Last year, scammers launched a deepfake videos of popular Nigerian social media influencer, Martins Vincent Otse, through a fraudulent Facebook campaign. The cybercriminals superimposed a fabricated AI-generated audio on a genuine video of Otse from early 2024.

Martins Vincent Otse

In the cloned voice, Otse introduced an online chatting scheme where viewers could earn up to $1,000 monthly. By cloning the influencer’s face and reputation, the scammers not only harmed Otse’s reputation as a respected influencer, they also preyed on the poverty of many Nigerians, tricking them into becoming a member of an online platform created to appropriate their personal information and savings.

On October 5, to show that the issue still mattered, the Nigeria Information Technology Development Agency (NITDA) warned organisations and staff against sharing sensitive data with public AI tools.

As the federal government shapes  its National Artificial Intelligence Strategy with the aim to leapfrog development, doctors and nurses use clinical Al tools without standardised verification methods and companies run systems without Al oversight structures. Banks move transactions to automated and Al-driven solutions to combat soaring transaction fraud and manage financial risk, while

87.5 percent of  fintech companies now deploy Al for automated fraud detection. In theory, the Nigeria Data Protection Act gave citizens the right to object to automated decision-making; in practice, the abuse of enforcement renders it non-justiceable when Al systems cause harm.

For instance, during the lead up to the 2023 presidential elections, a simulated audio clip went viral, purportedly featuring former Vice President Abubakar Atiku and other members of his party  discussing plans to rig the election. Fact checkers labelled the audio false and generated by artificial intelligence tools. Investigations opened, but there was no prosecution. Defamation laws exist in Nigeria, but none existed to assign blame for the creation of a lie by generative models.

In another instance, unverified screening AIs missed the detection of over 1,247 tuberculosis cases, as well misclassified 342 high-risk pregnancies, with 186 preventable patient deaths linked to Al diagnostic errors. The tools involved were not registered as medical devices by the National Agency for Food and Drug Administration and Control ( NAFDAC), but registered as “wellness” apps.

Driven by the Ministry of Communications and NITDA, the National AI Strategy sets normative ambitions — to align AI deployment with human rights. But with internet penetration now at 68.3 percent penetration at 148.2 million subscribers, that ambition is no longer aspirational; the state has a functional duty to protect citizens from algorithmic harm at scale.

The issue remains implementation. NITDA has put together a draft Al Code of Practice, and the Nigerian Data Protection Commission is enforcing data protection. But broader analytics from NITDA suggest that more than 56 percent of government IT projects fail to meet expectations due to procurement, design and implementation flaws. This exposes sensitive consumer data to hackers, increasing the risk of identity theft and digital fraud from design flaws.

In the UK and EU, high-risk AI systems are subject to mandatory pre-deployment conformity assessments. In Nigeria, where banks now use AI-powered tools to identify and prevent fraudulent transactions, no equivalent requirement exists for auditing models against local fraud patterns before they go live.

This institutionalises systemic weaknesses. Angela Olusola was a 46-year-old mother of three. After an unaudited AI flagged her successive transaction as suspicious fraud behaviour, it froze her account. Unfortunately, she lacked financial buffers, with the 48-to-72-hour account freeze meaning she could not buy food, pay her transport fares or restock her business. Since she lived from hand-to-mouth, the algorithm error triggered an immediate poverty crisis.

Algorithmic biases in micro-lending are not abstract. A poor trader who uses a cheap phone and constantly changes SIM cards is flagged as “high risk”. The algorithm sees her behaviour as an indicator of low financial standing or attempts to cheat, resulting in instant loan rejections.

The training data is heavily skewed to salaried, urban, Android users in Lagos.

Because the algorithms are fed historical datasets consisting mostly of formal salary earners of people belonging to this class, the AI learns to flag the variable income streams of the informal majority as a “high risk of default,” effectively shutting them out. If any person’s dataset deviates from those of salaried, urban, Android users in Lagos, the AI treats  this as high-risk indicators or fraudulent tendencies, leading to exclusion.

With time, automated errors force the poor into the cash-only economy, their financial security stripped from them, leaving them at the mercy of inflation and crooks.

A second, quieter weakness may be happening at the other end of the value chain. Google and Microsoft have started collecting Nigerian data to improve their Large Language models (LLMs). Projects like Meta and the Gates Foundation’s African Next Voices use local hubs to clean up voice data. The average Nigerian spends a lot of their monthly income on these bundles, paying for the network connection required to give their data away for free.

Who owns a Nigerian’s voice when it is used to train voice cloning? The local hubs sign release forms, meaning the ordinary Nigerian surrenders their voice ownership rights, their voice becoming an asset of the funding enterprise. Nigeria doesn’t benefit from this, since it has no data trust, no licensing regime for foundation models trained on its digital identity. We give them the raw materials, we shoulder all of the risk, and we spend a lot of our money purchasing the product.

Another weakness is state power. The federal government of Nigeria has spent over $470 million to purchase Al-driven cameras with automated license plate and facial recognition. Many note that these systems operate in a legal vacuum. In the absence of Al-specific laws, this brings a situation where the ordinary citizen’s everyday life could be constantly logged, analysed and stored.

Nigeria does not need to ban AI to protect her citizens. It needs to carry out three basic things that other African nations are already considering.

Addressing dataset bias is now a technical necessity, if it wasn’t in the past. Since current models are trained predominantly on salaried, urban, Android users in Lagos, they systematically misclassify the 76.7 percent of Nigerians in the informal sector. Funding localised datasets would correct this distributional skew while lowering adaptation costs for indigenous tools.

A constitutional remedy is required on the issue. Section 37 of the 1999 Constitution, which guarantees privacy, must be amended to expressly cover biometric and digital video metadata. Without that, state procurement of surveillance AI will continue to operate in a regulatory vacuum.

Mandating consumer-protection-by-design in fintech is now imperative. A significant share of automated loan rejections among low-income borrowers stems not from actual default risk but from unadjusted device-proxy metrics — such as SIM swapping to secure network coverage — being misread as fraud signals.

Without these interventions, Nigeria’s AI ambitions risk not halting underdevelopment but automating it. Martins Vincent Otse’s video was flagged as fake—saving many people from taking part in the online chatting scheme. For many  without that information, losing money becomes final.

The central question is not whether Nigeria should pursue AI development, but how liability is distributed when these systems fail. Currently, that burden is borne disproportionately by ordinary citizens.