AI for Financial Inclusion: Balancing Alternative Data with Bias and Explainability
Summary
Walter Thompson and Gwendolyn Carter discuss using AI to extend credit to the 1.3 billion unbanked adults globally by leveraging alternative data like mobile phone usage and utility payments. They highlight that while machine learning can assess creditworthiness in populations like Kenyan farmers, these models risk creating proxy variables for protected classes based on location or device type. The speakers warn that the 'black box' nature of deep learning creates accountability issues under regulations like the Equal Credit Opportunity Act, which requires specific denial reasons. They argue for explainable AI and human-defined governance 'red lines' to prevent exploitation of data correlations that harm vulnerable demographics. The hosts conclude that reputational and legal risks from opaque models may outweigh marginal performance gains.
Key takeaways
- 1 Alternative data sources for credit assessment include prepaid mobile phone top-ups, utility payment patterns, and digital social network metrics, moving beyond traditional ledger entries.
Walter Thompson · 1:11
- 2 Picture a smallhold Hold farmer in rural Kenya who's never walked into a bank, a decade of mobile money use, bills paid like clockwork, an algorithm can read that as reliable, where the old system saw static. That's the dream: finding creditworthiness in everyday behavior, things that used to look like background noise.
Walter Thompson · 1:30
- 3 AI models may inadvertently create 'proxy' variables for protected classes by correlating repayment with non-financial signals like postcode, contact networks, or late-night airtime top-ups.
Walter Thompson · 2:10
Episode sections
- 0:03 – 0:51
Introduction to AI and Financial Inclusion
Walter Thompson and Gwendolyn Carter introduce the topic of using AI to address financial inclusion for the unbanked population. They highlight the limitation of traditional credit scores and the potential of alternative data sources like utility payments and mobile phone patterns to assess creditworthiness.
- 0:51 – 2:30
Alternative Data and Proxy Bias
The speakers discuss how AI models use non-traditional data points, such as prepaid top-ups and bill payment rhythms, to evaluate individuals without credit histories. They warn that these models may inadvertently use proxies for class or status, such as location or device type, leading to biased outcomes.
- 2:30 – 5:17
Explainability and Regulatory Compliance
The conversation addresses the 'black box' nature of deep learning models and the legal challenges they pose under regulations like the Equal Credit Opportunity Act, which requires specific denial reasons. The speakers argue for explainable AI to ensure accountability and allow for appeals.
- 5:17 – 8:29
Governance and Ethical Guardrails
The hosts emphasize the need for human oversight to set ethical boundaries and define fairness in algorithmic decision-making. They suggest that while AI can drive inclusion, human judgment is required to prevent exploitation of data correlations that harm specific demographics.
Claims worth verifying
The World Bank puts the unbanked at roughly 1.3 billion adults.
Walter Thompson ·0:51
How to verify: Search World Bank publications or data portals for 'unbanked adults' statistics to confirm the 1.3 billion figure and the reporting period. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
The Equal Credit Opportunity Act demands a specific denial reason.
Walter Thompson ·2:50
How to verify: Review the text of the Equal Credit Opportunity Act (ECOA) and Regulation B (12 CFR Part 1002) for requirements regarding adverse action notices and specific reasons for denial. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
A deep net might buy half a percent less default.
Walter Thompson ·4:10
How to verify: This is a hypothetical illustrative claim by the speaker; verify by searching for industry case studies or white papers on the performance delta of deep learning vs. traditional models in credit risk assessment. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
There are 1.3 billion unbanked adults.
Walter Thompson ·0:51
How to verify: Check the World Bank's Global Findex Database for the most recent data on the number of adults without access to financial services. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Do your own research
- Verify the World Bank's current estimate of the unbanked population against their latest 'Global Findex' database.
- Review the text of the US Equal Credit Opportunity Act (ECOA) and Regulation B to understand specific requirements for adverse action notices.
- Research case studies of AI-driven credit models in Kenya or similar emerging markets to assess real-world accuracy and bias outcomes.
- Check for regulatory guidance from the CFPB or OCC on 'Explainable AI' and acceptable use of alternative data in lending.
- Analyze the cost-benefit analysis of implementing XAI (Explainable AI) frameworks versus the marginal reduction in default rates claimed.
- Investigate recent legal challenges or fines against financial institutions for algorithmic bias in lending decisions.
- Look for industry standards or frameworks for 'fairness metrics' in algorithmic decision-making, such as those proposed by NIST or IEEE.
- Review the specific alternative data providers mentioned (e.g., mobile money operators) for their data privacy policies and consent mechanisms.
Full transcriptexpand
0:03Walter Thompson: Artificial Intelligence Empowering Financial Powerhouses. Welcome to a fascinating intersection of artificial intelligence and finance. Quick note up front: I'm Walter Thompson, here with Gwendolyn Carter. We keep boiling our financial lives. down to a couple numbers: a credit score, a balance, a paycheck, when the fuller story might be hidden in things like utility payments or even mobile phone patterns.
0:38Gwendolyn Carter: You jumped right to the pot that hooked me years ago. For billions, the usual story just isn't there. No history, no file, no nothing. So the banking system can't see them. AI says it can.
0:51Walter Thompson: And maybe serve them, not just notice them. The World Bank puts the unbanked at roughly 1.3 billion adults. That's more people than any country has residents. If we're serious about changing that, we have to talk about the data these systems feed on, the so-called alternative stuff. Right, the how. Building a picture with no traditional trail isn't magic. it's different inputs, not the usual ledger entries. Exactly. Forget FICO scores. Think prepaid top-ups on your phone, the rhythm of your electric bill payments, even the texture of your digital social ties. Some models peak there. And that can be stunningly helpful. Picture a smallhold Hold farmer in rural Kenya who's never walked into a bank, a decade of mobile money use, bills paid like clockwork, an algorithm can read that as reliable, where the old system saw static. That's the dream: finding creditworthiness in everyday behavior, things that used to look like background noise. But here's the thing I can't shake. What else does the model latch onto? Maybe it notices most of her calls cluster in an area with high default rates, or it tags the fact she uses an older, cheaper handset, and that becomes a stand-in for income. Now we're in proxy territory. Harmless looking inputs morph into coded signals for class or
2:10status. Exactly. The system isn't trying to be biased, it's hunting correlations with repayment. If postcode, contact network, or late night airtime top ups correlate with default, the model will use them. It doesn't know society, it knows patterns. And that's chilling. You're evaluated not by your choices, but by a statistical silhouette stitched from a thousand data crumbs you didn't realize mattered. That's how you recreate redlining with new plumbing. It's not the train tracks anymore, it's metadata. And the worst part? The sealed box. Decisions you can't pry open. Right, with deep nets, even builders can't fully say why this person got to a know. no. The loan officer just has. computer says no. That isn't inclusion, that's a digital gate, which raises a basic accountability crisis. If no one can explain it, there's no appeal, only an opaque verdict. And a legal mess, how does that square with the Equal Credit Opportunity Act's demand for a specific denial reason? Because the computer felt like it doesn't fly. That's the case for explainable AI, not just accuracy, but the ability to show. your work, or at least the model's logic. In practice, what are we talking about? A breakdown like denied, 30% from phone age, 20% postcode, 50% late-night top-ups? Strip down, yes. Highlight which features carried the
3:30most weight so a human can trace the decision path. But then what? Does explainability cure the sickness or just label it neatly? We might just be documenting discrimination with better charts. Fair push. I still think transparency is step one. If you can't see bias, you can't correct it. If a model leans hard on location, regulators can say, nope, pull that lever out. So an audit crowbar. Exactly. Turning pitch black into grey. gray. Imperfect, but at least you get a handle. My worry is incentives. The most predictive setups are often the most opaque. If a deep net buys you, say, half a percent less default, that's real money at scale. True. Interpretability can cost performance, but firms are learning the long run risks of chasing the last basis point. You mean the get sued into the ground risk? That and reputational pain when it turns out your system quietly starves single parents or specific neighborhoods of credit, that stain can cost more than the 0.5% you gained.
4:36Gwendolyn Carter: Point taken.
4:37Walter Thompson: Now, if the decision is explainable, the person at least gets something to fix. Say the reason is two late utility payments in five years. That's actionable. Right. You can aim at a target. Stay current for a year, try again. It turns a void into a to-do list. But what if the reason is mushy? Low social network stability score? What are they supposed to do? Ask friends not to move? That's how you build a digital caste system. Rules you can't meaningfully act on. That's the nightmare. Explainable yet still indefensible. Seeing the logic doesn't make it legitimate. Which brings us straight to governance. Call it oversight, call it values, same destination. Human judgment setting the guardrails. Yes, someone has to declare certain signals off limits. If we don't, the machine will happily discover and exploit them. We draw the red lines. In other words, the top top-level level algorithm is still human, which data. are allowed, what fairness means, and how process beats raw lift. I'm still torn. The upside, bringing a billion people into formal finance, is enormous. That's generational stuff. And we can't lose sight of the actual wins. Families getting working capital loans that
5:57change trajectories. Totally. But the good doesn't cancel the harm by default. I don't know. trust profit motives to prioritize fairness without pressure. We need muscular, tech literate rules, not afterthoughts, but baked in. I land a notch more bullish on sunlight. Regulation is necessary, absolutely, and so are audits. But public scrutiny and clear explanations can move mountains when fines don't. So you're betting on markets. with a bright flashlight, I'm betting on regulators with a wrench. Both tools, same job. Build something that's genuinely inclusive, not just more efficiently exclusive. And we should admit, someone listening to our earlier drafts knows parts of this better than we do. If they audited my notes, they'd find turtle problem scribbled in the margin, and I still don't know what I meant by that. The coffee helps. Here's my petty gripe while we're confessing. Model validation memos that switch between month first and day first dates in the same table, in a risk document. I stand by my pettiness. That's not petty. That's survival. Clarity is a control.
7:12Then maybe that's the headline. Clarity as a control. Explain enough to be challenged, restrict the features that cross ethical lines, accept a small accuracy haircut, and keep a human in the loop who can say, This outcome doesn't pass the smell test. And when the model points at a fragile proxy, we name it out loud. That simple move, calling the thing what it is, can change the whole conversation. So we can't keep interrogating the system and ourselves. Not a tech problem at bottom, a human one.
7:45Unidentified speaker (UNKNOWN): Hmm.
7:45Walter Thompson: Agreed. Let's pick this up next time with concrete playbooks, monitoring, recourse workflows, and what an actually actionable reason code looks like in the wild.
7:57Unidentified speaker (UNKNOWN): Hmm.
7:58Walter Thompson: Good. Because if we don't set those lines, the model will draw. them for us, and we might not like where it puts them.
8:06Unidentified speaker (speaker_2): This show was assembled using AI tools.