Three months ago, I sat in a co-working space on Telok Ayer Street in Singapore, watching a product demo from a credit-scoring startup. The founder, a former Grab engineer, was walking through his company’s risk model. It was elegant. Clean UI, fast inference, and a dataset trained primarily on Southeast Asian mobile usage patterns. The algorithm weighed things like app download frequency, phone model, mobile top-up behavior, and GPS consistency to generate a creditworthiness score for the “unbanked.” Investors loved it. The pitch deck had the words “financial inclusion” on four separate slides.
I asked a simple question: what happens when this model gets licensed to a lender in West Africa?
The founder paused. He said the model was “geography-agnostic.” That it captured “universal behavioral signals.” I wrote that phrase down because I knew, with near certainty, that it was wrong. And over the following weeks, I traced exactly how wrong it was, following the path of a single algorithmic framework as it moved from a Singaporean startup’s servers to a Nigerian fintech’s loan approval pipeline, and eventually into the rejection notices received by small business owners in Lagos and gig workers in São Paulo.
What I found is a story about infrastructure, assumptions, and the quiet export of one society’s norms disguised as neutral technology. This is what digital colonialism actually looks like when you zoom in close enough to see the wiring.
The Architecture of Assumption
The credit-scoring model I tracked was built on a training dataset drawn from roughly 1.2 million users across Singapore, Malaysia, and Indonesia. These are countries with high smartphone penetration, reliable cellular infrastructure, and populations that interact with digital platforms in patterns shaped by those specific economies. In Singapore, where I live, it is normal to have a single phone with a consistent number, to top up data plans monthly, and to download apps from a curated set of mainstream platforms. The model learned these patterns and, essentially, encoded them as markers of financial reliability.
But consider Lagos. A GSMA report from 2024 found that roughly 54% of mobile subscribers in Sub-Saharan Africa use multiple SIM cards, often switching between networks to chase cheaper data rates or better coverage in different neighborhoods. Phone-sharing among family members is common. GPS data is erratic because people move between formal and informal economies that don’t map neatly onto fixed locations. Top-up behavior is lumpy and opportunistic, tied to irregular income streams rather than monthly salary cycles.
Every single behavioral signal the Singaporean model treated as a proxy for stability reads, in Lagos, as a proxy for risk. Multiple SIMs? Flagged. Inconsistent GPS? Flagged. Irregular top-ups? Flagged. The model doesn’t know why someone in Mushin switches SIM cards three times a week. It just knows that pattern doesn’t match its training data’s definition of a reliable borrower.
When I contacted the Nigerian fintech licensing this model (they asked not to be named, as they’re in the process of switching providers), their head of data science told me something revealing. “We knew the model wasn’t perfect for our market,” he said. “But the alternative was building from scratch, which takes years and millions we don’t have. So we calibrated. We adjusted thresholds. But the architecture of the model, the features it prioritizes, those stay the same.”
That word, “architecture,” stuck with me. You can adjust a thermostat, but if the building’s walls are made of glass, you’re still going to overheat. The architecture of assumption embedded in the model wasn’t a bug to be patched. It was the foundation.
How São Paulo Enters the Picture
The same model family, adapted by a different intermediary, surfaced in Brazil. A workforce management platform operating in São Paulo had integrated a behavioral risk score into its contractor vetting process for delivery and logistics companies. The score was marketed as a way to reduce fraud and identify “high-reliability” workers. In practice, it was being used as a pre-screening filter: score below a certain threshold, and your application never reached a human reviewer.
I spoke with a researcher at the University of São Paulo’s Center for Artificial Intelligence who had been studying algorithmic hiring tools in the gig economy. She described a pattern she called “infrastructure bias”: workers in favelas, where internet connectivity drops frequently and shared devices are the norm, were systematically scored lower than workers in wealthier neighborhoods. Their digital footprint looked, to the algorithm, like the footprint of unreliable people. In reality, it was the footprint of poverty itself.
“The algorithm doesn’t discriminate by race or income directly,” she told me. “It discriminates by the digital artifacts of structural inequality. The effect is the same. The deniability is better.”
That last sentence has stayed with me. The deniability is better. This is the core mechanism of what scholars like Catherine D’Ignazio and Lauren Klein have called the “data power” framework: systems that reproduce existing hierarchies while appearing technically neutral. The algorithm in question never asked about race, never asked about income, never asked about neighborhood. It simply measured patterns that are themselves products of centuries of racialized economic exclusion, and then penalized people for exhibiting them.
The Pipeline of Power
Here’s what makes this a story about colonialism rather than just algorithmic error. The direction of flow matters. The model was conceived in Singapore, trained on data from Southeast Asian economies that occupy a specific position in the global digital hierarchy, funded by venture capital from the United States and Europe, and then exported to markets in Africa and Latin America that had no role in shaping the model’s assumptions, no seat at the table when its features were selected, and often no realistic alternative.
The Nigerian fintech head of data science put it plainly: “We are consumers of these models, not producers. We integrate. We don’t architect.” This is the same dynamic that defined earlier forms of colonial extraction, where raw materials flowed in one direction and finished goods (along with the rules for how they’d be used) flowed in the other. Today, the raw material is behavioral data from billions of people in the Global South, and the finished goods are algorithmic models built in a handful of cities, mostly in Asia, North America, and Europe, that then get re-exported as infrastructure.
I’ve written before about how technology platforms create dependencies that are difficult to see until they’ve calcified. This pipeline of algorithmic influence is one of the most consequential examples. When a startup in Singapore or San Francisco builds a model, they’re encoding a theory of human behavior. When that model gets deployed in Lagos or São Paulo without fundamental rearchitecting, that theory becomes a governing system. People’s access to capital, employment, housing, and insurance gets mediated by a framework they had no part in building and no meaningful ability to contest.
The Convenient Fiction of “Geography-Agnostic”
I want to return to that phrase the startup founder used: “geography-agnostic.” Because it’s doing an enormous amount of ideological work while appearing to say nothing at all. The claim of geographic neutrality is, in practice, a claim of universality. It says: the way people in Singapore use phones is the way people everywhere use phones, or at least the way they should use phones if they want to be considered creditworthy. It takes one context’s norms and promotes them to the status of natural law.
This is a familiar move. Colonial administrations did the same thing with legal systems, educational curricula, and economic models. The British didn’t say they were imposing British norms on Nigeria; they said they were bringing “civilization,” a concept framed as universal but built entirely on British assumptions about property, governance, and personhood. The mechanism is identical. The vocabulary has been updated.
I’m not writing from a position of purity here. I live in Singapore. I benefit from the city-state’s position as a global technology hub. Brown Brothers Media operates within systems that profit from exactly the kind of cross-border digital infrastructure I’m critiquing. The co-working space where I watched that demo charges more per month than many of the Lagos small business owners I later spoke with earn in the same period. The complicity is structural, and I’m part of the structure.
But acknowledging complicity and refusing to look are different things. And what I saw, when I traced this single algorithmic decision from conception to consequence, was a system that operates with remarkable efficiency and almost no accountability.
What Accountability Could Look Like
There are people working on this. The African Union’s Continental AI Strategy, adopted in early 2024, explicitly calls for “data sovereignty” and the development of locally trained models. Brazil’s national AI legislation, still making its way through Congress, includes provisions for algorithmic impact assessments that would require companies to demonstrate that models perform equitably across demographic and geographic lines. India has been pursuing similar regulatory frameworks.
But regulation is slow, and the deployment pipeline is fast. By the time a government assembles a task force, the model has already been integrated into the infrastructure that processes millions of loan applications, job screenings, and insurance assessments. The asymmetry of speed is itself a form of power.
What I think matters more, in the near term, is a shift in how the technology industry talks about these products. The language of “inclusion” has become a marketing category. When a venture-backed startup says its credit-scoring model promotes financial inclusion, it is making a claim that deserves the same scrutiny as any other corporate claim about social impact. Who was included in the model’s design? Whose behavioral patterns does it reward? Whose does it penalize? And who profits from the deployment?
These are questions that can be asked in due diligence. They can be asked by journalists, by regulators, and by the fintech companies in Lagos and São Paulo who are choosing which models to license. The answers won’t always be damning. Some models are genuinely built with local context in mind. But the default, the path of least resistance, is to import a framework from a more powerful economy and deploy it with minor adjustments. And that default, replicated across thousands of companies and billions of decisions, is the machinery of digital colonialism.
The Human End of the Pipeline
I want to end with a specific person. Her name is Adunni (she asked me to use only her first name). She runs a small tailoring business in the Yaba district of Lagos. In early 2024, she applied for a microloan through a mobile lending app to buy a new sewing machine. She was denied. The app gave no explanation beyond a generic message about her “risk profile.”
Adunni has been running her business for seven years. She has a consistent customer base, pays her rent on time, and has successfully repaid two previous loans from a community savings group. By any reasonable, context-aware assessment, she is creditworthy. But she shares a phone with her teenage daughter. She switches between two SIM cards. Her top-up pattern is irregular because her income is seasonal, peaking before holidays when customers order new clothes.
To an algorithm trained in Singapore, Adunni looks like a risk. To anyone who has spent an afternoon in Yaba, she looks like the economy’s backbone.
The distance between those two assessments is the distance the algorithm traveled. And until the people who build these systems reckon with the assumptions embedded in that distance, the word “inclusion” will remain what it has too often been in the history of powerful institutions reaching into less powerful places: a polite word for extraction.
Feature image by Google DeepMind on Pexels
