The poor and vulnerable are the people fairness and consent matter most to, yet they have the least power to demand either. And it’s worth saying outright that being a brokeass isn’t a license for anyone to take advantage of you. Unfortunately, history is full of exactly that. Wherever there’s a gap in power or information, someone finds a way to profit from the side that can’t push back.
Sadly, in most parts of the world especially in third world countries, poverty and access to credit are like siamese twins, eternally tangled up together in ways that make life a living hell. A large share of the world’s poor stay poor partly because the one thing that could pull them out, formal credit, stays locked behind a door they can’t open.
So hand them credit, then. Problem solved? Not so fast. Most of them show up to that “door” with lean credit history or none at all, no track record for a lender to judge them by, and this gap alone can be enough to seal their fate.
On the contrary, if we take a look at more advanced nations, say the US for example, FICO built the gold standard for scoring, and while it doesn’t get things right every time, it works often enough that an entire financial system was built on top of it. What FICO does well is synthesise a huge volume of structured data into a single, defensible number, pulling from schemas refined over decades, with fields that map to risk in ways banks have trusted for generations.
Such infrastructure doesn’t exist across most of Africa, Latin America, and Southeast Asia. It also hardly exists for a new immigrant building a life in a Western country with no local credit history to their name. The data required to make a sound lending decision is either thin, scattered across systems that don’t talk to each other, or simply absent. This is the exact problem I’ve spent years thinking about at Lendsqr. That is, until AI-driven consumer credit scoring models became mainstream.
Where AI fits into the story
AI’s role here is a specific answer to a specific data problem. To see why, it helps to pull apart two words people use interchangeably: alternative data and unstructured data. They are not the same thing, and confusing them is where a lot of the AI-and-lending conversation goes wrong before it starts.
Alternative data is any data source outside the traditional credit-reporting menu. Mobile-money transaction histories, airtime top-up patterns, utility-payment records all qualify. Plenty of it still arrives in rows and columns, the same shape as a bank statement, just pulled from a source a credit bureau never touched. A model can read six months of mobile-money inflows and outflows almost the way it reads a repayment history, because the underlying structure isn’t so different.
Unstructured data sits on a separate axis entirely. SMS messages, call logs, voice recordings, photos, a free-form conversation with a loan officer or a chatbot, none of it arrives ready to plug into a scorecard. Something has to read it, listen to it, or parse it first and turn it into a signal a model can use, and that’s where AI earns its keep. A sheet of utility payments could, in principle, be scored by a decades-old statistical method. A folder of someone’s SMS inbox could not, not without a system built to make sense of language at scale.
Data can be alternative and structured at once, like the mobile-money example. It can also be alternative and unstructured at once, like the SMS example. Two different labels, with two different sets of tools required, and treating them as one blur called “AI risk” is how the conversation gets lost.
Fairness is about what the model does with what it sees
Fairness, here, has nothing to do with intent. Nobody at Lendsqr sits down to build a model that penalizes people for their race, gender, or location. Unfortunately, bias moves in more roundabout ways. Say a model discovers that people who let their prepaid airtime balance run down to almost nothing before topping up are more likely to default. Fine, the pattern holds up inside the data.
But what is it really capturing? It could be income volatility or could be a proxy for a region or income bracket where that habit is simply how people manage scarce cash. The model doesn’t know the difference, and left unchecked it will punish an entire group for a behaviour that has more to do with their circumstances than their creditworthiness.
That’s the difficult question, and it isn’t “does AI have bias.” Every model, human or machine, carries some. The difficult question is whether a model can stay genuinely useful and predictive while still producing outcomes we’re willing to call fair, and that has to be asked variable by variable.
Consent is about what a lender is entitled to learn
Where fairness asks what a model does with information someone already handed over, consent is more of what a lender is entitled to learn from that information in the first place, and AI makes the question uncomfortable in a way older systems never had to take into account.
Take for instance a borrower who agrees to let a lender read their SMS inbox to verify transaction alerts, believing the request is that narrow. If an AI model, built to extract every available signal, is run through the same inbox, income stability can be revealed, other loans the borrower is juggling, spending habits, signs of financial distress, even relationships they never meant to disclose. None of that was on the form they signed, but it sat inside data they’d already agreed to share, waiting for a system capable enough to find it.
So the question stops being “did I consent to you accessing this data.” It becomes “did I also consent to everything you’re capable of figuring out from it.” Those are two different permissions, and most consent flows, ours included, weren’t built with that distinction in mind.
Now for the side of this that gets less airtime. If fairness and consent risk make us uneasy enough to pull back hard on AI and alternative-data scoring, people with thin or non-existent credit history don’t land softly in some cleaner, fairer traditional system. I’ll be honest with you, there is no such system waiting for them.
A little history on how we got to the AI frontier
The questions around fairness, consent, and who gets left out didn’t begin with today’s credit scoring systems. Every generation has had its own version of this fight. But, to understand where we are now, it helps to look at how we got here.
In 1956, that’s when modern credit scoring began, when engineer Bill Fair and mathematician Earl Isaac built a system on the idea that subjective lending decisions needed to give way to statistical analysis. For years the formula stayed a black box the customer never saw. That began to change in 1974, when the United States adopted the Equal Credit Opportunity Act, prohibiting lenders from denying credit on the basis of race, religion, national origin, sex, marital status, or age. It put scoring systems on notice that however technical they looked on paper, someone still had to answer for the outcomes they produced.
A more consequential shift, for the purpose of this article, was happening far from Wall Street. In 2012 the Commercial Bank of Africa and Safaricom launched M-Shwari in Kenya, giving people with mobile-money accounts but no formal credit history access to loans scored on their transaction and airtime behaviour, reaching millions of first-time borrowers within its first year.
By 2014 a wave of app-based lenders, Tala (then still called Mkopo Rahisi) and Branch among them, had entered the market, pushing further into unstructured signals like phone metadata and app usage to drive underwriting. This is the tradition Lendsqr sits inside, and it’s where the fairness and consent questions in this article stopped being theoretical for a lot of people who had never had a credit history.
Regulators caught up rather quickly even if not consistently. The EU’s GDPR, in force from 2018, gave consumers new rights over automated decisions made about them and pushed the question of whether an AI model owes a person an explanation into mainstream law.
In the US, the CFPB confirmed in 2022 that federal anti-discrimination law requires lenders to explain the specific reasons behind a denial, even when relying on complex algorithms, and followed up in 2023 to close the loophole of generic, checklist-style reasons that didn’t reflect why a model made its decision. Using a more powerful model doesn’t buy a lender the right to a vaguer answer. AI is best read as the latest chapter in that history, arguably its most capable one.
So where does that leave us
This debate, unfortunately, doesn’t resolve easily, and I’ve made my peace with that. This piece has sat inside several tensions: alternative data against unstructured data, fairness against consent, exclusion against imperfection. Each one comes down to the same discipline: staying honest about what a model can see, what it can infer, and who could get hurt because of it.
The risk will never completely disappear and I don’t think pretending otherwise helps anybody.
Planes are not perfectly safe either because people still die in plane crashes. Yet nobody seriously argues that the answer is to stop flying. Instead, we continue to make planes safer, again, and again, and again. Today, based on the last five years of commercial aviation data, roughly 99.99998% of flights do not end in a fatal accident. The remaining risk is still real, and every accident still matters, but the answer has been relentless improvement rather than abandoning flight.
By the way, nobody is going to drag me on a helicopter – the risk of that is simply unacceptable to me.
I think AI-driven credit scoring deserves the same treatment.
For millions of people without a conventional credit history, there isn’t a pristine scoring system waiting in the wings if we decide AI makes us uncomfortable. In many cases, there is simply no score, no credit, and no opportunity. So the sensible response cannot be to demand that AI carry zero risk before we’re willing to use it. It should be to build guardrails that keep pushing that risk down.
Explainability is one of them.
If an AI system recommends declining someone, “the model said no” should never be an acceptable answer. A lender should be able to see something closer to: based on signals X, Y, and Z, this borrower has an estimated B% probability of default, compared with K% for comparable borrowers. The reasoning should be visible enough to interrogate, challenge, and, where appropriate, override.
And humans should remain capable of overriding it.
That matters because a score is ultimately a prediction, not a commandment. At Lendsqr, this is already part of how we think about responsible and ethical lending: technology should make the lender better informed, not remove judgment or accountability from the lender.
There should be other guardrails too. Regulators can require explainability. They can require lenders to audit models for unfair outcomes. They can draw a hard line between using data to understand repayment risk and using someone’s vulnerability simply to extract more from them. And as these systems become more capable, those rules will have to become more capable alongside them.
None of this will make AI perfectly fair. None of it will make every inference comfortable. And none of it guarantees that nobody will ever be harmed by a lending decision.
But perfection is the wrong destination to drive to. What should matter is whether we can make these systems safe, explainable, accountable, and fair enough that the enormous benefit of giving previously invisible borrowers a genuine shot at credit outweighs the risks that remain.
I believe we can.
So I will keep building but also stay willing to be told where we’re wrong, because that is how you make the next version safer than the last one.
Ó dàbọ̀.
Discover more from Adedeji Olowe
Subscribe to get the latest posts sent to your email.
