorite dishes at a restaurant. But there is a habit we, as humans, usually possess: if the pastry or dish is extremely delectable, we ask for the recipe. Moreover, before going to the bakery or restaurant, we assess whether it adopts hygienic practices, uses quality ingredients, and so on. You guessed it right — inquisitiveness and skepticism make humans an extraordinary species.Similarly, while leveraging Artificial Intelligence, which quickly and efficiently provides solutions to a myriad of problems, having a curious and questioning mind is natural.Responsible AI and the Principle of ExplainabilitySince AI assists humans in solving problems and making decisions, it is imperative to ensure that people's trust is consistently maintained. Furthermore, the decisions and solutions must inevitably be ethical and law-abiding. There is a dire necessity for AI tools to act responsibly, and extensive deliberation on 'Responsible AI' thus finds a significant place in today's technology landscape. As AI continues to become indispensable for businesses, understanding how it derives a specific solution or reaches a particular conclusion becomes crucial. The Austrian-American consultant and educator Peter Drucker beautifully says, "The purpose of information is not knowledge. It is being able to take the right action." Responsible AI hence brings forth the principle of explainability to foster confidence in AI systems, equipping us to make informed decisions that translate into actions with the right impact.Model Performance and Explainability: Balancing Between the TwoWhile discussing Artificial Intelligence, we often come across the terms AI models and algorithms. AI models are programs empowered by AI algorithms to make decisions based on data analysis. The major challenge to explainability in AI stems from many AI models being opaque. Significant breakthroughs in AI have resulted in the emergence of numerous opaque models, as many incorporate deep learning. Deep learning models employ multi-layered neural networks that comprehend patterns based on vast datasets by adjusting weights across several layers. Instead of each step in decision-making being clearly specified, neural networks find their own path to draw conclusions.Consider a scenario wherein such an AI model examines CT scans to predict the chances of lung cancer in a person and estimates an 80% chance of malignancy. If it does not provide any plausible explanation regarding the rate of metastasis, highlighted tumor areas, or shape or texture-based characteristics, it certainly does not adhere to the principle of explainability. Consequently, it would delay the process of further diagnosis and treatment, which may prove fatal.In yet another scenario, if we request a refund for a damaged item bought online and the e-commerce site, using a chatbot for customer support, responds, 'Your request is denied' without specifying any reason, such as 'No refund possible after 20 days,' it essentially does not qualify as explainable AI.To counter such problems, AI models have begun to include explainable AI approaches like LIME (Local Interpretable Model-agnostic Explanations), Integrated Gradients, SHAP (SHapley Additive exPlanations), Activation Maximization, etc. Although the adoption of such techniques is rising, achieving explainability to a satisfactory level continues to be a Herculean task. As businesses witness the evolution of AI, striking the right balance between model performance and explainability becomes paramount. Continued research and improved methodologies can enable organizations to implement explainability appropriately, marking a significant leap toward 'Responsible AI'.By Jigyasu Pant, Application Developer, IBM.Note: This article is a part of ETCIO's Brand Connect Initiative.