By Dr Rais Hussin
On Sept 29, OpenAI unveiled “always-on” autonomous agents designed not merely to answer prompts but to pursue goals across applications. With that, artificial intelligence (AI) has crossed from a technology of assistance and content creation into an infrastructure of delegated agency, and the world is scrambling to decide how to regulate its risks, constrain its excesses and assign responsibility for its decisions.
However, while many institutions still approach governance largely as a software issue, the challenge reaches much deeper, into the political economy that determines who owns the models, commands the infrastructure, allocates the capital, sets standards, retains expertise and, ultimately, preserves human agency.
One thing is becoming increasingly clear: We are regulating AI at the wrong altitude.
Until recently, a country could reasonably imagine AI governance as a legal exercise: Parliament legislates, regulators impose disclosure requirements, agencies develop ethical standards, procurement rules are tightened and companies are penalised for violations.
However, that model assumes something significant without usually saying so: the regulator has sufficient visibility into, access to and leverage over what it regulates.
With conventional industries, that assumption is broadly valid. A government can inspect a factory. A banking regulator can demand records from a licensed bank. A medicines regulator can require clinical evidence before approving a drug.
AI complicates this enormously. When consequential parts of a system consist of foreign-owned models running on foreign-controlled clouds, trained through proprietary processes on inaccessible datasets, dependent on imported accelerators, updated by firms outside the jurisdiction and increasingly capable of autonomous action, legal authority may exist without equivalent technical agency.
A great gap therefore lies between jurisdictional sovereignty and effective sovereignty. A state can retain the first while steadily losing the second. And AI may become one of the clearest cases yet.
AI governance is moving through three layers, with the first being output governance: Misinformation, hallucination, discrimination, harmful content, intellectual property, deepfakes. Second is system governance: Model evaluation, red-teaming, accountability, auditability, human oversight and agent permissions. Third, and increasingly consequential, is infrastructural governance: Compute, chips, cloud access, model weights, training processes, energy, data provenance, cybersecurity, supply chains and the technical means to inspect or intervene in advanced systems.
The earlier debate concentrated heavily on the first layer. Attention is now moving upstream towards the second and third, where sovereignty over the technology stack matters far more because the capabilities required for regulation cannot simply be legislated into existence.
A country can prohibit an output without owning a model, or write an ethical principle without manufacturing a chip. But once governance requires inspecting training processes, auditing frontier systems, tracing computational dependencies, securing sensitive datasets or controlling autonomous permissions, technical capability itself becomes regulatory capacity.
That creates a second-order problem receiving far too little attention: AI governance risks becoming recursively unequal.
The countries and firms that accumulated AI capacity first are increasingly those with the technical knowledge, compute access, model visibility and institutional expertise needed to define what “safe”, “auditable” or “responsible” AI means in practice.
Technological advantage thus converts into governance advantage, and the two reinforce one another. Infrastructure produces expertise; expertise shapes standards; standards influence market access; market access strengthens incumbents, which accumulate still more infrastructure and knowledge. The feedback loop reaches beyond the familiar AI divide. It becomes a pure sovereignty flywheel.
Dependency reproduces itself just as effectively in reverse: foreign infrastructure constrains domestic expertise, weak expertise limits inspection capacity, regulators become reliant on vendor assurances, and standards are increasingly shaped elsewhere.
This is precisely why EMIR Research’s earlier warning on vendor lock-in carried implications beyond procurement efficiency. In “Sovereign AI, not a GPU Arms Race”, we argued that the strategic question was who could reduce single points of failure and avoid dependence on proprietary providers. We can now see the further implication: vendor lock-in can become regulatory lock-in, because systems we cannot adequately inspect, interrogate or substitute inevitably narrow the space for meaningful oversight.
That matters acutely for Malaysia. Hosting intelligence is not the same as owning capability. The country is attracting major foreign direct investment into data centres and computing infrastructure, but AI’s maturation forces a harder question: what, precisely, does that investment leave behind? Megawatts of compute on Malaysian soil do not automatically translate into Malaysian technological sovereignty.
A data centre becomes sovereign infrastructure only to the degree that sovereignty travels through it. Which models can we run? Who controls access? Can Malaysian researchers use the compute and regulators inspect relevant systems? Is technical knowledge transferred? Are local firms moving upstream? Do we control strategically important datasets, retain the ability to switch providers and keep essential systems operating during geopolitical disruption?
This is why EMIR Research has consistently advocated wider application of the Input-Output-Outcome-Impact framework, including in “AI in Malaysia: Caution Without Strategy is Its Own Mistake”. The meaningful measure is not how many billions arrive, but what strategic capability they create. FDI is the input. Data centres are the output. Sovereign technological capacity is the outcome. Strategic autonomy and regulatory sovereignty are the impact.
Yet sovereignty does not reside in servers alone. A country also needs people capable of understanding the systems being governed. If Malaysian regulators depend on vendors to explain whether a model is safe, the informational asymmetry is already severe: the regulated entity understands the system; the regulator understands only what the regulated entity discloses. That is not robust governance.
Sovereignty therefore has several intertwined dimensions: access to compute and infrastructure, control over strategically important data, the capacity to build, modify, inspect and substitute systems, and, crucially, enough domestic expertise to determine whether claims made by vendors and foreign institutions are actually true.
This connects directly to EMIR Research’s repeated concern over brain drain. Talent loss already erodes innovation capacity and sovereignty. In the AI era, it can become a governance vulnerability. Without engineers, evaluators, security researchers and technically sophisticated regulators who understand frontier systems, Malaysia may end up regulating technologies it cannot independently interrogate.
Autonomous agents make the problem sharper still. They do not merely advise; increasingly, they execute. Here even the familiar safeguard of “human-in-the-loop” begins to weaken: human oversight means little if the human no longer understands the loop. A nominal approver does not preserve agency if that person cannot understand the system, challenge its recommendation or meaningfully reverse its action.
We are therefore confronting the same question at two scales: can a person retain meaningful agency over an autonomous AI system, and can a state retain meaningful agency over an AI ecosystem?
If Malaysia wishes AI to reflect social priorities and public values, including justice, compassion, fairness and human dignity, as Prime Minister Datuk Seri Anwar Ibrahim has previously emphasised, sufficient control over the systems themselves becomes indispensable.
This leads inevitably to geopolitics. In “A Critical Perspective on the UN’s AI Governance Plan” EMIR Research previously warned that ostensibly universal AI-governance frameworks could reproduce existing geopolitical asymmetries. Now the concern runs deeper: AI governance itself may become an instrument of geopolitical power, as technical standards increasingly determine access to markets, compute, models and infrastructure.
For Malaysia, active neutrality therefore acquires a new technological meaning. It should mature into something closer to interoperable sovereignty: neither technological isolation nor structural dependence on a single ecosystem, but sufficient domestic capability to connect across technological blocs without becoming captive to any one of them.
Interoperable sovereignty is not autarky, nor choosing US over China, or China over US. Neither is it blind faith in open source. It means preserving the ability to choose, combine, inspect, modify and replace technologies across ecosystems. For a middle power, that is meaningful technological autonomy.
At the end, we arrive at the deepest paradox, though perhaps it was visible from the beginning. The AI race was never simply a contest over how intelligent our machines could become. The deeper test has always been whether the institutions and societies surrounding them retain sufficient intelligence, capability and sovereignty to determine what those machines are permitted to do.
Malaysia does not need to win a race for the largest model or the greatest number of GPUs. But it does need sufficient command over the systems entering its economy and institutions to remain the author of its own choices. In the age of delegated machine agency, sovereignty may ultimately mean something very simple: retaining the practical ability to say yes, to say no and, more importantly, to make either decision consequential.
The author is the President/CEO of EMIR Research, a think tank focused on strategic policy recommendations based on rigorous research.
