Responsible and Trustworthy AI in Digital Lending: From Principles to Practices

Posted On Date: March 29, 2025

Save Post

Abstract

The interest and adoption of Artificial Intelligence (AI) in financial services is rising consistently. Latest industry reports suggest that the financial sector, specifically fintech, is leading the adoption of AI in India.[i] It also appears that among the different sectors in the economy, fintech is one of the best poised to make commercial gains from the adoption of AI.[ii]

The potential gains from the adoption of AI in financial services have been well documented. To summarize, the gains from AI could accrue on three counts. First, enhanced data-processing abilities, including the ability to process qualitative and audio-based data could lead to a deeper understanding of customers’ needs and improve the fit of products. These insights could also significantly improve the customer’s journey by sensing their needs and offering customized, relevant, and timely support in a customer-friendly format along the customer journey. In addition to deeper personalization of product design and customer support, AI systems also help financial service providers (FSPs) build stronger defenses against fraudulent activity. Second, AI systems exhibit a high degree of scalability and flexibility.[iii] They appear to be capable of solving the wicked dilemma of offering hyper-personalization at scale. Finally, AI systems help realize efficiency gains from process-rationalization,[iv] where providers are able to eliminate duplication of tasks by deploying AI. When AI works as intended, it can deepen financial inclusion and enhance the relevance of financial services at population-scale. For instance, Generative AI (GenAI) can significantly enhance the ease of opening accounts for uninitiated customers. It can also nudge them to improve account usage, promote budgeting, and deepen financial literacy through relevant, customized and timely content. In the case of credit, fueled by big data, algorithms could do a better job in predicting creditworthiness of thin-filed customers and credit invisibles.[v]

These gains, however, are tempered by attendant risks. The first set of AI-related risks arises from the advanced processing of rich and personal data. This includes the risk of infringement of privacy, the risk of proliferating bias and discrimination, the AI systems generating ‘hallucination’ and misinformation, and the reduced robustness i.e., inconsistent accuracy of the AI system. These risks can take away from the gains presented by deeper processing capabilities. Further, these risks are aggravated by the relative scalability of algorithms.[vi] Just like the benefits, the risks also get scaled with the model, affecting vast swathes of customers at once. Finally, the difficulties in explaining complex algorithms create anxiety at two levels: On the one hand, it makes the algorithm impermeable and therefore hard to assess for accuracy. It significantly reduces the ability of the customers to question the algorithm, identify mistakes or seek remedies. On the other, it pronounces the philosophical tension between human autonomy and agency and automated decision-making. Inscrutable algorithms chip away at human agency, expecting humans to comply with machines they do not fully understand, nor can completely control. While these risks appear to be esoteric, they can trigger adverse systemic shifts in the financial system and jeopardize customer protection. For instance, reliance on similar, off-the-shelf machine learning tools could encourage herd behaviour among lenders and create procyclical vulnerabilities.[vii] Similarly, a less-than-fit algorithm could lend to borrowers who may not have the wherewithal to repay, causing distress and default. It could affect the borrowers’ credit scores, cutting them off formal credit markets and severely erode the lenders’ portfolio quality.[viii]

An appreciation of the risks and benefits of using AI is imperative for an informed policy stance. This knowledge of the risks and benefits must also be buttressed by an understanding of their underlying causal mechanisms. Collective wisdom[ix] now suggests that AI governance must adopt a lifecycle or a value-chain approach. A value-chain approach to AI focuses on improving the visibility over the various components of an AI system and the stakeholders responsible for each. It allows us to unpack the origins of risks and benefits along each link of the chain and allocate proportionate responsibilities to every actor. This visibility over the different components, their potential implications and the respective stewards also offers comfort to the regulator. This approach also allows for the distribution of legal responsibility. A governance framework anchored in the value-chain approach is most likely to be proportionate and both preventive and remedial.

Finally, this value-chain approach to governance is a necessary precondition to craft an AI governance framework that focusses on Responsible and Trustworthy AI. The paradigm of Responsible and Trustworthy AI is fast gaining traction and is poised to guide the way for ethical, safe and inclusive application of AI.[x] A quick review of literature surfaces two essential components of Responsible and Trustworthy AI (RTAI). First, it concerns itself with the AI system as-a-whole and not just the outcomes of the AI. Second, it requires AI systems to align with socially desirable values such as non-discrimination, fairness, that they be technologically robust and safe and amenable to being held to account. It appears that sometimes the terms responsible and trustworthy are used interchangeably. This Whitepaper, however, differentiates the two attributes by requiring that the processes associated with the design and deployment of AI be responsible so that the conduct and the outcomes of the AI system thus designed, are trustworthy. Thus, in this Whitepaper, responsible describes the processes and procedures put in place to ensure that the conduct, decision and outcomes of the AI systems are trustworthy.

This Whitepaper unpacks what RTAI would mean in the context of digital lending. While the term Responsible AI has been adequately conceptualized in academia and elsewhere, it still needs to be coherently contextualized to specific domains. Therefore, building on a systematic review of literature, the first section compiles principles of RTAI along with its essential components. The next section maps relevant tools to each principle. Put another way, these tools are practice-recommendations to digital lenders that can help them perform better on the specific characteristics of RTAI. The unique value addition of this publication is a distance map that allows digital lenders to gauge how far their current AI safeguards are from the desired level and how they might close the gap. The distance map takes the form of a checklist; designed for the technology teams of digital lenders. They should be able to run down the checklist without external supervision and reflect on the overall intensity of their AI safeguards. It serves as a diagnostic tool surfacing relevant areas and offers direction to lenders on how they might further strengthen their AI practices. The distance map is anchored in the understanding that AI in finance needs to be governed not for AI’s sake but for the sake of finance. Responsible AI in lending is necessary for responsible lending. Therefore, to AI experts, this list of tools may appear non-exhaustive because it may not capture all the upcoming practices and measures. However, they have been deliberately curated to prioritize financial concerns and to be compatible with the technology being used by digital lenders. By distilling the characteristics of RTAI into specific practices that are amenable to the existing technological capabilities of digital lenders, this Whitepaper contributes to narrowing the principle-practice gap that exists in the research.

The full report is available here.