Beyond Algorithms: Navigating Fairness in India's Lending Landscape

Algorithmic discrimination in digital lending occurs when algorithms unfairly privilege certain individuals or groups, leading to unequal access to credit. To address this, the Reserve Bank of India mandates non-discrimination in lending practices and is examining ethical AI use. Achieving fairness requires societal awareness, legal frameworks, and responsible AI initiatives.

Fairness is perceived in different forms that serve different contexts. When used as a direct measure to address underlying systemic discrimination, it brings value and equity to society. This shows that if equity were the goal, fairness could be a standard and consistent measure across geographies.

So, when a woman entrepreneur launching her small business in India seeks credit, can the same measures adopted for fairness in the West be applied in the Indian context? The simple answer is no.

Digital India is as diverse as the country. When Amarjit, 29, a self-employed contractor in the construction industry in Bathinda, a city in the northern state of Punjab, and Srikanth, 30, a gig economy freelancer in Bengaluru, an urban metropolis, both with a similar educational background and similar income levels, were offered credit, but at different limits from the same digital lending app, we were intrigued. While the lender priced credit differently for competitive advantage, how do you solve digital inequity across the country?

Algorithmic discrimination occurs when an algorithm offers the privilege to any individual or demographic over another in ways different from the intended function of the system, resulting in unfair outcomes. In digital lending, this could result in a particular individual or a group they belong to being denied credit, offered less credit or credit at unequal terms, based on a decision primarily influenced by Automated Credit Decision Systems. Traditional credit scoring systems may perpetuate bias coupled with alternative data sources and AI/ML models; such systems can reinforce existing biases.

There is a renewed focus on financial inclusion, promoting financial education and literacy, and making credit available to all productive sectors of the economy. The Reserve Bank of India, the country's banking regulator, has mandated banks and non-banking finance companies to establish a 'Fair Practices Code' for lending. This includes ensuring non-discrimination based on race, caste, religion, gender, marital status, age, financial status, or physical disability. The regulator has notified recently that ethical AI and non-discrimination in algorithmic decisions are areas they are examining currently.

Pre-existing social context

Humans have biases. Human society has biases influenced by individual and cultural perceptions regarding age, gender, marital status, location, education, and income levels. Any individual's digital footprint is a mirror of these perceptions.

The data on which Automated Credit Decision Systems is trained might contain pre-existing social biases. Data engineers and data scientists address them using a structured approach to examine the data and assess the need to balance them. In most cases, spotting which data trends are unbalanced due to pre-existing social context is complex. It's pertinent to note that social context varies across the 28 states and eight union territories in India at any point in time and over time.

For example, although India decriminalized same-gender relationships with a historic Section 377 judgment, it is yet to allow registering marriages in the LGBTQA+ community. The banking system is slowly adapting to such change, and the gaps in historical data, including inadequate data on these individuals, may cause discrimination.

The causality and reliability conundrum

Data scientists use document classification algorithms for classifying collaterals, extraction algorithms to extract data from bank statements or identity documents, tree-based algorithms for lending decisions, fraud prediction, income and credit limit determination, and optimization algorithms for specific objectives. Automated Credit Decision Systems use data that include personal identity information, sources of income, location, and multiple credit bureau scores, besides the nature of collaterals in determining credit eligibility and affordability. Given the nature of data gathered and the kind of predictions arrived at, discrimination among specific individuals or groups may occur with the model outcomes.

For instance, some digital lending apps require access permissions to all contacts and call logs on the smartphone, quite simply to identify potential fraudulent customers, based on the concept of homophily; so that connections to known fraudulent entities could be flagged; however, such models are prone to false positives. In India, people routinely own more than one device and mobile number for convenience; the model assumes that the phone number to which short messages are received is indeed the number tied to banking transactions. Models also use short messages to determine the income or affordability levels of the prospects for a loan. Further, some models assess the geographical proximity of location to previously identified customers as a collections risk indicator. This again may have causality problems, besides reliability issues associated with such predictions.

Affirmative action policy

Affirmative action policies primarily focus on disadvantaged populations and contain measures to ensure financial inclusion. For example, a State may allow a subsidy in loan interest for a scheduled caste or scheduled tribe, officially designated groups of people among the most disadvantaged socio-economic groups in India. At the same time, the Central government may not have such a subsidy. Similarly, there are schemes for loans to the differently abled, loans to women via self-help groups, and loans to women-run businesses in rural areas. A one-model fits-all approach is unlikely to be effective!

While Amarjit did not get the credit on equitable terms, and raised the issue with the digital lending app, only to be asked to apply again citing lack of sufficient data on his credit file. While concepts of parity and equal odds to ensure fairness in an algorithmic context appear relevant, they require substantial other considerations to enable fairness. Using an accountability aligned process (as detailed below) can help in identifying potential fairness gaps and accounting-ably support in addressing them early on. As regulations and practices evolve, fairness considerations in India may undertake adaptive approaches that consider constraints specific to the Indian context and ground realities.

Given the complexities associated with dealing with bias, follow the 8 steps given below for consistently managing them on an ongoing basis:

• To assess and address bias in training datasets, compile relevant datasets and carefully analyse them to identify any instances of exceptions or outliers. These exceptions may include lending decisions that are influenced by regulatory (limit lending in specific geographies where potential terrorist threats exist) or business policy constraints (increased focus on women entrepreneurs), which should be eliminated to ensure fairness. Examine and understand the ways in which social biases are embedded in the data.
• Compare the statistical characteristics of protected variables (such as gender, religion, or disability) in the organization's target market with those present in the training or real-world data. This comparison helps to understand any disparities and potential biases that exist.
• Establish key metrics that will be used to evaluate bias and set clear thresholds for measuring them consistently. Defining these metrics is essential to monitor and track bias throughout the development and deployment of algorithms. Take adequate care to understand the relevance of the metrics in Indian context and accordingly align the variables for measuring them consistently.
• Define measures for handling biased results, considering both process and algorithmic approaches. Process measures may involve incorporating human validation or human-in-the-loop considerations to ensure fair decision-making. Algorithmic approaches should be implemented to address and mitigate bias effectively.
• Implement a mechanism to address exceptions identified based on the established metrics and ongoing monitoring of automated decision algorithms. This shall also include identifying failure modes where the metrics or monitoring mechanism provides significant false positives. This mechanism should be designed to rectify and correct any biases detected in the decision-making processes and address errors in bias detection metrics or measures.
• Establish a mechanism that provides adequate interpretability and explainability of algorithmic decisions. This allows customers and users to understand and provide feedback on the decisions made by the algorithms, ensuring transparency and accountability.
• Conduct periodic bias audits or testing to identify any exceptions or inconsistencies that may arise over time. Regular assessments help to proactively identify and address biases that may emerge as the algorithms are deployed and used in real-world scenarios.
• Analyse the overall trends of bias and assess the effectiveness of the current mechanisms in mitigating bias. This examination of biases and the impact of mitigation efforts helps organizations refine their approaches and ensure continuous improvement in reducing bias in algorithmic decision-making.

Establishing fairness in digital lending would require social awareness of potential discrimination, legal frameworks defining discrimination, and Responsible AI initiatives in the industry intertwined with algorithmic metrics aligned to Indian societal realities.

Keywords: AI Ethics, Algorithm Bias, Lending System, Responsible AI

Disclaimer: This article is an opinion of the authors.