In this episode, Dr. Fola Adeleke, Executive Director of the Global Center on AI Governance, talks us through what a realistic path to AI adoption in Africa might look like.

Below is a transcript of the conversation.

Meenakshi Dalal:

Hello and welcome to the AI Elsewhere podcast. My name is Meenakshi Dalal and in this first episode, we’re going to ask a very big question about a very big part of the world. What does a realistic path to AI adoption in Africa actually look like?

What got me thinking about this was this steady drip of headlines. Projections that AI will add $1.2 trillion to the African economy by 2030. New national AI adoption strategy after new national AI adoption strategy. A continental initiative that’s taking an Africa-first approach to AI. Billion-dollar data center investments.

At the same time, many places across the continent are still managing more basic issues like clean water and reliable electricity. To me, that contrast begs the question, what’s realistic?

With the entire world racing to adopt AI as quickly as possible and African countries being no exception, what approach is actually going to move countries forward given some of their constraints?

To find out, I sat down with Dr. Fola Adeleke. He directs the Global Center on AI Governance and its African Observatory on Responsible AI.

Here’s our conversation.

Meenakshi Dalal:

Why don’t we start with a little bit of an overview for people who may not be familiar with this part of the world. Which African countries would you say stand out in AI conversations right now?

Four countries central to AI conversations

Fola Adeleke:

We’ve primarily zoned in on Nigeria, South Africa, Kenya, and to some extent Rwanda. Rwanda comes in because the President of Rwanda currently leads the Smart Africa Initiative. Nigeria, Kenya, and South Africa have been the countries of priority for most big tech institutions, largely for economic reasons and largely for geopolitical reasons.

South Africa has the most robust infrastructure on the continent in relation to AI. So the availability of infrastructure is one thing that drives that interest. Microsoft recently pledged AI infrastructure in South Africa. Google also opened its own cloud services hub within South Africa. Cassava Technologies intends to invest just under a billion dollars in partnership with NVIDIA on building the first Africa AI factory in South Africa. And the idea is that South Africa will serve as this infrastructure hub that services the rest of the continent.

Kenya’s own competitive advantage is their tech startup ecosystem. A lot of innovation in terms of AI has been led by Kenya. For a very long time Kenya had this nickname of being Africa’s own version of Silicon Valley.

And then Nigeria. Nigeria’s advantage is its youth bulge. It is the largest country on the continent so in terms of sheer numbers, your biggest market is always going to be Nigeria.

Meenakshi Dalal:

So broadly speaking, big companies are looking at South Africa for infrastructure, Kenya for its tech startups and Nigeria for its large consumer base. And then Rwanda comes in for policy leadership.

Let’s now zoom out to the rest of the continent. Can you give us an idea about what African policymakers are most concerned about right now?

Digital sovereignty and geopolitical pressure

Fola Adeleke:

Policymakers are preoccupied with this idea of digital sovereignty. But achieving this will be difficult given most countries heavily rely on external actors, including technology that has been developed by big tech firms elsewhere.

These countries, need to deal with how to stay away from being locked into geopolitical competition, including this broader US-China AI race. Some of these countries have Chinese-built infrastructure which US companies don’t want to use and so you risk this fragmentation, geopolitical fragmentation.

But their neutrality is not risk-free.

So diversification of partners is incredibly important for these countries if we are going to really fulfill this objective of digital sovereignty, which is not something that has been led solely by these four countries. It’s also been pushed by the African Union.

Meenakshi Dalal:

If I’ve understood correctly, on the one hand, African countries want to adopt AI technology at scale. On the other, depending on how they do it and whose models and which infrastructure they use, they can face consequences because of longstanding US-China tensions. And all of that makes digital sovereignty a very serious priority.

But how does that play out in practice? I mean, is it actually happening?

Fola Adeleke:

African ministers are saying the right thing around the need to centralize infrastructure because African economies on their own are too small to attract foreign investments. If you are truly going to fulfill this objective of digital sovereignty, you are better off speaking as one because banding together potentially gives you more leverage.

Conversations around cross-border data flows, making sure that your regulations are essentially in sync is something that African regulators are also increasingly speaking about. But I think it’s a different thing to actually claim that they are actively working towards this.

The African Union in its AI Strategy laid the groundwork for us to achieve some of these things. But we are yet to see the implementation plan so we can’t really speak to how African states are implementing this commitment for now. But I can certainly speak to how the private sector is trying to assist them.

Meenakshi Dalal:

How is the private sector trying to assist them?

Infrastructure gaps and uneven access

Fola Adeleke:

For example, there’s a growing subsea capacity on the continent, large subsea cable systems such as 2Africa, continent spanning projects that are meant to connect the entire continent. But in reality, the rural urban divide in most African countries is still quite significant.

African states are certainly not equal in their levels of infrastructure development or access to finance for building infrastructure. So to maximize the benefits of projects like 2Africa, policy interventions on digital inclusion are still needed to reduce affordability barriers.

Meenakshi Dalal:

So even with better connectivity, the rural-urban divide means the benefits won’t be felt equally. What has to change?

Fola Adeleke:

Some of these initiatives assume widespread internet penetration, reliable electricity, and affordable mobile broadband. But the major barrier, is the affordability of mobile devices. In some low-income countries, the retail cost of a smartphone can be as high as 70% of their average monthly income, compared to a global average of 25-26%.

Meenakshi Dalal:

Right, no matter how connected somewhere is, if people can’t afford a smartphone, that connectivity doesn’t mean very much. So are you saying that governments are getting ahead of themselves by focusing on AI investments?

Are governments getting ahead of themselves?

Fola Adeleke:

It makes sense that African countries are increasingly amending their digital transformation strategies to essentially prioritize AI investments because you are trying to chase the money, right?

And even though a lot of African countries still have some serious basic infrastructure problems you still want some kind of signaling to investors that you are aware of these new developments and you are open for potential investors.

You could also argue that some of these national AI strategies and ambitions of states is essentially putting the cart before the horse

Meenakshi Dalal:

Okay, so investments in AI aren’t necessarily premature, but it’s a matter of which ones countries are prioritizing. How would you suggest they move forward?

How local startups can drive AI adoption

Fola Adeleke:

Don’t get too excited about AI just yet. You still need to prioritize some of your basic infrastructure problems. The fact that you shouldn’t get too excited doesn’t mean there isn’t anything you shouldn’t be doing now. Rather focus on developing an enabling environment for your thriving and innovative tech startups.

So while I’m skeptical about having a national AI strategy where every single African country is saying that they want to be the leader, the leader of AI in Africa… focus on more concrete objectives.

Yes, infrastructure investment is very expensive. It’s going to take you several years to be able to build up your national grid capacity. But a simple thing like having a Startup Act and making sure that your tech startups can thrive by giving them access to financing, giving them tax relief, making sure that they’re protected in terms of foreign competition, giving them additional rights in terms of intellectual property. It helps in the meantime. It helps in the meantime. And then they start talking about bigger issues around data centers and things like that.

Meenakshi Dalal:

I don’t fully understand that part. How does focusing on tech startups push African countries ahead when it comes to AI?

Fola Adeleke:

Because if you’ve developed an ecosystem where your tech startups are thriving and your tech startups are attracting significant foreign investments, then that builds the economy as a whole.

You won’t find yourself in situations where your startup founders are selling up and leaving the continent. So you are creating new businesses, you are hiring locally and you are still bringing in significant foreign investments because the reality is when you are developing local solutions for local problems you can’t sell off.

So even if a foreign investor is keen, that technology has been developed for a specific local context…so they will just be investing in you to grow the business, right? And you’ve seen that happen successfully in Nigeria, for example, in Kenya as well, M-Pesa, that mobile money program, those are local solutions for local problems where there’s no risk of a company acquiring that and then all of this is transferred elsewhere.

African leaders, ministers keep saying this, “local AI, local AI, local AI.” If your local startup system is thriving, there will be increased appetite for more people to come into this system to build more innovative products. Local AI is truly supporting your startups.

Meenakshi Dalal:

You mentioned Startup Acts are an important policy solution. I see how that works for a country like Kenya, which is known for its tech startups. How does that translate to other African countries?

Fola Adeleke:

Across the continent there so many tech startups. And the African Union developed a Startup Model Law Framework. Not every country has a startup law. Nigeria has one. Tunisia was the first to have one. Other countries like Cote d’Ivoire Togo, Senegal, they now have Startup Acts. And the idea of a Startup Act really is to support that young ecosystem.

Meenakshi Dalal:

So the argument here is that the real path to adopting AI is to create the right conditions for tech startups to work on local problems. And more often than not, AI technology will be a part of those solutions. But of course, all that depends on money.

Funding and real-world AI innovations

Fola Adeleke:

I think one thing that we also need to recognize is that the funding landscape for AI is a little bit more diverse than what you would find in other industries. So if we are a tech startup where we are trying to develop a new local AI innovation that is going to improve access to healthcare or improve in education. Suddenly your funders are not just venture capitalists or private funders. You can potentially get financing for that from a foreign aid agency or from a foundation because of the social impact of innovation.

And the innovation is amazing. We have someone who developed voice to text chat bots to help illiterate, small businesses to develop finance sheets to access short-term loans in banks. I’ve seen innovation around using AI to protect women against gender-based violence. I’ve seen people stay on HIV treatment using AI technology. So the innovation is really exciting out there that has been developed by Africa-based startups.

Meenakshi Dalal:

I think that’s a wonderful note to leave it on.

We started this episode trying to understand what a realistic path for AI adoption in Africa might look like. And what I’m taking away from this conversation is that big ticket AI infrastructure projects may make more sense later. And in the meantime, creating the right conditions for local tech startups to thrive, especially ones using AI, may be the most beneficial path forward.

Thank you so much, Fola, for your time and perspective. And thank you for listening to the first episode of the AI Elsewhere podcast. More soon.

—

Special thanks to Hasham Cheema for the original music and for mixing this episode.

Saillant is an AI researcher who advises governments and international institutions on AI policy, including work with the United Nations Industrial Development Organization, the OECD, and the United Kingdom. Originally from the Dominican Republic, his work focuses on how AI can advance economies across Latin America and other developing regions.

In our conversation, Saillant expands on how AI-ready special economic zones, or AI-SEZs, would work and why they offer a way for developing economies to establish effective AI regulation through incentives, rather than deregulation.

This interview has been edited for clarity and length.
Interviewer: Meenakshi Dalal (MD)
Interviewee: Said Saillant (SS)

The Problem

MD: What problem are you trying to solve?

SS:

• Right now, AI development is concentrated primarily in the U.S., Europe, and China. This makes most countries consumers of AI with limited pathways to become builders or regulators.

• That concentration is a problem because it means that only a small set of countries capture the compounding benefits of creating AI systems, like money, talent, and agenda-setting power. The goal is to identify a repeatable way for more countries to build expertise, test AI systems in the real world, and accumulate governance know-how.

MD: I understand why the current geographical concentration is an issue from a benefit-sharing point of view. But if AI systems are being developed quickly and effectively in the U.S., Europe, and China, why does it matter who builds them?

SS:

• It matters for two reasons. First, the current geographical concentration can counterintuitively slow innovation over time. Second, it shapes which ideas get developed in the first place.

• The strong institutions and mature regulatory frameworks in advanced economies are valuable. But those rules can be cumbersome and difficult to adapt in a timely fashion, which can hinder innovation by limiting opportunities to experiment with and deploy new AI systems.

• Beyond regulatory friction, it also influences which ideas ever get a chance to surface. When frontier AI work is clustered in just a few countries, access to jobs, compute, data, networks, and real-world deployment opportunities also become clustered.

• As a result, many promising ideas, especially those rooted in different languages, sectors, and lived realities, never get explored. Over time, that will lead to fewer breakthroughs and AI systems being shaped by the priorities of a narrow slice of humanity.

The Proposed Solution

MD: To address those problems, you’ve proposed something you call an AI-SEZ. What is it and how does it differ from a traditional SEZ?

SS:

• What I’m proposing is a geographically defined area with permanent regulatory status for AI-related activities. It would operate within broad guardrails set by national constitutions and international obligations.

• Traditional SEZs offer regulatory and infrastructure advantages, but their main incentive to draw in investments has historically been major tax breaks.

• An AI-SEZ is different. It’s not a tax holiday for companies. Its primary draw is maximum regulatory autonomy, which allows faster experimentation and deployment of AI systems within clearly defined limits.

MD: Maximum regulatory autonomy in less advanced economies sounds like a loophole for big tech companies to essentially do whatever they want.

SS:

• Not at all. Regulatory autonomy doesn’t mean an absence of regulation. It means structuring regulation differently so innovation can move quickly without crossing constitutional, legal, or international red lines.

• We know that AI systems evolve faster than traditional regulatory processes can adapt. Rather than rewriting laws every time the technology changes, an AI-SEZ creates a stable environment where experimentation and scaling can happen within limits set by national constitutions, international obligations, and market requirements.

MD: Can you walk me through how maximum regulatory autonomy would be structured in an AI-SEZ? How does it both speed up innovation and expand who shapes AI systems?

SS:

• Regulatory autonomy removes the constraints imposed by existing regulatory frameworks. When companies are no longer bound by the rules of their home market, they can step back and think more creatively about how to structure their operations, their technology, and their business models.

• That creativity is intentional. Rather than the state prescribing a single model, companies or zone operators would propose how they think activities should be organized within the zone, based on what they believe creates the most value. This creates a space where governments, firms, and local institutions learn together by running pilots, building operational capacity, and developing governance practices that can later be applied more broadly. That means local actors would directly shape how the systems are built, deployed, and governed.

  An AI-SEZ is different. It’s not a tax holiday for companies. Its primary draw is maximum regulatory autonomy, which allows faster experimentation and deployment of AI systems within clearly defined limits.

• Those proposals would then be reviewed by an oversight body within the country, which would decide whether they align with national priorities, attract the right kind of investment, and advance the issues the country cares about.

• In practice, this works because regulatory authority within the zone is centralized. Instead of navigating multiple ministries, companies interact with a single office—a “one-stop shop”—that has already been given the necessary powers. This model is already used in places like Panama and Costa Rica, and it’s a key reason those zones are able to operate faster and more efficiently.

The (Hypothetical) Paraguay Example

MD: You mentioned that right now no AI-SEZs with this kind of regulatory autonomy exist yet. How does a country create one?

SS:

• There is no turnkey recipe because every country will have a different process as it navigates politics, its legal system, financing, permitting, available infrastructure, ability to enforce, and so on.

• To give a very simplified example, I can walk you through how it might look in a country like Paraguay.

• The starting point is a feasibility study. Before creating any zone, the country needs to identify whether there is international demand for a particular kind of regulatory flexibility, given what that country can actually offer.

• Paraguay is a useful example because it has very low-cost, 100 percent renewable energy, as well as significant water reserves. Those features make it potentially attractive for energy-intensive parts of the AI value chain, such as data centers that need both power and water.

• If a data-center operator is interested in setting up there, they may also be looking for greater flexibility around environmental guardrails. That’s where the reciprocal exchange comes in. Regulatory flexibility would not be offered unconditionally, but in return for clear benefits to the country itself.

• Those benefits would be defined through performance metrics. For example, a company might commit to helping build out Paraguay’s electrical grid in phases, and continued autonomy would depend on actually meeting those commitments.

• Once demand is established and the terms for the AI-SEZ are clear, the country could then create the zone. The most likely pathways for this are through legislation or a presidential decree. This really depends on the country.

• The AI-SEZ would then rely on an oversight committee, or its “one- stop shop”, to evaluate proposals and make sure investments are anchored in the local economy and aligned with national priorities.

Other SEZs Designed Around Speed And Capability

MD: Are there examples of SEZs that operate with a similar logic?

SS:

• While AI-SEZs with the regulatory autonomy I’ve described don’t exist yet, several governments have announced SEZs for the AI sector, like Oman, India, and Tajikistan. Public information is a bit limited, but so far none of them appear to be operational. As far as we know, they also don’t commit to the kind of regulatory autonomy that we’ve been discussing in terms of having the authority to set and enforce data rules, keep up with the speed of change, or maintain oversight across political cycles.

• There are some special economic zones in other sectors that are prioritizing innovation velocity over tax incentives, and they tend to be among the most successful globally.

• Zones in Costa Rica, Panama, and the United Arab Emirates consistently rank among the top performers because they focus on efficient regulation, strong infrastructure, and high-quality services that attract leading tenants.

• In Costa Rica, for example, that approach helped attract part of the operations of Databricks, a major AI services firm. And in the UAE, the model has gone a step further. They’ve begun exporting their zone framework, most recently through a partnership with Ghana, to help de-risk AI investment and build regional AI hubs.

• Those cases show that when zones are designed around speed, predictability, and capability they can attract serious AI activity and scale beyond their original borders.

Not a Race to the Bottom

MD: AI-SEZs clearly make a country more attractive to investors. But how do they ensure that local populations actually benefit? Isn’t there a risk that countries hungry for investment end up prioritizing tech companies instead?

SS:

• That concern is understandable, but regulatory autonomy does not mean competing on weaker standards. The idea is that, if these zones are designed correctly, market forces can actually push standards upward from the start.

• Many jurisdictions are starting with no AI-specific statute at all. An AI-SEZ lets a country build rules in real time. The zone can establish a clear baseline for how AI systems are governed, covering issues like data handling, audits, incident reporting, and user redress. Operators within the zone can also run supervised pilots and measure outcomes. What proves effective can be translated into national policy, allowing governments to learn what works before applying it nationwide.

• Furthermore, many AI companies want to serve highly regulated markets, such as the EU. To do that, they need to meet strict data-protection and governance requirements wherever they operate. That creates economic pressure to adopt stronger data rules, not weaker ones, even inside an AI-SEZ.

• The same logic applies to talent. In many countries, the absence of AI rules creates a one-way pipeline: ambitious engineers, researchers and founders must leave to level up. That functions as a de facto “race to the bottom” for local capacity. Each wave of departures shrinks the domestic talent base and weakens opportunities for mentorship or entrepreneurship. Over time, that makes the next generation even more likely to leave. AI-SEZs aim to reverse that dynamic by creating opportunities for highly-skilled workers at home.

• Finally, autonomy in an AI-SEZ is conditional. Companies receive regulatory flexibility in exchange for meeting performance requirements tied to national priorities, and an oversight body ensures those commitments are enforced over time. The benefits to companies are explicitly linked to longer-term benefits for the local economy.

Why This Matters Now

MD: Why are AI-SEZs relevant now?

SS:

• AI ecosystems are forming quickly, and early policy decisions are already shaping where innovation, talent, and investment will be concentrated for decades to come. Countries that wait, risk becoming permanent consumers of AI technologies developed elsewhere.

• These zones can build local know-how by creating a focused place to pilot AI systems, train workers, and develop practical governance approaches that can later be applied across the wider economy.

MD: Which countries are AI-SEZs best suited for?

SS:

• I’ve mainly focused my thinking on Latin American countries, especially my home country of the Dominican Republic. That’s because we already have the basis for a strong AI-SEZ with a long history of special economic zones, talent base, and strategic geographical location.

• That being said, this concept can work across emerging and scaling economies provided a few basic preconditions hold. A country must have: 1) reliable power and connectivity; 2) a basic pool of skilled workers; 3) enough rule-of-law and administrative capacity to enforce commitments.

• Fragile states typically miss one or more of these prerequisites because their governments must prioritize security and basic services over tech adoption.

• It would be less appropriate for very advanced economies that have mature institutions and dense, economy-wide regulation.

• Even where national indicators lag, many developing countries have certain areas that are already relatively strong, like industrial parks, major ports, and universities or research centers. AI-SEZs focus on these sites, bringing together the infrastructure and talent needed to run supervised pilots that generate evidence governments can later use to scale rules nationally.

• Ultimately, AI-SEZs give countries a practical way to participate in the AI transition without rewriting the entire legal system at once.

AI Elsewhere is produced by Meenakshi Dalal. For collaboration, story ideas, or related work, you can reach her at info@aielsewhere.com

Brazil has:

● The world’s 10^th highest GDP & largest economy in Latin America

● 180+ data centers—more than any other country in Latin America

● Clean energy infrastructure that appeals to tech companies

● Third-biggest ChatGPT user base, after India and the U.S.

● Fifth-largest internet user base globally

Basically: Brazil is BIG in every possible way. A country with such a huge user base and economy, along with its potential to become Latin America’s data center capital, is very likely to have a significant influence on how AI is shaped.

To explore this, I spoke to Luis Enrique Urtubey De Césaris, who has spent the past year working with policymakers, civil society, and other stakeholders regarding several AI governance initiatives in Brazil. He directs the newly founded Centro de Estudos em Governança de Inteligência Artificial, serves as an AI expert for the Observatory for Catastrophic Global Risks and has spent the last decade as a technology-focused super-forecaster* at Good Judgment.

This interview has been edited for clarity and length. 
Interviewer: Meenakshi Dalal (MD) 
Interviewee: Luis Enrique Urtubey De Césaris (LEU)

Brazil’s AI Bill and Why It Matters Now—Even If You Don’t Live in (or Anywhere Near) Brazil

MD: Can you explain briefly what this bill is?

LEU: Bill 2338 is the main piece of AI legislation currently under debate in Brazil. It establishes a framework to govern AI in the country, designed to mitigate AI risks of every kind.

The bill classifies AI systems by risk level, creates a national governance structure for oversight, and grants regulators several powers to try to guide the development of the technology in a positive direction. It’s also further along in the legislative process than almost any of the other 120+ AI-related bills in the country.

It is important to underscore that the bill is under very active debate and, in case of becoming a law, the specific final wording and ensuing regulatory process is likely to be of great importance.

MD: What does it apply to?

LEU: If it comes into force, it would apply—in a way or another—to all AI systems available for commercial, economic, or public use. Eventually, other laws would be expected to cover other potential uses.

MD: And how close is Bill 2338 to becoming a law?

LEU: It’s about halfway there. The bill was introduced to the Senate in 2023. An amended version passed there in December 2024 and moved to the House for debate in 2025. A newly amended version is expected to go to a House floor vote at the end of 2025.

If it passes, the bill returns to the Senate for a final vote before landing on the president’s desk. Everything else being equal, it’d likely be signed into law in the first half of 2026.

Some articles of the law would enter into force as soon as it’s signed into law, but most provisions would only go into effect six months to two years afterward.

A Roadmap of the Legislative Process for Bill 2338 to Become a Law in Brazil

MD: The number of pieces of AI legislation floating around the world is overwhelming. Why does this bill matter so much both in Brazil and for people and companies nowhere near the country?

LEU: Several actors intend for Brazil’s AI policy to serve as an example for the rest of the world, particularly among countries in the Global South. This goal is explicitly stated in the country’s National AI Plan (the other main AI policy document at present), which is more focused on the development of the technology as such.

It would be analogous to the so-called Brussels Effect,* in which other countries and regions take up, or are inspired by, the standards set in other jurisdictions.

MD: And is it correct to say that Bill 2338 is an example of the Brussels Effect in action?

LEU: Yes, to some extent. The easiest, shorthand way to understand this bill is to know that it is inspired by the European Union's AI Act, but with evolving local differences.

How Brazil’s Bill and the EU’s AI Act Compare

MD: Can you walk me through how Bill 2338 is similar to the EU AI Act?

LEU: In broad terms, Bill 2338 takes core elements from the EU AI Act to address similar concerns. It takes a preventive, harm-reduction approach grounded in transparency and fundamental rights.

Like the EU’s law, it follows what we call a ‘risk-based’ approach. It lays out an AI system’s possible uses and categorizes them as posing an excessive risk, high risk, or other. Each classification carries a set of obligations. For example, a ‘high-risk’ AI system—like self-driving cars—would have to undergo a preliminary risk assessment before it would be allowed in Brazil.

Brazil’s proposed law—like the EU AI Act—also covers systemic risks* from General-Purpose AI* models.

It also sets up a framework for serious incident reporting, establishes some whistleblower protections, and the possibility of third-party audits. The bill creates a national AI governance system, coordinated across agencies and ministries, that would oversee and enforce regulations. In the most extreme case, if a model is deemed too dangerous, it could be banned outright. This would most likely occur if other preliminary measures didn’t reduce risks satisfactorily.

MD: What are the main differences between Brazil’s proposed bill and the EU AI Act?

LEU: While using some central principles from the EU AI Act, lawmakers are also adapting them to local conditions, particularly in how the AI governance system would be structured. The European Union has its own version, with an office at the European level supported by national agencies.

In Brazil, it is a national system designed in a way which also mixes elements of both centralized and decentralized governance, looking to adapt to the needs of each sector, industry, or application. In this sense, they’re trying to make this law as future-proof as possible.

At present, the Brazilian legislature is in the process of trying to strike a more explicit balance between innovation and regulation, compared to previous drafts, showing that both can operate without being at odds with each other.

This is not a straightforward process. A lot of this approach comes from recent public-administration theories and practices—what could be called experimentalist governance.* It’s in part also a political process, so how much of this survives once the rubber hits the road is to be seen.

There are differences around exactly how systems would be regulated and to what extent. For example, the EU’s law has four risk levels, while Brazil’s has three. Also, while both pieces of legislation cover systemic risk, the EU’s definition of systemic risk is more precise than Brazil’s and links to AI systems’ potential to cause catastrophic harm more explicitly. It also addresses AI model size as a key variable, which Brazil’s bill currently does not.

In Flux

MD: Okay, so there are quite a few differences between the EU’s AI Act and Bill 2338. Are any of those differences part of the current debate?

LEU: Yes, they are. For example, definitions are central to ongoing discussions, and we’re looking towards more specificity. Lawmakers are debating which AI uses belong in each category and the processes for updating them.

MD: Why are these important?

LEU: These definitions matter because they determine which AI systems would be regulated and how. If these aren’t clear enough, it may make enforcement difficult later on, as well as creating unnecessary uncertainty.

MD: What else is in flux?

LEU: For example, copyright protection is included in the bill right now, but the extent of those protections is under debate. The final form the governance system will take is an open issue.

MD: I read that the Minister of Finance is courting tech companies to build more data centers in Brazil. From what I understand, this would come through a not-yet-public National Data Center policy that offers both tax incentives and access to Brazil’s renewable energy infrastructure. Does that affect the bill?

LEU: It might, directly or indirectly. The desire to regulate and develop local industry could be in some tension, depending on how it is structured. But we won’t know until the Ministry of Finance publishes the proposed policy, likely sometime in September. It is expected to be focused, in part, on tax exemptions for the data center industry.

Concern and Optimism

MD: During the past year, you’ve talked at length to many people who have an interest in how this bill turns out. What’s one thing you would say people are particularly concerned about?

LEU: Especially when it comes to small and mid-sized companies, there’s concern about having the right responsibilities and obligations assigned to appropriate actors, while simultaneously allowing for the development and deployment of the technology. For example, upstream* and downstream developers* would need clear definitions about what each could do.

MD: What are people optimistic about?

LEU: People are optimistic that this bill and other policies could make Brazil a global leader in the development, deployment and governance of AI.

There’s also excitement when it comes to potential in the public sector: Brazil could create both a robust process of training government staff on AI and enable innovation more broadly. Regulatory sandboxes* could play a role in this.

The primary thing the bill does is try to create the conditions for different stakeholders to engage productively in regulation, and therefore hopefully shape the trajectory of the technology in a way which may serve the public interest.

Very concretely, there’s also optimism that this bill might lead to the prevention of crime and other types of misuse, such as stopping scams caused by AI deepfakes—something that has become a major issue both in Brazil and around the world. The thought is that if successful in reducing harm, Brazil would be a leading jurisdiction developing these regulations and contributing to an emerging global effort.

MD: As a proponent of ethical and safe AI, do you want this bill to pass?

LEU: Yes, I do. The bill is directionally good enough to be a strong starting point, and I expect that adjustments will be made between now and a final vote which will make the bill stronger.

In its current language, the bill gives regulatory bodies the flexibility to adapt the fine print as risks evolve, within a governance system which tries to be accountable and responsive. The bill opens the possibility of mitigating a broad spectrum of negative outcomes, including the most negative. And that’s a good thing.

My Two Cents: The choices Brazil makes in how it balances AI regulation, innovation, and investment will have ripple effects, especially if other countries follow its example. If Brazil offers AI developers leeway that the EU doesn’t, it could result in developers producing different versions of their models, one to meet the EU’s requirements and another to meet those of Brazil and countries adopting similar regulations.

Brussels Effect—When other countries adopt regulations modeled on the European Union’s standards.

Downstream developers—People or organizations that take a general-purpose AI model and customize it for a particular use, like healthcare or education.

Experimentalist Governance—A flexible, trial-and-error approach to policymaking that adjusts as new evidence emerges.

General Purpose AI (GPAI)—An AI system that can perform many different tasks, instead of just one. The same model might be used for image or speech recognition, generating audio or video, detecting patterns, answering questions, or translation.

Super-forecaster—Someone who consistently makes unusually accurate predictions about future events.

Systemic Risk—Large-scale threats that can affect entire societies or economies.

Regulatory sandbox—A controlled environment where new technologies can be tested under lighter rules. It lets companies experiment while giving regulators a chance to observe and figure out how to manage risks.

Upstream developers—People or organizations that build the original, general-purpose AI models before anyone adapts them for specific uses.

Edward Tsoi walked me through what AI in Southeast Asia looks like. The region is quickly becoming central to global AI systems when it comes to adoption and infrastructure, but investments in governance are lagging. 

Ed is the co-founder of AI Safety Asia, a nonprofit working to strengthen safe and effective AI governance across the region through training, research, and policy work with governments and civil society. During the past year, he’s led seven regional roundtables, and the organization is about to launch an AI governance training program in Indonesia for 30 senior officials in partnership with the Ministry of Communication and Digital Affairs. Originally from Hong Kong and now based in London, he sits on the boards of two university AI divisions and previously founded an education NGO in Myanmar. Long story short: Ed knows what’s up.

This interview has been edited for clarity and length. 
Interviewer: Meenakshi Dalal (MD) 
Interviewee: Edward Tsoi (ET)

AI Optimism and Investment in Southeast Asia

MD: Is what I’ve been reading true: that people in Southeast Asia see AI as a net good? I came across a recent Stanford University study showing that 80% of people in Indonesia and 77% in Thailand believe the benefits of AI outweigh the risks—nearly double the percentage in the U.S. and Canada.

ET: People see technology as a force for good in a way that is often taken for granted in the Global North. It’s viewed as something that can quickly boost quality of life. In Southeast Asia, digital tools like messaging or mobile payments have leapfrogged entire stages of development, saving people hours of travel just to make a call or transfer money. This kind of ease helps explain the eagerness around adopting AI in Southeast Asia.

MD: That sounds very positive. How else is AI showing up in the region?

ET: Southeast Asia is already part of the global AI value chain. Companies rely on real people to do something called data annotation, which involves reviewing and labeling the information that AI models are trained on. It’s created a lot of job opportunities, particularly for gig workers.

We’re now also seeing major tech companies invest heavily in the region, in a way that will make Southeast Asia a much bigger player in the global AI system. We’re talking about upwards of $30 billion in committed investments to build data centers and other infrastructure projects in countries like Singapore, Malaysia and Thailand.

MD: But?

ET: What we are not seeing is matching investment on the governance side, which is critical for AI to be adopted safely. The real issue isn’t that Asia will somehow fall behind when it comes to AI: it’s that it’s already central and strong governance needs more attention.

Why Investing in AI Governance in Asia Matters

MD: Let’s start with why investment in safe AI governance matters in the first place.

ET: People often frame growth and regulation, innovation and safety, as opposites. They’re not. We can—and must—have both.

Take cars, for example. They’re powerful machines that improve everyday life but can also be dangerous. That’s why we have seat belts, traffic lights, and road laws. Without seat belts and traffic regulations, driving will be a dangerous mess, and it would be bad for the car industry. Regulations don’t eliminate risk, but they reduce it significantly. These guardrails benefit both the public and the automotive industry.

That’s our argument at AI Safety Asia: adopting AI safely through a mix of governance, regulation, and legislation is a win for both the industry and society.

MD: What are you worried about when you say ‘risk’?

ET: We are concerned about a spectrum of risks, from those we are seeing right now including bias, discrimination and disinformation, to risks that are coming soon with early warning signs from empirical AI safety research. These include biosecurity, loss of control, and alignment problems.

MD: I’ve heard people agree wholeheartedly on the need for safety and then argue that the biggest gains in AI safety governance are going to come from the West. What do you say to them?

ET: That’s a bit shortsighted. AI is a global technology, and its risks manifest differently depending on the context. Misinformation in London looks very different from misinformation in Yangon.

Sixty percent of the world’s population lives in Asia. It’s a diverse region with distinct political, social, and technological environments. We can’t afford to operate blindly there or apply another region’s lens without deep consideration. The challenge is that right now governance, research, technical talent, and safety infrastructure across Southeast Asia are uneven—advanced in some countries and just beginning in others. That’s why we started AI Safety Asia: to help build the policy, technical, and research capabilities needed to govern AI safely.

How Southeast Asia is Approaching AI Governance

MD: AI Safety Asia started its work in Southeast Asia. Why?

ET: It’s where we could have the biggest impact early on. Of all the sub-regions in Asia, it had a less developed AI ecosystem. One thing that stands out after hundreds of hours of roundtables and listening exercises is that most countries in this region—aside from countries like China and Singapore—are not trying to build the next ChatGPT or Claude. They fully accept that, at least for now, they will be the users of this technology, not the producers.

MD: What’s political will like?

ET: Like everywhere, AI is a popular topic for politicians. Everyone understands that AI is going to reshape their countries in big ways.

MD: And how does that understanding translate into governance?

ET: Unevenly. In Southeast Asia we’re looking at a wide spectrum of abilities, priorities, financial resources, and progress. Six countries have already set up their national AI roadmaps or strategies. Some of them are even starting to consider setting up governance boards and research centers.

MD: Can you elaborate a bit more on that unevenness?

ET: On one end, you have Singapore, which is moving fast when it comes to adopting AI policies, building infrastructure, and helping people adopt the technology in a constructive way. On the other, you have places like Myanmar, where an ongoing civil war makes AI governance a low priority. Or, countries like Laos or Cambodia that haven’t begun drafting legislation yet.

In between, there’s a lot of variation.

Take Indonesia: it’s actively working on AI governance but facing challenges that could severely limit progress. Its electrical supply per capita is about a third of Singapore’s and the government is in the middle of a massive (and expensive) effort to relocate the capital city, which is currently sinking.

In the Philippines, the government is laser-focused on the job displacement angle. About 8% of the country’s GDP comes from the outsourcing industry and AI has the potential to disrupt that.

What’s Holding Effective AI Governance in Southeast Asia Back

MD: What do you make of these initiatives so far?

ET: I see two main issues.

First, even when leaders are pushing pragmatic AI policies, coordination across government departments is often weak. Every department wants to be the champion of AI, which ironically can stall progress.

Secondly, many national AI strategies are super high-level and vague, sometimes more slogan than substance. They often lack the concrete detail needed to guide meaningful policy.

MD: What do you think is driving these challenges?

ET: A lot of it comes down to familiarity and understanding of AI—not a lack of will. At the roundtables, many officials expressed feeling a bit lost. Some said they’re not familiar with how this technology works at a very basic level, while others want a better handle on the full spectrum of risks and opportunities that AI brings.

They’re also trying to absorb the latest research on technical safety and AI governance to understand the key lessons learned in other parts of the world and figure out what makes sense to incorporate into their own national contexts.

What’s next?

MD: What will build the momentum needed to move safe governance in Southeast Asia forward?

ET: First, there needs to be more investment in building know-how among the people making the laws. No matter whether the civil servants work in digital affairs or science and technology, they need a solid understanding of how this technology works, its risks, and what governance options exist globally, which AI Safety Asia is supporting through our AI governance training program.

Secondly, the 2026 AI Impact Summit, which will be hosted by India, is a big opportunity. It’s a chance to bring equitable AI, and the priorities of Southeast Asia (and other countries in the Global South) to the center of the global conversation. Those contributions can’t be surface level.

Finally, collective and harmonized influence matter. Individual countries will have a hard time gaining traction alone. If Southeast Asian countries want to drive meaningful AI policies that genuinely serve their people, especially when navigating the priorities of powerful multi-trillion-dollar tech companies, they’ll need to come up with a harmonized approach. Tough, but not impossible.

Follow AI Safety Asia’s work via their LinkedIn page and website. The organization will be publishing its first research report with the Brookings Institution on AI Governance in Southeast Asia in the coming month. They also run the Southeast Asia AI Observatory that tracks policies and legislation coming out of the region.