In an exclusive interview with «l’express», Dan Banik, Professor of Political Science at the University of Oslo, discusses artificial intelligence, its concrete uses in the Global South, the risks of widening inequalities between nations, and what Mauritius should prioritize to govern AI responsibly.

? Beyond the hype, what are the concrete use cases where generative AI is genuinely improving lives in the Global South?

The hype cycle around generative artificial intelligence (AI) tends to obscure the fact that genuinely useful (and often quite mundane) applications are already changing lives in lowand middle-income contexts. The clearest examples sit at the intersection of three constraints that the Global South knows well: scarce specialists, weak infrastructure, and fragmented information. In healthcare, computer-aided detection tools such as CAD4TB now triage tuberculosis from a chest X-ray in roughly five seconds and can be used in remote settings without internet access or trained radiologists. Ruby Health uses a smartphone photo of a fingernail to screen for anemia non-invasively, lowering the cost of a first-line check enormously. In agriculture, Farmerline’s Darli, recognized on TIME’s Best Inventions list in 2024, lets smallholder farmers in Ghana access regenerative-farming advice, weather updates and cropdisease detection in their own language through a simple voice or WhatsApp interaction. In education, QANDA, used by around eight million students each month across Asia, gives a personalized explanation from a photo of a math problem in seconds, and SolarSPELL provides solar-powered, offline digital libraries to communities without reliable connectivity. Tools like Be My AI and Voiceitt2 are quietly expanding agency for blind users and people with non-standard speech. What unites these cases is not novelty but fit: they meet people where they actually are, i.e., on a basic phone, in a clinic without specialists, in a classroom without enough teachers. That is where generative AI, in my view, is doing real work.

? You advocate for a “rights-based approach” to AI deployment. What does this look like in practice in a developing country?

A rights-based approach means treating AI not as a neutral technical tool but as a socio-political phenomenon that touches on dignity, equality, due process and freedom of expression. In practice, in a developing country, it has four concrete components. First, transparency in two senses: algorithmic transparency, so that people can understand how a system arrived at a decision affecting them, and institutional transparency, so that citizens know what data was collected, who is using it, and on what legal basis. Second, contestability, i.e., a citizen denied a benefit, flagged by predictive policing, or excluded by a biometric system must have a meaningful and accessible route to challenge that decision and receive a remedy. Third, inclusion in design, i.e., the people most affected by a system, including marginalized linguistic, ethnic and economic groups, should have a voice in how it is built and evaluated, not only in how grievances are processed afterwards. Fourth, baseline protections against the most dangerous uses (e.g., realtime mass facial recognition, social-scoring schemes, electoral microtargeting) either through national legislation or through regional frameworks. None of this requires a developing country to copy the EU AI Act line by line. But the underlying logic – that some uses are off-limits, others require independent oversight, and citizens always retain rights of explanation and redress – is portable, and it is the right starting point.

? Is AI likely to widen the gap between rich and poor countries, or can it be a tool for catching up?

Honestly, both are happening at once, and which one wins depends on choices we are making right now. The widening pressures are real and well documented. For example, the UNDP’s 2025 Human Development Report shows that global progress on human development has slowed and that the long-running narrowing of the gap between very high and low HDI countries has reversed. Roughly a third of the world’s population is still offline; in low-income countries, fewer than 30 percent of people use the internet, against more than 93 percent in high-income countries. Frontier AI models, the data centers that train them, and the chips that run them are concentrated in a handful of firms in the Global North. If a country has neither the connectivity, the data, the compute, nor the regulatory leverage, it risks paying for AI services without shaping them.

At the same time, the catching-up case is genuine. Mobile telephony showed that leapfrogging is possible. Locally developed tools (e.g., Farmerline in Ghana, QANDA in Asia, mobile credit scoring across East Africa) demonstrate that you do not need to build your own foundation model to deliver real value. You need to build the application layer that fits your context, your languages, and your sectors. AI will widen the gap by default and narrow it by design. The deciding variables are public investment in skills and connectivity, smart procurement, regional cooperation, and a willingness to set rules rather than only follow them.

? How can small island states like Mauritius position themselves against tech giants who control AI?

Small states will not out-compute Google, Microsoft or OpenAI, and they should not even try. The strategy has to be asymmetric. First, Mauritius can leverage the application layer rather than the model layer. The most valuable AI work for Mauritius is not training a frontier model. Rather, it is finetuning, integrating and contextualizing existing models for problems Mauritius understands best, including its bilingual public administration, its tourism and financial services sectors, its ocean economy, climate adaptation for a small island, and multilingual education in Creole, French and English. That is where competitive advantage and local sovereignty actually lie.

Second, use procurement as policy. Government and parastatals are major buyers of digital services. Contractual requirements on data localization, explainability, audit rights, and exit clauses give a small state surprising leverage over very large vendors–provided those clauses are written, defended and enforced. Third, pool sovereignty. Mauritius alone is small. But Mauritius within the African Union’s emerging AI framework, the Indian Ocean Commission, the Commonwealth, and bilateral partnerships with the EU is much less so. Common standards, shared regulatory expertise and joint infrastructure investments can typically multiply leverage. Fourth, invest in human capital relentlessly. And not only in AI engineers, but also in lawyers, regulators, journalists, auditors and civil servants who understand AI systems well enough to govern them. The countries that will be least dependent are the ones whose institutions can ask hard questions and read the answers.

? In your research, have you observed authoritarian abuses linked to AI deployment in certain countries?

Yes, but this is not research I have undertaken myself; however, the pattern in the available literature on the topic appears clear enough that we should treat it as a category of risk rather than a list of incidents. There are three recurrent patterns. The first is the normalization of mass surveillance– such as facial recognition in public space, predictive policing and social-scoring schemes–that, taken together, can suppress dissent and shrink civic space long before they are used to make any specific arrest. The second is electoral manipulation. We have seen deepfake video and audio used to impersonate political leaders. Research also finds that large language models combined with personality inference can generate persuasive and individually-tailored political messages at scale, with no human in the loop. The third pattern is more subtle: the use of AI-enabled biometric voting and identification systems where transparency, error rates and data governance are weak. In some countries, biometric voter authentication has produced documented false-rejection rates and inconsistencies in data handling that have contributed to public mistrust and, in some cases, to the contestation of electoral outcomes. The point is not that these technologies are inherently authoritarian. Rather, it is that without independent oversight they can be turned against the citizens they were meant to serve.

? Data is AI’s fuel. How do we protect citizens in developing countries from data exploitation?

Treat data as a public-interest matter, not as a side-effect of using a service. A clear and enforceable data protection law with an adequately resourced regulator is a sensible starting position. However, enforcement capacity is what gives a law teeth. People should know what is being collected, by whom, for what purpose, and for how long, in language they actually understand. There should be restrictions on the most invasive practices, including limits on real-time mass facial recognition and on the use of personal data for political microtargeting. And data localization and audit rights in public-sector contracts are needed so that sensitive citizen data (e.g., health, biometric, financial) cannot simply be exported with no recourse. Without these protections, the political economy of AI tilts further toward a small number of foreign vendors who already control the models, the infrastructure and increasingly the data.

? Do we need international AI regulation, or should each country set its own rules?

We need both, and they have to do different jobs. National rules are essential because AI is deployed in specific institutional contexts (e.g., elections, courts, schools, hospitals, welfare systems) whose design varies enormously across countries. A one-size-fits-all global regime would either be too lax to bite or too rigid to fit. National regulation is also where democratic legitimacy lives. The rules that affect citizens should be made by people accountable to those citizens. But AI itself is borderless. Models are trained in one jurisdiction, hosted in a second, deployed in a third, and produce harms in a fourth. Without some international coordination on the most serious risks (including autonomous weapons, large-scale electoral interference, cross-border surveillance, child safety, and environmental costs of large-scale training), national regulators will simply be outflanked. The current landscape – with the EU AI Act on one side, a more fragmented US approach on the other, and large parts of the world without any framework at all – is creating regulatory gaps that are easy to exploit. Crucially, the global layer cannot be designed by Global North countries alone. If it is, it will lack legitimacy and will not survive.

? Can AI genuinely strengthen participatory democracy, or is that an empty promise?

It can, but only if we are honest about what “participation” actually means. AI tools can help governments process and synthesize large volumes of citizen input from participatory budgeting, citizens’ assemblies and public consultations. Chatbots and interactive platforms can explain a draft law, a budget allocation or an entitlement in accessible language (and in the local language) to people who would otherwise be locked out of the conversation. Moreover, independent fact-checking and deepfake detection tools can support the informational environment on which democratic deliberation depends. However, although AI can scale listening, it cannot replace the political work of deliberation, contestation and accountability. Used well, it is a cognitive scaffold that helps citizens participate more effectively. Used badly, it is a sophisticated form of theatre. The difference is institutional: who controls the system, who audits it, and whether the resulting input has any binding effect on what governments actually do.

? Which sectors should be priorities for AI adoption in a country like Mauritius: education, health, agriculture?

All three are genuinely promising, and I would resist the framing that forces a choice. Each sector addresses a different bottleneck. Education is probably where the per-rupee impact is highest in the short term. Adaptive learning tools can personalize instruction for students at very different levels in the same classroom. Translation and tutoring assistants can extend the reach of strong teachers. And the cultural fit for Mauritius (with its multilingual population) is unusually good. Health is where AI can most directly extend access: low-cost diagnostic tools, AI-supported telemedicine for the outer islands, predictive analytics for non-communicable diseases that are rising sharply in Mauritius, and decision-support for primary-care nurses. Agriculture matters disproportionately for climate resilience. Precision tools, weather-driven advisories, and pest and disease detection can help farmers adapt to a climate that is changing faster than traditional knowledge can keep up with.

I would also like to highlight public administration, where AI can streamline routine bureaucratic tasks and free up frontline staff, as Estonia has shown. And also, the ocean economy, where Mauritius has a natural specialization that almost no other country can claim at the same scale. The question is less “which sector” and more “what is the highest-impact, lowestrisk application within each sector,” and then fund and regulate those applications properly.

? How do we prevent AI from reproducing or amplifying existing biases in our societies?

That is a good question. Perhaps we should start by recognizing that this is not primarily a technical problem. AI systems reproduce bias because they are trained on data drawn from societies that are themselves unequal, and because their development teams and benchmarks are still overwhelmingly Euro-American. Most AI ethics literature reflects that imbalance, which means African, Indian Ocean and other lived experiences are routinely sidelined.

Representation in the design and evaluation of systems (linguistic, ethnic, gender, disability, geographic) is important. A facial recognition system that has never been benchmarked on dark skin tones will fail dark-skinned users. And a translation tool that ignores Mauritian Creole will not serve Mauritian citizens. In addition, independent auditing (mandatory, periodic, and with real consequences) of high-stakes systems used in policing, welfare, education and employment is also crucial. Audits should look not only at average accuracy but at performance across subgroups, because that is where bias hides. Finally, civic and digital literacy must be emphasized and prioritized so that the people interacting with these systems (including citizens, journalists, civil servants, judges) can recognize a biased outcome when they see one and know what to do about it.

? Do Global South countries have the technical and financial capacity to develop their own AI, or will they remain dependent?

There are several issues here relating to fine-tuning opensource models on local languages and local data, building applications that serve specific sectors and populations, producing high-quality local datasets that did not previously exist, auditing and stress-testing imported systems before they are deployed in sensitive areas, and training engineers, data scientists, regulators and lawyers. Several Global South initiatives, such as QANDA in Asia, Farmerline’s Darli in Ghana, and the Kwame learning assistant on the African continent, show that locally led AI is not aspirational; it is happening now. A country can be dependent on external sources for compute and base models while being highly autonomous in deployment, regulation and data governance. That mixed position is realistic and, frankly, where almost everyone outside the United States and China currently sits, including most of Europe. The risk is not that the Global South will be “dependent” in some abstract sense; it is that countries will accept terms of dependency that they could have negotiated more firmly. Pooled regional capacity, open-source ecosystems, and sustained investment in human capital are the way out.

? What advice would you give Mauritian policymakers who want to adopt AI responsibly?

Lead with the public interest, not with the technology. Ask which problems Mauritius most needs to solve (e.g., climate adaptation, an ageing population, public-sector productivity, education quality) and ask where AI is genuinely the best tool. Some problems do not need AI; some do; pretending otherwise wastes money and erodes trust. But also, build the institutional plumbing first: a modern data protection regime with a properly resourced regulator, clear procurement standards, an independent body to audit high-risk publicsector AI, and transparent registers of where AI is being used in government. Without these, even good intentions produce bad outcomes. Invest in digital and civic literacy across the population.

The countries that govern AI well are the ones whose institutions can ask informed questions. Cooperate regionally and internationally. Mauritius is small, but its diplomatic and institutional reach is significant. The most important AI rules of the next decade will be written in regional and international forums; being in the room, with prepared positions, matters far more than being first to deploy any particular technology.

? What advice would you give Mauritian businesses who want to deploy AI in their daily processes?

Pilot small and evaluate honestly. Take data seriously before taking on models. Most disappointment with AI in business comes from poor data, not from poor models. But also, be transparent with customers and employees. If AI is making or shaping decisions that affect a customer – including pricing, eligibility, and recommendations – say so. If AI is being used to monitor or evaluate employees, be explicit about it and create channels for them to push back. Trust, once lost on these issues, is expensive to rebuild. The firms that get the most out of AI are typically the ones that use it to make existing staff more effective, not the ones that use it to fire and rehire.