As Mauritius accelerates its digital transformation, artificial intelligence is moving from buzzword to policy priority. In this interview with Nad Sivaramen, AI expert Dr. Kwet Ng outlines the opportunities, risks and regulatory choices facing our country, drawing on his proposed framework for a Mauritian AI Act...

?What does artificial intelligence actually mean for the average Mauritian today — and how might it reshape daily life over the next five years?

For most Mauritians, AI is majorly present in terms of Generative AI. This is so called because it generates text, images or videos after receiving instructions from the user. Unfortunately, for the general public, AI is mainly GPTs or more commomnly known as chatbots. However, in real world, chatbots such as Chatgpt is only a small part of AI and now the trend is to move from Large Language Models (LLM), foundation for chatbots, to predictive AI and real world AI. Big name researchers such as Yann LE Cun, considered as one of the godfathers of AI, has clearly stated that LLM is reaching its limit and real world AI models is the next step if we want to achieve Artificial General AI (AGI) where AI will be able to mimic the human behaviour in terms of thinking and reasoning. This theoretical state has not been reached yet and in my opinion, mainly because of the excessive focus on LLMs and chatbots. Said focus can easily be explained by the need for all major tech companies to monetize AI as quickly as possible and for its use to be profitable.

AI is a huge umbrella under which many misconceptions and false attributions can be hidden. As mentioned earlier, most Mauritian entities using AI are mainly at the basic level of chatbots. There are even financial institutions which are claiming that they use AI for credit granting while the use of chatbots is not efficient for that particular purpose. The average Mauritian, excluding chatbots, will not perceive the use of AI directly in his/her daily life but will be told that the responsible for many tasks is an AI system. Just blame the AI especially when things go wrong!

If the proper investment is made, the average person will be impacted indirectly such as in the hospitals, banks or public entities . A simple example would be that a bank using an AI predictive model, not chatbot, will be able to detect fraudulent transactions, send an email to the account holder who can validate if the transaction is truly fraudulent or not. The user will not know if this was done by AI or not but this will be really useful to avoid those unpleasant surprises in our personal finance. Similarly, in terms of healthcare, use of images prediction models such as Joint Embedding Predictive Architcture(JEPA) could predict how a particular disease is expected to evolve and which treatments should be applied accordingly. Obviously, the patient will not be aware of the use of AI but will certainly benefit from it.

Another major impact will be on education and already Mauritius Telecom has developed an in house LLM for students. This is a great initiative though safeguards in terms of legislation need to developed to define clearly responsibilities in terms of misuse and thus the need for a Mauritian adapted version of the EU AI act. In Europe, we have had cases of suicide due to excessive and abusive use of chatbots developed by the big tech or recent cases of generative naked images of individuals through the misuse of those chatbots. Prompt injection, jailbreaking, and adversarial prompt design are core to AI security and LLMs are notoriously vulnerable to such attacks. Therefore, it should be clear who is responsible for what and this should be clearly defined in legislation especially with regards to big tech companies. . The practical texture of daily life will shift in subtler ways too. Job applications will be filtered by AI screening tools before any human reads them. Social-benefit eligibility decisions will increasingly be shaped by algorithmic assessments. Traffic signals will respond to real-time flow data rather than fixed schedules. These changes raise a question the Policy Blueprint asks directly: does the average Mauritian understand when an AI system is making or influencing a consequential decision about them, and do they have meaningful recourse? The honest answer, in 2026, is still no — which is precisely the governance gap the proposed Act seeks to close.

By 2030, under the Digital Mauritius 2030 Strategic Plan, the government's stated ambition is for AI to be integrated into 50 percent of public services. Whether that target materialises as genuine improvement or as technocratic opacity without accountability will depend almost entirely on whether the regulatory architecture proposed in these papers is enacted and staffed with genuine technical capacity.

?You describe Mauritius as being at a 'critical juncture' in its digital transformation. What makes this moment decisive, and what risks does the country face if it fails to act now?

The critical juncture argument rests on convergence of three forces that are unlikely to align again in the same configuration.

First, the EU AI Act entered graduated enforcement from August 2024, with prohibitions on unacceptable-risk systems already binding, high-risk obligations phasing in through 2026, and full enforcement expected by mid-2027. Mauritius's International Financial Centre business model depends on its attractiveness to multinational corporations and fund managers headquartered in Europe. Those entities are now required to document and govern the AI systems embedded in their operations. If Mauritius cannot demonstrate a credible, compatible governance framework, it becomes a regulatory blind-spot rather than a trusted gateway — a reputational risk that no marketing campaign can reverse.

Second, the 2025–2026 National Budget explicitly allocated MUR 25 million for a Public Sector AI Programme, establishing a dedicated AI Unit within the Ministry of Information Technology. This is the first moment in Mauritius's policy history when political will, budgetary commitment, and legislative readiness have intersected. The Foundation Paper notes that the Mauritius Artificial Intelligence Council proposed in the 2018 AI Strategy was never constituted with statutory authority — a lapse that allowed seven years to pass without enforceable governance. Repeating that pattern now, when AI adoption is growing exponentially, would produce a far more damaging vacuum.

Third, domestic AI deployment is already outpacing governance capacity. Smart city surveillance systems incorporating computer vision and facial recognition are operating with no comprehensive legal framework. Algorithmic trading and robo-advisory services in the financial sector raise explainability and liability questions that current legislation cannot answer. Credit-scoring AI affects access to mortgages and business loans without any obligation to explain rejections to applicants.

The cost of inaction is not hypothetical. It includes discriminatory outcomes embedded and amplified at scale, erosion of trust in digital public services, regulatory arbitrage by foreign technology firms seeking the least-scrutinised jurisdiction for deployment, and eventual forced legislative catch-up under external pressure — on less favourable terms than a proactively designed framework would secure.

?Mauritius aspires to position itself as an AI innovation hub in Africa. What tangible advantages does it hold — and what structural weaknesses still constrain that ambition?

The Foundation Paper catalogues Mauritius's genuine structural advantages with appropriate rigour. The country has a GDPR-aligned Data Protection Act 2017, a multi-lingual workforce capable of serving Francophone, Anglophone, and increasingly Asian markets, stable democratic institutions, and a well-capitalised financial services sector that constitutes 13 percent of GDP. The Mauritius Africa Fintech Hub, the Financial Services Commission's Fintech Innovation Lab, and the Bank of Mauritius's Open Lab provide functioning sandbox infrastructure that many larger African economies lack. Huawei's USD 20 million AI lab, Bank One's USD 100 million payment platform, and approximately 30 AI-focused fintech startups by 2024 represent a nascent but real ecosystem.

The country's geographic position at the intersection of African, Asian, and European trade routes gives it a convening role that exceeds its population size. It is party to the African Continental Free Trade Area, the Southern African Development Community protocols, and bilateral agreements that create natural corridors for AI services exports. Its relatively low sovereign risk and established common-law legal system provide predictability attractive to technology investors.

The structural weaknesses, however, are equally clear. Technical talent is the binding constraint. The country's entire ICT and BPO sector employs approximately 34,500 people across 975 companies — a pool insufficient to staff a credible national AI regulator, a research university with frontier capability, a thriving startup ecosystem, and major corporate AI deployments simultaneously. The Brain drain to Europe, the Gulf, and North America is persistent. The 2025–2026 Budget's ambition to certify 5,000 AI professionals by 2024 (en route to 10,000 by 2030) is directionally correct but almost certainly insufficient at the pace of global AI advancement.

A second constraint is data infrastructure. AI systems require large, diverse, and well-governed datasets. Mauritius's domestic market of 1.3 million people limits the volume of locally generated data, and the country lacks a coherent national data-sharing framework that would allow public-sector datasets in health, transport, and agriculture to be made available for research and innovation under appropriate safeguards.

A third major constraint is energy and water. For huge datacenters, critical for AI development, huge amount of energy and water is needed and both are scarce commodities especially in Mauritius where there is clearly a lack of energy supply and the use of domestic water will quickly lead to social problems especially in drought periods which are now more common due to climate change.

By April 2026, the emergence of open-weight frontier models — including Meta's LLaMA series and Mistral's releases — has partially democratised access to powerful AI capabilities, reducing Mauritius's dependence on expensive proprietary API access. This shift creates an opportunity for local fine-tuning on domain-specific data, particularly in multilingual applications and agriculture, where the Foundation Paper identifies a comparative advantage in tropical-environment computer vision and local-language natural language processing.

?For entrepreneurs and investors, where are the most immediate AI-driven opportunities — across finance, agriculture, healthcare or public services — and what kind of returns or transformation should they realistically expect?

The Foundation Paper's sector analysis points to financial services and agriculture as the two domains where the combination of existing infrastructure, clear use-cases, and measurable outcomes makes near-term returns most credible.

In financial services, the highest-immediacy opportunities lie in regulatory technology (RegTech) and fraud analytics. Mauritius's role as an International Financial Centre means that anti-money laundering and Know Your Customer processes generate enormous compliance burdens. AI systems that automate transaction monitoring, adverse-media screening, and sanctions-list checking can deliver measurable cost reductions within a twelve-month deployment cycle. The fintech lending segment — following FundKiss's model — offers a second high-return opportunity: alternative credit scoring using psychometric, mobile behaviour, and transaction data to serve the estimated 30 percent of the adult population without conventional credit histories. Realistic returns in this segment are those of a speciality lender with lower origination costs, not a technology multiple.

In agriculture, the Smart Agriculture Project's documented results — 20 percent reduction in water usage, 15 percent yield increase adding USD 30 million to annual exports — validate the business case for AI-powered precision farming. The extension of these systems from sugarcane to vegetables, fruits, and aquaculture is underway. Investors should treat this as infrastructure-style returns: capital-intensive, moderate risk, long payback periods, but supported by government subsidy frameworks and export revenue guarantees.

Healthcare's opportunity set is real but time-horizoned by regulatory complexity. AI diagnostic support for radiology and pathology has the clearest clinical pathway, but deployment requires medical device validation protocols that Mauritius has not yet formalised. The dengue surveillance model — a public health analytics system rather than a clinical AI — is replicable across vector-borne and non-communicable disease monitoring and represents a lower-regulatory-barrier entry point.

In public services, the MUR 25 million Public Sector AI Programme creates a government procurement opportunity for AI-powered document processing, citizen service chatbots, and fraud detection in social benefits. These are contractual revenues rather than equity stories, but they offer stability and reference-client value for startups seeking to scale regionally.

The honest expectation-setting point the Foundation Paper implies, even if it does not state it directly, is this: Mauritius's market size limits purely domestic addressable markets. The real value proposition for AI investors is Mauritius as a base for building AI services products that serve African and Indian Ocean markets — a jurisdiction combining regulatory quality, tax efficiency, and linguistic reach. That story requires the AI Act to exist and be credible.

?You propose a Mauritian AI Act inspired in part by the EU AI Act. Why, in your view, is regulation not a brake on innovation but a condition for it — particularly in a small economy?

The Policy Blueprint addresses this directly and its logic deserves full elaboration, because the innovation-versus-regulation framing is the most persistent obstacle to legislative action in small developing economies.

The first argument is market access. Mauritius's largest export markets — the European Union, the United Kingdom, and increasingly the Gulf — are implementing or have implemented AI governance requirements. A Mauritian AI company seeking to deploy a credit-scoring or insurance-underwriting system for European clients must demonstrate compliance with the EU AI Act regardless of what Mauritius chooses to legislate domestically. The choice, for Mauritius, is not between regulation and no regulation — it is between designing regulation that serves Mauritius's economic interests or importing compliance obligations designed in Brussels with no Mauritian input.

The second argument is investor confidence. The Foundation Paper notes that the absence of legal certainty about AI liability, data governance, and conformity assessment creates a due-diligence problem for institutional investors. Venture capital and private equity funds allocating to African markets apply a regulatory-risk discount to jurisdictions without clear AI frameworks. A well-designed Act with proportionate requirements reduces that discount, lowering the cost of capital for Mauritian AI startups.

The third argument is trust infrastructure. Consumer adoption of AI services — whether in digital banking, telemedicine, or agri-tech — depends on public trust. Trust requires visible accountability: a regulator that investigates complaints, redress mechanisms that function, and transparency obligations that allow citizens to understand consequential decisions. Without that architecture, adoption plateaus at the technologically curious early-adopter segment and never reaches the mass-market penetration that generates transformative economic returns.

The fourth and most underappreciated argument is that regulatory sandboxes — a centrepiece of both the Policy Blueprint and the Foundation Paper — are themselves innovation infrastructure. The Financial Services Commission's Fintech Innovation Lab and the Bank of Mauritius's Open Lab have already demonstrated that structured supervised experimentation accelerates product development and reduces time-to-market. The proposed MAIRA-operated AI sandboxes extend this model to all sectors, providing startups with a legally protected space to test systems that would otherwise carry unquantifiable liability exposure.

In small economies, where no single company has the scale to absorb catastrophic regulatory penalties, the certainty provided by a clear framework is more valuable than the freedom provided by an absence of rules.

?If AI systems were allowed to develop without proper safeguards in Mauritius, what would a worst-case scenario look like in practical terms — for citizens, businesses and the state?

The Foundation Paper's governance-gaps analysis, read against Mauritius's specific sectoral deployment patterns, produces a coherent worst-case scenario that is neither speculative nor remote.

For citizens, the most immediately damaging pathway runs through financial services. Credit-scoring AI trained on historical data that reflects past discrimination will perpetuate and amplify exclusion of lower-income, rural, and Creole-community applicants — denying mortgage access, business loans, and insurance on the basis of proxies for protected characteristics that no applicant can challenge because no explanation is required. The Data Protection Act's Article 38 on automated decision-making provides a theoretical remedy but — as the Foundation Paper notes — gives no guidance on what constitutes adequate explanation, making enforcement practically inoperable.

In public administration, AI-powered eligibility screening for social benefits without human-review requirements creates a pathway to systematic exclusion of the most vulnerable populations — those least able to navigate administrative appeals. The Ministry of Social Integration's stated interest in AI systems for identifying vulnerable populations represents, without safeguards, a scenario where the same system that is meant to help could exclude on the basis of opaque algorithmic scores.

For businesses, the worst-case scenario is regulatory fragmentation followed by forced catch-up. As the EU AI Act's extraterritorial reach tightens through 2026 and 2027, Mauritius-based companies serving European markets that have not invested in compliance infrastructure face simultaneous exposure: EU enforcement action from abroad, and domestic liability claims that Mauritius courts are unprepared to adjudicate because no AI-specific legal framework exists.

For the state, the Foundation Paper identifies smart city surveillance as the highest-risk domain. Computer vision and facial recognition systems deployed for traffic management or public safety, without a legal framework governing their scope, retention, access, and oversight, represent the infrastructure of a surveillance state that Mauritius has not debated and has not chosen. The 2021 controversy over proposed ICT Act amendments for social media monitoring — which attracted international human rights condemnation and was eventually modified — demonstrated that Mauritius's democratic institutions can push back. But that pushback requires the issue to become visible. AI-embedded surveillance scales quietly.

The systemic risk scenario — relevant given Mauritius's financial centre role — involves algorithmic trading contagion. Coordinated AI-driven trading systems operating without circuit-breaker requirements or explainability obligations could propagate market shocks from global exchanges through Mauritius's financial infrastructure to African clients, triggering cross-border liability questions that no existing legal framework can resolve.

?Your proposal includes the creation of a Mauritius AI Regulatory Authority (MAIRA) with significant technical powers. Given the country's limited pool of AI expertise, how do you ensure such an institution becomes effective rather than merely symbolic?

This is the most operationally demanding question in the entire policy framework, and both papers address it with uncommon frankness. The Foundation Paper explicitly states that effective AI regulation demands technical capacity currently limited in Mauritius — a concession that distinguishes rigorous policy design from aspirational declaration.

The solution framework has several interlocking components. The first is competitive compensation. Regulatory salaries in Mauritius's public sector are structurally below private sector equivalents for technical roles. A MAIRA that pays data scientists and ML engineers civil-service rates will lose every recruitment contest to the fintech and BPO sectors. The funding model proposed — combining government budgetary allocation, registration fees from AI system providers, sandbox participation revenues, and international development assistance during the initial phase — is designed specifically to give MAIRA financial independence sufficient to set market-competitive salaries within its technical units.

The second component is the secondment and partnership model. Rather than attempting to build all expertise in-house, MAIRA should operate a rolling programme of secondments: regulatory staff spending twelve months embedded in the Financial Services Commission, University of Mauritius AI research groups, or international counterparts including the UK AI Safety Institute and Singapore's AI governance division. The reverse flow — private sector and academic AI specialists spending fixed terms within MAIRA — transfers institutional knowledge and builds regulatory credibility simultaneously.

The third component is scope discipline. A newly established regulator with limited staff that attempts to enforce every provision on day one will fail visibly and lose authority. The Foundation Paper's phased implementation strategy is, in effect, a capacity-building plan: Year One focuses only on prohibited practices and transparency obligations; Years Two and Three introduce high-risk conformity assessment; full enforcement begins in Year Three. This sequencing allows the regulator to build competence, case law, and institutional confidence before taking on the most technically complex assessments.

The fourth component is the Advisory Council. The proposed multi-stakeholder body — comprising academia, industry, civil society, and sectoral regulators — functions as a continuous technical advisory service, compensating for gaps in MAIRA's own expertise. For specialised assessments, accredited third-party auditors operating under MAIRA oversight perform conformity assessments, following the EU's notified body model. MAIRA's role in those cases is oversight and enforcement of auditor standards, not direct technical assessment.

By April 2026, the UK AI Safety Institute and Singapore's Model AI Governance Framework have published detailed technical evaluation methodologies that MAIRA could adopt and contextualise for the Mauritian environment without incurring the development cost. International cooperation agreements with these bodies, proposed in the Policy Blueprint, make that knowledge transfer concrete rather than aspirational.

?Your framework imposes strict obligations on high-risk AI systems. In an SME-driven economy, is there a danger that compliance costs could stifle local innovation and tilt the playing field towards larger foreign actors?

The concern is legitimate and both papers engage with it directly rather than dismissing it. Mauritius's ICT and BPO ecosystem is built on approximately 975 companies, the majority of which are small or micro enterprises. An AI compliance framework modelled entirely on EU obligations — designed for Volkswagen and Deutsche Bank — would be prohibitively expensive for a five-person fintech startup in Ebène Cybercity.

The Policy Blueprint's response is the SME-relief architecture embedded within the framework itself. This includes extended compliance timelines for startups and microenterprises, simplified documentation templates replacing full technical dossiers, fee waivers for conformity assessment costs, free advisory services from MAIRA on regulatory interpretation, and sandbox access without the capital requirements that conventional market deployment would entail. The explicit intent is that reduced requirements should not compromise safety or fundamental rights — particularly for high-risk systems — but that the path to demonstrating compliance should be proportionate to organisational capacity.

The Foundation Paper adds an important structural observation: the risk of tilting the playing field toward larger foreign actors exists with or without an AI Act. Without a framework, large multinationals deploying AI systems in Mauritius from their home jurisdictions face no local obligations at all, while domestic startups bear reputational and liability exposure without the legal certainty of a defined compliance regime. The Act, by establishing requirements for all providers regardless of origin, actually levels the playing field by forcing foreign actors into the same compliance architecture rather than allowing them to operate under a regulatory double standard.

The voluntary certification scheme for minimal-risk systems — where providers meeting voluntary standards receive official recognition — creates a positive market incentive without mandatory burden. Startups building low-risk AI tools (recommendation engines, content classification, customer service automation) can voluntarily certify and use the certification as a competitive differentiator with enterprise clients and government procurement processes without bearing the full cost of high-risk conformity assessment.

The realistic risk that remains is implementation asymmetry: large foreign firms with dedicated compliance teams absorb regulatory requirements as fixed costs and adapt quickly, while small domestic firms face a disproportionate management burden even with simplified pathways. This argues for MAIRA investing heavily in its compliance assistance function — not just issuing guidance documents, but providing active advisory support — in the first two years of operation.

?You also seek to regulate AI systems developed abroad but deployed locally. How can a small island state realistically enforce such rules on global technology firms without jeopardising access to their tools?

This question sits at the heart of every small-state regulatory design challenge, and the Policy Blueprint addresses it through a combination of legal mechanism, diplomatic strategy, and pragmatic sequencing.

The legal mechanism is the authorised representative requirement. Following the EU AI Act model, foreign providers placing AI systems on the Mauritian market or deploying them within Mauritius must designate an authorised representative — a legal entity or individual resident in Mauritius — responsible for compliance obligations and serving as the contact point for MAIRA investigations. This single requirement transforms extraterritorial enforcement from a diplomatic problem into a domestic enforcement action: MAIRA pursues the local representative, not the foreign developer.

The diplomatic mechanism is mutual recognition and information-sharing agreements. The Policy Blueprint proposes memoranda of understanding with the EU AI Office, the UK AI Safety Institute, and Singapore's AI governance framework. Under these arrangements, if the EU has already conducted conformity assessments of a system deployed in Mauritius, MAIRA can accept that assessment rather than duplicating it — dramatically reducing the enforcement resource required. Conversely, Mauritius can contribute intelligence about deployment patterns and incidents in its jurisdiction to international regulatory databases, making it a net contributor to global oversight rather than a free-rider.

The pragmatic sequencing argument is this: Mauritius does not need to enforce its full framework against every global technology firm simultaneously. The initial enforcement focus on prohibited practices — which are narrow and severe — targets the clearest violations. High-risk system obligations, where the international regulatory context is already robust, rely on the shadow of EU enforcement to produce compliance in systems that serve cross-border markets.

There is an important calibration point the Foundation Paper is candid about: Mauritius cannot directly regulate international GPAI providers — OpenAI, Google, Anthropic, Meta — through domestic legislation alone. What it can do is establish clear legal expectations for providers serving Mauritian markets, require local deployers using GPAI APIs to conduct downstream risk assessments, and participate in international regulatory coalitions where collective leverage is greater than any small state's individual authority. The African Union's Digital Transformation Strategy, which progressed materially through 2025 and 2026, creates exactly that collective forum.

The jeopardising-access concern is real but overstated. Global technology firms do not exit markets over reasonable governance requirements. They exited Russia over sanctions, not over AI regulations. They adapted to GDPR rather than withdrawing from Europe. The more relevant historical precedent is that clear, predictable rules increase rather than decrease long-term market participation by technology firms that prefer regulatory certainty to legal ambiguity.

?The blueprint promises transparency, explainability and avenues for redress. In practice, how can ordinary citizens challenge opaque algorithmic decisions, particularly in sensitive areas such as finance or public services?

The transparency and redress architecture in both papers is among the most carefully designed elements of the framework, precisely because the authors understand the gap between rights on paper and rights in practice.

The immediate right — disclosure — requires that citizens be informed when an AI system is making or significantly influencing a consequential decision about them, and be provided with contact information for the responsible party. For a mortgage rejection, credit application refusal, or social benefit denial, this means the decision notice must identify the AI system involved and name the entity accountable for it. This is the precondition for any further challenge: without knowing an AI system was involved, no appeal can be framed.

The right to explanation requires, in the language of the Policy Blueprint, reasons for consequential decisions and accessible appeals. The Foundation Paper's analysis is more technically nuanced here: it acknowledges that for complex neural networks, true explanations of decision processes may be technically infeasible. What the framework requires is not explainability in the technical sense of full causal transparency, but meaningful explanation in the legal and administrative sense — a statement of the primary factors that led to the outcome, sufficient for a person to understand why they were treated differently from others and what they could change.

The right to human review is the most operationally powerful provision. For high-impact decisions — credit denial, benefit refusal, employment rejection — citizens may request review by a human decision-maker who is not simply rubber-stamping the AI output but is genuinely empowered to override it. The Policy Blueprint requires that high-risk AI systems include override and stop capabilities accessible to oversight personnel.

The complaint mechanism runs through MAIRA with coordination from sectoral regulators. A citizen who believes their financial services AI treatment was discriminatory can file with MAIRA, which has investigative powers including access to AI system documentation and, critically, source code. The burden-shifting mechanism proposed in the Foundation Paper — where a person who establishes prima facie harm and AI system involvement shifts the burden of proof to the defendant — is a pragmatic solution to the information asymmetry problem. Citizens cannot access training data and model weights; regulators and courts can.

For collective harm, the class-action mechanism enables groups of individuals affected by the same system to pursue remedies jointly, dramatically reducing the per-person cost of legal action. Legal aid provisions ensure this is not available only to those who can afford private counsel.

The honest caveat is that all of this requires MAIRA to be staffed and operational, the complaint process to be genuinely accessible in Kreol and French as well as English, and the courts to develop AI-litigation competency. These are implementation challenges, not design flaws — but they are challenges that require sustained resourcing to overcome.

You outline a phased implementation between 2025 and 2030. In a rapidly evolving technological landscape, is Mauritius at risk of moving too slowly — and what level of political commitment will be required to ensure this does not remain a paper exercise?

The speed-versus-capacity tension is the central implementation dilemma, and the Foundation Paper addresses it with more realism than most comparable policy documents. The honest answer is: yes, there is a risk of moving too slowly, and the risk is asymmetric — the cost of inadequate governance compounds over time as deployment scales, while the cost of a brief delay in enactment is comparatively contained.

The 2025–2030 phasing is not a leisurely timetable. It reflects a binding constraint: regulatory capacity cannot be willed into existence faster than talent can be recruited, trained, and made effective. A framework enacted overnight with no technical staff, no guidance documents, no sandbox infrastructure, and no enforcement budget would be worse than the status quo — it would create legal obligations without compliance pathways and deter investment without protecting rights.

The adaptive governance mechanisms built into the framework are the primary hedge against technological obsolescence. Technical annexes enumerating high-risk domains and prohibited practices are designed for two-year review cycles with streamlined amendment procedures, allowing MAIRA to add AI capabilities — such as autonomous weapon systems or AI-generated financial advice — to the high-risk list without requiring full parliamentary legislation each time. Sunset clauses require periodic reauthorisation of specific provisions, creating mandatory legislative attention rather than passive drift.

By April 2026, the global AI governance landscape has moved faster than most 2024 policy papers anticipated. The release of GPT-4o, Claude 3.5 and subsequent versions, Gemini 2.0, and multiple open-weight frontier models has compressed the timeline between laboratory capability and consumer deployment. The EU AI Act's GPAI provisions, designed with GPT-4-era systems in mind, are already being tested against autonomous agent capabilities that did not exist at drafting. Mauritius's framework, if it is to avoid the EU's already-visible adaptation problem, should build even more explicit future-proofing: principle-based rather than taxonomy-based definitions of AI systems, and MAIRA advisory council horizon-scanning as a formal annual obligation rather than an optional activity.

On political commitment: the Foundation Paper's conclusion is characteristically direct. Legislation alone cannot ensure responsible AI innovation. What is required is sustained cross-party commitment over a five-year period that will span at least one general election cycle. The history of Mauritius's well-designed but under-implemented strategies — the 2018 AI Strategy's proposed MAIC never constituted; the Digital Mauritius 2030 targets progressing unevenly — suggests that the critical failure mode is not legislative design but post-enactment resource allocation and political attention.

The minimum political commitment threshold is: a MAIRA budget that cannot be raided for other priorities, a recruitment mandate with salaries set above civil-service bands, a direct accountability line from the MAIRA Commissioner to the National Assembly's oversight committee, and a standing item in every budget cycle for AI governance resourcing. Without these structural commitments, the Act becomes a declaration of intent rather than an operational framework.

?Mauritius has no shortage of well-crafted strategies that fall short in execution. What must political leaders, institutions and universities do differently this time to turn AI into a genuine engine of economic and societal transformation?

The Foundation Paper closes with precisely this challenge, and its final paragraphs contain the most important observation in either document: the journey matters as much as the destination. The following analysis draws directly on its prescriptions and extends them with the lessons visible from April 2026.

Political leaders must break the strategy-without-accountability cycle. Every Mauritian digital strategy since 2001 has included ambitious targets, ministerial launches, and international conferences. What has been absent is a transparent, publicly reported accountability framework that names responsible officers, publishes progress against targets quarterly, and triggers budget consequences for non-delivery. The AI Act must include mandatory annual reporting by MAIRA to the National Assembly — not a glossy publication for the Economic Development Board, but a technically detailed account of enforcement actions, compliance rates, sandbox outcomes, and capacity gaps, tabled in Parliament and available to the public.

Institutions must move from coordination to integration. The fragmentation of digital governance across the Data Protection Office, CERT-MU, Financial Services Commission, Bank of Mauritius, ICTA, and multiple ministries has produced regulatory gaps precisely where they are most damaging — at the intersections. The AI Act's coordination mandate for MAIRA is necessary but insufficient without dedicated joint investigation units, shared technical platforms, and secondment programmes that build personal relationships across institutional boundaries. Coordination that relies on formal committee meetings alone will fail.

The University of Mauritius and partner institutions face a specific challenge: building the talent pipeline for a governance and innovation ecosystem simultaneously, without the research funding or faculty density of European counterparts. The honest prescription is specialisation over breadth. Rather than attempting to replicate a full AI curriculum across all sub-disciplines, Mauritius's universities should identify three to five domains where they can develop internationally recognised research capability — tropical-environment AI applications, multilingual NLP for Indian Ocean languages, AI governance and policy for small island developing states, financial AI for emerging-market applications — and concentrate resources accordingly. Industry partnership agreements that fund PhD positions and provide real-world datasets make that research productive and retain talent domestically.

The private sector's role is underspecified in both papers. The thirty AI-focused fintech startups, the Huawei AI lab, and Bank One's payment platform are assets. What is missing is a private-sector AI governance council — distinct from government — that develops industry codes of practice, shares incident data across companies, and provides the regulatory ecosystem with market intelligence that MAIRA cannot generate internally. In Singapore, the financial sector's AI governance frameworks were developed with extensive industry co-design. That model is replicable in Mauritius at lower cost, given the smaller number of significant AI deployers.

Finally, and most fundamentally: the measure of success is not the number of AI systems deployed or the GDP contribution of the digital sector. It is whether an elderly woman in Mahébourg whose social benefit application was rejected by an algorithm can find out why, challenge the decision, and get a fair hearing. It is whether a young entrepreneur from Rose-Hill can access a regulatory sandbox, test an agricultural AI product, and bring it to market without navigating a compliance burden designed for multinationals. It is whether Mauritius's AI governance model is cited in Kigali, Nairobi, and Dakar as proof that a small developing state can do this well.

The foundation has been laid through years of strategic planning. The Policy Blueprint and the Foundation Paper provide a technically rigorous, institutionally realistic, and values-coherent roadmap. What transforms a roadmap into a journey is the decision, taken by specific people in positions of authority, to walk it.

Interview by Nad SIVARAMEN

****** 

Biodata

I am a researcher in quantum AI with different doctorates and masters who is mainly involved in the application of AI in the banking sector. My specialized field is application of quantum AI, using quantum computers together with AI but has been heavily involved in design of AI systems in less sophisticated banks in credit scoring, money laundering and detection of fraudulent transactions with predictive AI. I run my own company quantumaiconsultancies and we have just launched our local implementation in Mauritius.» Born in Port-Louis (Ward IV), Dr Kwet Ng is 58 years old.

****** 

Key references

Primary Sources (Referenced Throughout)

• Ng, K. (2025). Toward a Mauritian AI Act 2025: Policy Blueprint for Responsible Innovation. Quantum AI Consultancies.

• Ng, K. (November 2025). Foundation for a Mauritian AI Act: Building on Global Standards for Island Innovation. Quantum AI Consultancies.

• European Parliament and Council. (2024). Regulation (EU) 2024/1689 — EU Artificial Intelligence Act.

• Government of Mauritius. (2017). Data Protection Act 2017.

• Government of Mauritius. (2021). Cybersecurity and Cybercrime Act 2021.

• Government of Mauritius. (2018). Digital Mauritius 2030 Strategic Plan.

• Ministry of Technology, Communication and Innovation. (2024). ICT Industry Blueprint.

• Treasury, Republic of Mauritius. (2025–2026). National Budget: Innovative Mauritius.

• African Union. (2020). Digital Transformation Strategy for Africa 2020–2030.

• OECD. (2019). OECD AI Principles.

• UNESCO. (2021). Recommendation on the Ethics of Artificial Intelligence.

• UK AI Safety Institute. (2024–2026). Technical Evaluation Methodologies for Advanced AI Systems.

• Singapore PDPC. (2020). Model AI Governance Framework (Second Edition).