Abstract

Digital lending promises to expand access to finance for borrowers excluded from traditional credit systems. Yet women in India continue to face a striking paradox. They outperform men on conventional measures of creditworthiness but receive less than a quarter of the country’s outstanding credit, with only14 per cent of India’s 453 million credit-eligible women accessing formal credit. This paper argues that these disparities cannot be explained by borrower risk alone. It conceptualises digital lending as a five-stage socio-technical pipeline comprising data generation, verification and linkage, borrower profiling, algorithmic scoring, and credit allocation, showing how gendered inequalities become embedded and compounded across successive stages. Bias is co-produced through data, infrastructure, institutional practices, and model design, and accumulates across the pipeline, making technical fixes at the scoring stage insufficient. The paper proposes policy pathways across the lending lifecycle to reshape data practices, model governance, and credit-risk allocation.

Introduction

Digital lending platforms are transforming how creditworthiness is evaluated, allowing financial services to reach populations historically excluded from formal banking systems. Driven by rising mobile phone ownership, internet connectivity, and rapid uptake of digital financial services, bank account ownership increased from 51 per cent in 2011 to 79 per cent globally in 2025. The spread of digital technology has also expanded the range of information available to lenders, enabling them to extend credit to thin-file borrowers who lack formal financial histories. 

Traditionally, creditworthiness was assessed using a relatively narrow set of indicators, including collateral ownership, credit scores, employment records, banking histories, and prior repayment behaviour. While these approaches worked reasonably well for borrowers with established relationships with formal financial institutions, they often excluded individuals with limited credit histories, informal, cash-based incomes, or no collateral. Digital lending emerged in response to this gap, catering to the vast population that was invisible to the traditional lending system. Models such as point-of-sale lending, platform-integrated credit, invoice discounting, and marketplace lending exemplify this shift, relying on transaction-level and platform-derived data to inform lending decisions. 

Yet creditworthiness is not simply a neutral reflection of repayment capacity. Rather, it is produced as an outcome of how data is selectively captured, structured, and interpreted. Borrowers whose economic lives align with platform-mediated, digitised patterns are more likely to be visible within these systems, while those operating in informal or intermittent economies may be excluded or misclassified. Women often face structural barriers to digital participation and are less likely to generate the forms of data increasingly used to assess creditworthiness. As a result, digital lending systems may reproduce existing inequalities even while expanding access to finance. This paper argues that gender bias in digital lending emerges not at a single point of algorithmic decision-making but cumulatively across the lending pipeline – from data generation and verification to profiling, scoring, and credit allocation. By conceptualising digital lending as a socio-technical process, the paper shows how existing inequalities become embedded within data systems and reproduced through lending decisions over time.

India, occupying a unique position in the evolution of digital lending, provides an especially important context in which to examine these dynamics. Unlike many high-income countries, where lending systems still rely heavily on legacy credit bureaus and fragmented financial data, or in low- and middle-income countries where digital credit is built around mobile money and telecom data, India has developed an integrated Digital Public Infrastructure (DPI) ecosystem. This includes Aadhaar for digital identity, UPI for payments, DigiLocker for digital documentation, and the Account Aggregator framework for consent-based data sharing. This environment has enabled lenders to access new forms of digital lending that draw on diverse data sources, including agricultural records, digital payment histories, small business transactions, and field-level assessments.

At the same time, international evidence suggests that digital lending also introduces new risks. Studies from Kenya and Tanzania show that access to digital credit has been concentrated among younger, urban, and more digitally connected populations, and raise concerns about the risk of overindebtedness, algorithmic bias, and demographic discrimination. These concerns underscore the need to examine not only whether digital lending expands access to credit, but also whose economic lives become visible within data-driven systems and on what terms.

Digital Lending as Socio-Technical Infrastructure: How do Data, Algorithms, and Gendered Biases Intersect?

Alternative Data for Construction of Creditworthiness

The growth of digital lending has been accompanied by the increasing use of alternative data: information beyond conventional financial data points, including digital payment histories, mobile phone usage, platform transactions, utility payments, geospatial information, and behavioural indicators. In India, lenders draw on a wide range of such alternative data sources. Social fintechs, such as Vest in Villages, combine household-level financial and social information collected by field agents to construct borrower profiles, while agri-tech platforms, such as Dvara e-Registry, use satellite imagery, land productivity indicators, and climate-related information to estimate agricultural incomes. Other lenders, such as Kaleidofin, employ behavioural scoring systems that draw on patterns of financial activity and digital engagement.

Despite their differences, these models share a common assumption: that economic life can be understood through the data trails it generates. Data, therefore, becomes the basis through which borrowers are rendered visible within lending systems. However, some have argued that data is not a neutral representation of reality. Gitelman (2013) observes that “raw data” does not exist independently of the systems that collect and organise it, while Van Dijck (2014) describes “datafication” as the process through which social life is translated into measurable and trackable forms. Digital lending systems therefore do more than observe economic activity; they shape what forms of economic activity become visible as evidence of creditworthiness.

This has important implications for financial inclusion. Economic activities that are irregular, cash-based, home-based, or community-mediated often leave weaker digital traces than platform-mediated or formally documented transactions, despite contributing substantially to livelihoods and household welfare. Consequently, digital lending systems may privilege borrowers whose economic lives are already embedded within digital formal infrastructures while underrepresenting those operating outside them.

Gendered Inequalities in Data Generation and Financial Visibility

Women face a distinct set of structural constraints that affect both participation in digital systems and visibility within data-driven credit markets. In India, they remain less likely to own mobile phones (36 per cent compared to 58 per cent), access the internet on a mobile phone (39 per cent versus 58 per cent), or use digital financial services such as online bill payments (10 per cent compared to 20 per cent of men). These disparities extend beyond access alone. Women rely more on shared devices and SIM cards, possess lower levels of digital literacy, and face affordability, mobility, and social norm constraints that limit independent participation in digital ecosystems. Beyond digital access, women’s lower asset ownership (66 per cent of women do not hold land in India compared to 40 per cent men, restricted mobility, concentration in informal and home-based work, and weaker control over household finances often limit their ability to generate the regular, individualised, and digitised transaction histories that increasingly underpin algorithmic credit assessment.

India has significantly expanded financial inclusion through initiatives such as the Pradhan Mantri Jan Dhan Yojana, but access to accounts does not necessarily translate into active digital participation or financial autonomy. Similarly, despite the near universal coverage of Aadhaar in India, disparities in access to supporting documents, mobile-linked accounts and the ability to navigate digital verification systems remain substantial. Difficulties obtaining identity documents, meeting KYC requirements, and complying with SIM registration procedures further restrict women’s access to digital financial services. Gender, however, does not operate in isolation. Digital exclusion is shaped by the interaction of gender with caste, class, geography, disability, age, religion, and other forms of marginalisation. Research on caste-based digital inequality in India demonstrates persistent disparities in device ownership, internet access, digital skills, and digital participation between Scheduled Castes (SCs), Scheduled Tribes (STs), Other Backward Classes (OBCs), and dominant caste groups. Evidence from credit markets similarly shows that lower caste borrowers experience lower participation rates and less favourable lending outcomes even after accounting for socioeconomic characteristics. These inequalities shape who is able to generate the forms of data that digital lending systems recognise as indicators of creditworthiness.

Much of the literature relies on binary comparisons between women and men, leaving the experiences of transgender, non-binary, and gender-diverse individuals underexplored. Some evidence from the Global North suggests that gendered documentation requirements and biometric verification systems, for instance, can create mismatches between lived identities and official records, leading to exclusion or misidentification within digital systems. These challenges highlight the need for a broader understanding of how digital lending systems recognise, classify, and evaluate borrowers across diverse social identities.

Algorithms, Bias, and the Governance of Digital Lending

A central paradox lies at the heart of contemporary lending systems. In India, women borrowers outperform men on several conventional measures of creditworthiness. Around 57 per cent of women hold prime credit scores compared to 51 per cent of men, while their loan delinquency rate stands at 5.2 per cent compared to 6.9 per cent for men. However, women account for less than a quarter of total credit outstanding, and formal lenders serve only 14 per cent of the country’s estimated 453 million credit-eligible women. Women, therefore, appear less risky according to the very metrics that lending systems use to assess repayment behaviour, but continue to receive substantially less formal credit. This suggests that unequal lending outcomes cannot be explained by borrower risk alone. Instead, it raises questions about how creditworthiness is constructed, measured, and translated into lending decisions within digital financial systems.

As data generation translates into decision-making, algorithms play an increasingly important role in transforming borrower information into credit outcomes. Algorithmic systems can reproduce existing inequalities when they are trained on incomplete, historically biased, or unrepresentative datasets. Bias can emerge in various ways. Representation bias may occur when women are underrepresented within training data, while proxy bias arises when seemingly neutral variables such as device ownership, transaction frequency, or mobility patterns indirectly reflect gendered social constraints. Credit bureau records and prior lending outcomes also embed historical bias, causing past inequalities to be reproduced within future lending decisions.

At the same time, the relationship between algorithms and inclusion is not uniformly negative. Some studies suggest that fintech-based lending models can reduce forms of human discretion and bias that have historically disadvantaged women and other underserved borrowers. By incorporating broader forms of information than traditional underwriting systems, digital lending may identify creditworthy borrowers who would otherwise remain invisible within conventional credit markets. Evidence from the United States suggests that fintech credit algorithms can improve approval rates for women borrowers, particularly in institutions with historically high levels of gender bias. The benefits of algorithmic lending, however, remain contingent on the quality, representativeness, and governance of underlying data systems. Algorithmic fairness cannot be achieved through technical design alone but requires institutional mechanisms of accountability. Simply removing protected attributes such as gender or caste is often insufficient, as machine learning models can infer these characteristics through proxy variables embedded in behavioural and transactional data. Meaningful oversight, therefore, requires making disparities visible and auditable. This is particularly relevant in India, where initiatives such as Priority Sector Lending, Self-Help Group (SHG)-bank linkage programmes, and targeted credit schemes for SCs, STs, and women already recognise that access to finance is shaped by structural inequalities rather than individual financial behaviour alone.

Digital lending should thus be understood as a socio-technical system that reconstructs creditworthiness through data, algorithms, and social and institutional practices. While digital lending has the potential to expand access to finance, existing inequalities in digital participation, economic visibility, and data representation may also be reproduced within these systems. 

Existing research on digital lending tends to examine specific parts of the lending process, focusing on algorithmic bias, barriers to digital and financial inclusion, or regulatory responses. This paper brings these strands together by conceptualising digital lending as an end-to-end socio-technical process. We propose a process-based digital lending data pipeline that traces how social realities become digital signals, how exclusions and biases emerge across different stages, and how lending outcomes can reinforce these inequalities over time. Building on this framework, we identify policy interventions to promote more equitable digital lending systems.

Digital Lending as a Data Pipeline 

Literature on digital lending and algorithmic decision-making can broadly be grouped into three strands.

First, digital financial inclusion literature examines how individuals and enterprises move through different stages of access and adoption within digital financial systems. Frameworks developed by the World Bank and others focus on the progression from financial access to meaningful use of digital financial services, highlighting barriers related to infrastructure, affordability, and digital capability.

Second, algorithmic fairness literature focuses on how bias emerges within machine learning systems. Scholars have developed lifecycle-based frameworks that identify sources of bias across data collection, model development, deployment, and monitoring. More recent work extends these analyses to gender bias in AI systems. These approaches are valuable for identifying technical sources of discrimination but tend to focus either on data inputs or model design.

Third, socio-technical and digital inequality work situates digital systems within broader structures of power and exclusion. They argue that technologies shape social and economic participation. Digital inequalities emerge not only through unequal access to technology but also through differences in skills, usage patterns, and opportunities to participate in digital systems. In the context of digital lending, these inequalities shape who generates data, whose economic activity becomes visible, and whose behaviour is recognised as creditworthy.

Recognising that risks emerge across multiple actors and stages, ecosystem-level governance frameworks have also begun to examine digital lending as a distributed process. For example, CGAP’s (2024) Responsible Digital Finance Ecosystem framework demonstrates that consumer harms arise through interactions between lenders, platforms, regulators, and data infrastructures rather than from a single point of failure. However, existing frameworks rarely connect these different levels of analysis. Less attention has been paid to how gendered inequalities and biases accumulate sequentially across the lending process itself from data generation and verification to profiling, scoring, and credit allocation. Our digital lending pipeline below traces how social inequalities become translated into data, embedded within infrastructures, and reproduced through algorithmic and institutional decision-making.

Stage 1: Data Generation and Signal Formation

The digital lending pipeline begins when economic activity is translated into digital signals. However, not all economic activities get converted into data. Activities conducted through digital platforms, payment systems, formal financial accounts, and other structured data-collection processes become visible within lending systems while cash-based, informal, home-based, and community-mediated economic activity often remains outside the data environment altogether.

This creates a form of representation bias that emerges before any algorithmic assessment takes place. As Gitelman argues, data is always shaped by prior decisions about what should be measured and recorded. Women are disproportionately concentrated in informal, home-based, intermittent, and care-related work, much of which generates few digital traces despite contributing substantially to household and local economies. 

Gender gaps in smartphone ownership, internet access, and limited engagement with digital financial services further limit the ability of women to generate a digital footprint that increasingly serves as an input into credit assessment. As a result, subsequent stages of the pipeline operate on an incomplete representation of borrowers’ economic lives, embedding structural inequalities into the lending process from the outset.

Stage 2: Data Verification and Linkage

Now, digital signals must be verified and linked to a recognised financial identity before they can be used within lending systems. Data that cannot be successfully authenticated does not progress further through the lending pipeline. In digital lending systems, identity verification systems measure and recognise some users more easily than others. When verification tools like biometric systems, matching algorithms and KYC protocols perform differently for different demographic groups, they tend to introduce measurement bias by systematically misreading some users over others. India’s digital verification mechanism implicitly assumes continuous device access, consistent identity documentation, and the ability to independently navigate digital authentication processes. 

For women, gendered patterns of shared device use, limited control over mobile phones, lower levels of digital autonomy, and inconsistencies in identity documentation can create barriers to successful verification. Authentication systems often rely on one-time passwords, device-linked identities, or exact matches across records, making routine disruptions such as shared phone access or name changes after marriage appear as technical verification failures rather than manifestations of broader social constraints. Unlike Stage 1, where economic activity may never generate a digital signal, the data exists here but cannot always be successfully attributed to a verified borrower. Borrowers who experience authentication failures enter later stages of the pipeline with incomplete records, fragmented identities, or no usable data at all, further limiting their visibility within digital lending systems. 

Stage 3: Creation of Borrower Profiles

Once digital signals have been generated and authenticated, they must be transformed into something that lenders can interpret as evidence of creditworthiness. This involves a series of institutional decisions through which Lending Service Providers (LSPs), credit bureaus, and lenders aggregate, filter, and prioritise information to construct a borrower profile.

At this stage, multiple forms of data are combined, including bureau records, bank account histories, digital payment transactions, platform activity, device-related information, field-level assessments, and sector-specific indicators such as agricultural productivity. Institutions decide which data sources to include, identify signals predictive of repayment, and assign a weightage to each data point. The result is a curated profile built from selected indicators rather than a complete representation of a borrower’s economic life.

Because visibility is driven by particular forms of data, some forms of economic participation remain obscure. Even when women borrowers enter the pipeline, the informal, intermittent, and care-based activities are often translated into weaker or incomplete signals, creating a form of selection bias in how creditworthiness is constructed. Historical data can further reinforce these exclusions. Credit bureau records are commonly incorporated into borrower profiles as indicators of past financial behaviour, but bureau scores themselves reflect earlier lending decisions and prior access to formal financial services. Borrowers who have historically faced barriers to formal credit markets may therefore enter digital lending systems with thinner records or weaker bureau profiles, carrying forward the effects of previous exclusion. 

Stage 4: Algorithmic Scoring and Risk Classification

Once a borrower profile has been constructed, algorithmic models translate it into a quantified assessment of credit risk. Multiple data points are combined into a single score or risk classification that directly shapes lending decisions.

Although often presented as objective, these models are trained on data and assumptions inherited from earlier stages of the pipeline. As a result, existing inequalities can become embedded within risk predictions. Variables associated with fragmented financial histories, shared device use, limited formal credit experience, or other structurally constrained behaviours may be interpreted as indicators of elevated risk even when they do not reflect repayment capacity. In this way, historical and social inequalities are translated into seemingly neutral statistical predictions.

A further challenge lies in the opacity of many scoring systems. As Pasquale (2015) argues, commercial credit models frequently operate as “black boxes”, making it difficult for borrowers, regulators, and even lenders to understand how decisions are reached. Unless outcomes are regularly assessed across gender and other social categories, discriminatory patterns may remain hidden. For borrowers, these scores can influence not only loan approval but also loan size, pricing, and lending conditions, often with little visibility into how the assessment was made.

Stage 5: Credit Decisions and Feedback Loops

The final stage translates risk scores into lending outcomes, including approval or rejection, loan size, interest rates, repayment terms, and product type. While algorithmic scores play a central role, institutional discretion, business rules, and product design choices continue to shape how credit is ultimately allocated.

The defining feature of this stage is that lending outcomes generate new data. Repayment histories, defaults, transaction records, and borrowing behaviour become inputs into future credit assessments through bureau records, digital transaction trails, and model training datasets. The output of one lending cycle, therefore, becomes an input into the next.

This creates the possibility of recursive bias. Borrowers who receive smaller loans or are repeatedly denied credit have fewer opportunities to build the financial histories that future models interpret as evidence of creditworthiness. As a result, inequalities inherited from earlier stages can become reinforced over time. What begins as a one-time disadvantage may gradually evolve into a persistent pattern of exclusion, making past lending decisions appear validated by the very data they helped produce.

Policy Strategies for Gender Equitable Digital Lending

The analysis of the digital lending pipeline suggests that gender bias emerges across data generation, verification, profiling, scoring, and lending decisions. Because these exclusions are structurally embedded, equitable lending cannot be achieved through technical corrections at the point of algorithmic decision-making alone. It requires upstream interventions that reshape what data is collected, how identities are recognised, how models are governed, and how credit risks are distributed. In a highly regulated sector such as finance, these changes will require active government leadership alongside action by lenders, technology providers, and digital infrastructure operators. We therefore propose four possible pathways to advance gender-equitable digital lending.

Expand data sources used to assess creditworthiness 

Digital lending systems rely heavily on signals derived from formal, digitised, and platform-mediated activity, creating blind spots for borrowers engaged in informal work, collective savings, household enterprises, and cash-based transactions – forms of economic participation in which women are disproportionately represented.

While alternative credit-scoring models demonstrate the potential of broader data sources, many datasets that capture women's economic activity remain excluded from lending pipelines. These include SHG and Joint Liability Group (JLG) records, Direct Benefit Transfer (DBT) inflows into Pradhan Mantri Jan Dhan Yojana (PMJDY) accounts, utility payment histories in women's names, and repayment records from Kisan Credit Card programmes. Incorporating such datasets would expand recognised creditworthiness signals and improve the visibility of women's economic activity.

In parallel, community field officers can serve as data intermediaries, digitising informal records and providing field-based verification for borrowers facing authentication barriers. These approaches could be piloted through a joint Reserve Bank of India (RBI)-National Bank for Agriculture and Rural Development (NABARD) regulatory sandbox with standardised data collection protocols and gender-disaggregated reporting.

Redesign authentication and verification systems for low digital literacy users

Improving inclusion requires adapting digital public infrastructure to the needs of users with lower digital literacy and more constrained digital autonomy. The Account Aggregator (AA) framework in India has standardised consent for financial data sharing and has high adoption among Financial Information Providers (FIPs), but it can be expanded to serve cooperative banks, regional rural banks, and microfinance institutions. This would require amendments to allow individual account holders to share joint account data without requiring co-holder consent. This eliminates the current mechanism that grants male co-holders structural power over women's credit applications. Regulators should also mandate that, among priority sector lenders such as MFIs and RRBs, repayment histories and SHG records be included in the FIP ecosystem. Additionally, introducing inclusive interfaces to independently test voice-mediated and CFO-assisted consent flows for users with low digital literacy will complement the uptake. 

Mandate gender responsive algorithmic accountability 

Bias in digital lending cannot be addressed through technical fixes alone. Decisions made during data collection, feature selection, model training, and deployment can all contribute to discriminatory outcomes, even when protected characteristics such as gender are excluded from models. The current regulatory regimes mandate data privacy but lack fairness protocols. In the financial sector, self-regulation norms are not as effective and enforceable in addressing gender equity. 

Regulators should therefore undertake gender-disaggregated audits of algorithmic decisions to assess outcomes such as approval rates, loan ticket size, interest rate, and cases of false negatives. Institutions should further be obligated to address identified disparities by using explainability tools to detect proxy discrimination and implementing concrete remediation measures whenever models exhibit biased or miscalibrated performance. 

Disrupt the cycle of exclusion through blended finance instruments

The greatest risk in digital lending is not a single biased decision but the creation of self-reinforcing cycles of exclusion. Borrowers excluded earlier in the pipeline are denied the initial credit necessary for building a digital footprint, thus amplifying the discrimination at stage 5. 

Targeted risk-sharing mechanisms can help interrupt this cycle. Existing initiatives like RBI’s First Loss Default Guarantee framework (2023) offer risk-sharing systems, thereby providing a foundation for encouraging lending to underserved groups, but could be more explicitly aligned with gender inclusion objectives. 

Development finance institutions can further strengthen these incentives by linking first-loss coverage to gender-focused lending targets, such as requiring that a substantial share of the supported portfolio be held by women, thereby aligning risk-sharing arrangements with broader goals of gender-inclusive financial access. 

Way Forward

By tracing the movement of data within the lending pipeline from generation and verification to borrower profiling, algorithmic scoring, and credit allocation, this paper argues that gender bias is cumulative and systemic. At the same time, the analysis is constrained by the opacity of commercial lending models and an evidence base that often relies on outcome disparities rather than direct observation of algorithmic processes. Experiences of exclusion are also shaped by intersecting factors such as caste, class, geography, disability, and digital literacy, many of which remain underexplored. As digital lending systems continue to evolve, the central challenge is to ensure that the data, infrastructures, and governance arrangements on which they rely do not reproduce existing inequalities in new digital forms. Achieving genuinely inclusive finance will therefore require scrutiny and intervention across the lending pipeline as a whole.