In Nairobi’s digital-lending ecosystem, credit scores are not neutral outputs but products of continual technical and political work; practitioners and regulators co-shape models and markets using data workarounds, multiple risk framings, and legal strategies.

Main Finding

Algorithmic credit scoring in Kenya is not a purely technical or neutral refinement of risk measurement but a sociotechnical accomplishment achieved through ongoing alignment work. Practitioners create and translate alternative data, construct multiple operationalizations of “risk,” and negotiate model performance via technical and political interventions. Alignment operates bidirectionally: models are made “safe for worlds” while social, institutional, and market practices are reshaped to be “safe for models.” This stabilizes decision-making under persistent epistemic, modeling, and contextual uncertainty but produces significant distributional, privacy, and governance consequences.

Key Points

• Three interlinked forms of uncertainty disrupt credit scoring:
  • Epistemic uncertainty — limits on what can be known or measured about borrowers.
  • Modeling uncertainty — gaps between model claims and real-world deployment contexts.
  • Contextual uncertainty — discretion, strategic behavior, and social practices that alter model inputs and outcomes.
• Alternative data practices:
  • Firms (startups, telcos, banks) construct “alternative data” from mobile-money flows, airtime use, SIM- and device-level signals, and other nontraditional traces.
  • These data are often produced through technical and legal workarounds (data engineering, cross-party agreements, creative feature engineering) rather than through passive reuse of neutral records.
  • The making of alternative data is performative: data engineering choices reshape how borrowers are represented and how risk is enacted in markets.
• Negotiation of model performance:
  • Teams balance explainability, regulatory compliance, and business objectives; performance claims are validated and contested across engineers, product managers, legal teams, and regulators.
  • Non-technical levers (pricing, collection practices, product design) are used alongside algorithmic adjustments to manage defaults and perceived risk.
• Two-way alignment:
  • Alignment is not only tuning models to reflect social goals but also reshaping borrower behavior, institutional practices, and market rules to fit model affordances.
  • This mutual shaping underpins the apparent success of scoring systems while embedding power asymmetries (e.g., incumbents’ data advantages).
• Real-world context:
  • Kenya’s fintech ecosystem (e.g., dominant telco platforms, high mobile-money penetration) creates both opportunities and risks: rapid market expansion, high default rates, aggressive debt collection practices, and emergent regulatory responses.
  • Borrowers adopt strategies (SIM changes, refinancing, evasive practices) that feed back into model uncertainty and require further alignment work.

Data & Methods

• Design: Nine-month ethnographic study in Nairobi (Feb–Nov 2025).
• Sites and actors: Observations and engagements with digital-lending startups, established telcos and banks, product and data science teams, regulators, and borrowers/communities.
• Methods: Participant observation, shadowing of data science/product teams, interviews across organizational roles, and analysis of regulatory and company practices and artifacts (models, features, data pipelines, policies).
• Contextual inputs cited in the study: Kenya’s high mobile-money use (World Bank Findex 2025: ~32% borrowing via mobile-money; ~86% using mobile money for day-to-day needs), reports of high digital-loan default rates (up to ~40%), and Safaricom’s dominant market position in telecom/fintech data.
• Analytical frame: Draws on HCI and science-and-technology studies concepts of alignment (Fujimura, Dourish) and distinguishes epistemic, modeling, and contextual uncertainty as analytic categories.

Implications for AI Economics

• Market structure & data rents:
  • Incumbent platforms with privileged data (telcos, large banks) hold competitive advantages; alternative-data production is itself an economic asset and barrier to entry.
  • Economic models should treat data creation and legal/organizational work (data deals, consent engineering) as production inputs and sources of rents.
• Endogeneity and performativity:
  • Credit scoring is performative: the use of scores changes borrower behavior and market functioning. Econometric and theoretical models must account for feedback loops and policy endogeneity when estimating welfare or default dynamics.
• Pricing, segmentation, and fairness:
  • Risk-based pricing informed by alternative data may enable finer segmentation and price discrimination, raising distributional concerns. Regulators and economists should study who benefits and who is excluded.
• Strategic behavior and model fragility:
  • Borrowers and intermediaries adopt tactics (SIM switching, obfuscation, strategic borrowing) that reduce predictive validity and can generate arms races between firms and users. This implies higher ongoing costs for model maintenance and monitoring.
• Institutional and coordination costs:
  • Firms invest beyond algorithms: legal compliance, lobbying, collection practices, product redesign, and trust-building mechanisms are part of the “cost of alignment.” These non-algorithmic investments affect market equilibrium and should be included in cost–benefit assessments.
• Policy and regulatory design:
  • Effective regulation requires attention to how alternative data are produced/processed (not only which features are used). Transparency, auditability, limits on risky operational practices (e.g., public shaming collections), and standards for acceptable alternative-data sourcing are critical.
  • Consumer protection policy should consider both model outputs and the institutional practices that make scoring workable (collection methods, pricing strategies, dispute mechanisms).
• Research agenda for economists:
  • Quantify welfare impacts of alternative-data scoring on different socioeconomic groups.
  • Model dynamic feedbacks between scoring deployment and borrower behavior (endogenous selection and strategic responses).
  • Measure transaction and governance costs of alignment (legal workarounds, data-sharing contracts, monitoring).
  • Evaluate how regulatory interventions (data protection, transparency mandates, caps on interest/collections) alter incentives for data creation and model design.

Overall, the paper highlights that evaluating AI-driven financial products requires moving beyond predictive-accuracy metrics to include the economics of data production, institutional alignment work, strategic behavior, and regulatory feedbacks that collectively determine market outcomes and social welfare.