Geoffrey Hinton’s warning about artificial intelligence should neither be dismissed as technological alarmism nor accepted as prophecy. The deeper question is whether human institutions can keep pace with a technology whose capabilities, incentives and consequences are advancing faster than the rules designed to govern them.
By John Onyeukwu |
Humanity is building machines that may eventually become more intelligent and capable than their creators. The uncomfortable question is not simply whether that will happen. It is whether the institutions governing the race to get there will remain capable of managing the consequences when it does.
That question has acquired new urgency. Geoffrey Hinton, whose foundational work on neural networks helped make today’s AI revolution possible, has again warned that humanity should be deeply concerned about increasingly capable artificial intelligence. In a recent interview with ABC, Hinton backed calls for a slowdown in AI development, saying most experts believe systems smarter than humans could arrive within a decade. He also acknowledged a critical limitation: nobody knows how to calculate a reliable probability of an AI catastrophe.
That distinction matters. AI could contribute to catastrophic outcomes, including human extinction. That is a serious risk scenario. But it is not an established prediction that AI will destroy humanity within the decade.
The rational response to that uncertainty is neither panic nor complacency. It is to ask a more fundamental question: what kind of institutions should govern a technology when its potential consequences are enormous but our knowledge of those consequences remains incomplete?
Start with first principles
The first principle is simple: uncertainty does not eliminate responsibility. Governments routinely regulate technologies whose risks cannot be predicted with precision. They do so not because every possible harm is known, but because the potential consequences justify testing, monitoring, accountability and safeguards proportionate to the risk. AI should be approached in the same way.
The debate becomes clearer if we separate three categories of risk. The first is present harm: fraud, impersonation, deepfakes, cybercrime, privacy violations, misinformation and unreliable automated decisions. These are observable problems, not hypothetical futures. The second is systemic risk: the effects of AI when it becomes deeply embedded in labour markets, financial systems, public administration, critical infrastructure and information ecosystems. Here, failure can propagate beyond an individual user or organisation. The third is frontier or existential risk: scenarios involving systems sufficiently capable and autonomous to evade meaningful human control, facilitate catastrophic misuse or create irreversible consequences.
The 2026 International AI Safety Report treats loss of control as a hypothetical but serious risk. It also makes an important distinction: current systems do not possess the capabilities required for such scenarios, although researchers are observing progress in areas such as autonomous operation, planning and the ability to exploit weaknesses in evaluation systems.
These risks should not be conflated. Nor should governments wait to resolve the third before addressing the first two.
The evidence is serious but not conclusive
The concern about frontier AI is not simply science fiction. A major survey of 2,778 AI researchers found that the median respondent assigned at least a 5 percent probability to AI causing human extinction or similarly severe and permanent disempowerment, while between roughly one-third and one-half of respondents assigned at least a 10 percent probability, depending on the question. At the same time, many of those same researchers believed advanced AI was more likely to produce good outcomes than bad ones.
That combination is important. It tells us that concern about catastrophic risk does not necessarily imply opposition to AI. It also demonstrates why probability claims should be handled carefully. Expert estimates about unprecedented events are judgments under radical uncertainty, not statistical forecasts derived from a long historical record. There is no dataset from which anyone can calculate a scientifically established probability that superintelligent AI will eliminate humanity by a particular date.
Hinton’s warning should therefore be taken seriously without being converted into a countdown to extinction.
The stronger case for action does not depend on knowing the probability of extinction. There are already sufficient reasons to improve AI governance.
The race creates a governance problem
This is where the issue becomes less about technology and more about political economy. AI developers have strong incentives to increase capability. Companies compete for market share. Investors expect growth. Researchers want to advance the frontier. Governments want economic and strategic advantage. Military establishments see national-security applications. Consumers want better systems at lower cost.
The incentives point in roughly the same direction: build more and deploy faster.
But the benefits of moving quickly are immediate and concentrated, while some risks may be delayed, diffuse and borne by people who had no role in creating them. Stanford’s 2026 AI Index reports that global corporate AI investment more than doubled in 2025, with private investment rising 127.5 percent and generative-AI investment increasing by more than 200 percent.
This creates a classic collective-action problem. A company that slows development to strengthen safety may fear that a competitor will move ahead. A government that imposes stricter controls may fear that another country will capture the economic or strategic advantage.
Each actor can therefore behave rationally from its own perspective while collectively producing a race that nobody would necessarily choose if all participants could coordinate.
The critical question is not simply who can build AI fastest. It is who has the incentive and authority to slow down when slowing down is in the collective interest.
That is a governance question.
The geopolitical dimension makes it harder
The race is no longer only between technology companies. It is increasingly a competition among states for computing capacity, advanced chips, data, talent, capital and strategic advantage.
Stanford estimates that US private AI investment reached $285.9 billion in 2025, compared with $12.4 billion in China’s private investment, while cautioning that China’s state-linked funding means private investment figures do not capture the full picture.
This creates a difficult paradox. A government may believe that advanced AI carries serious risks while also believing that slowing its own development could weaken its position relative to geopolitical competitors. That makes unilateral restraint difficult.
Any credible international AI regime must therefore address incentives as well as safety. Rules that impose costs on one jurisdiction while leaving competitors unconstrained may simply relocate development rather than reduce risk.
AI governance is consequently becoming a question of international statecraft.
Regulation should follow risk, not rhetoric
There is little value in framing the debate as regulation versus innovation. Mature societies do not govern aviation, pharmaceuticals, banking or nuclear technology by banning innovation. Nor do they leave high-consequence industries entirely to voluntary restraint. They establish rules according to risk.
AI should be treated similarly. A system used to draft routine correspondence does not present the same governance challenge as an autonomous system controlling critical infrastructure, conducting cyber operations or making high-stakes decisions about citizens.
The International AI Safety Report notes persistent governance difficulties, including information asymmetry between companies and governments, uncertain liability, commercial pressure for rapid deployment and limitations in current safety evaluation.
The implication is straightforward. Voluntary responsibility can complement regulation, but it cannot replace independent accountability where the consequences are systemic. High-risk systems require stronger evaluation, incident reporting, cybersecurity, auditability, clear liability and regulatory capacity to intervene when necessary.
The accountability gap
Perhaps the most underappreciated AI governance problem is not whether machines become smarter than humans. It is whether responsibility becomes harder to locate. Consider an AI-assisted government decision: who receives a public benefit, which taxpayer is subjected to additional scrutiny, which application is rejected or which individual is flagged for investigation. In a conventional administrative system, there is usually a recognisable chain of authority. AI can fragment it.
Responsibility may be distributed among the public official, agency, software vendor, model developer, data provider and system operator. When something goes wrong, everyone can potentially point elsewhere.
That creates a fundamental governance risk: a system can produce a decision without producing an accountable decision-maker.
That principle should be non-negotiable. AI may assist public authority; it should not dissolve public accountability. Citizens must retain the ability to understand how consequential decisions were made, challenge them, obtain meaningful human review and identify the institution legally responsible for correcting harm.
The economic disruption is already more tangible
The extinction debate can also obscure a more immediate transformation: the changing economics of work.
The evidence does not support the simplistic claim that AI will simply eliminate jobs. AI is already increasing productivity in some settings, while the International AI Safety Report expects general-purpose AI to automate a wide range of cognitive tasks, particularly in knowledge work.
The more consequential question is distribution. If AI increases productivity while concentrating income, market power and ownership of productive assets, the social consequences could be significant even if aggregate economic output rises. If AI reduces the number of entry-level opportunities through which young people acquire professional experience, the effects could extend beyond employment into social mobility.
For countries with large young populations, this matters enormously.
The AI policy question is therefore not simply how much wealth AI will create. It is who captures the gains, who bears the transition costs and which institutions determine the distribution. That is political economy.
Nigeria cannot afford to be only a consumer of AI
For Nigeria, the AI debate should not begin and end with whether the country should “embrace AI.” That question has effectively been answered by economic reality. Nigeria will use AI. The more important question is whether the country will merely consume technologies designed elsewhere or develop the institutional, technical and economic capacity to shape how those technologies are deployed.
Nigeria is not starting from zero. Its National Artificial Intelligence Strategy is intended to harness AI for sustainable development, innovation, national productivity and human well-being, building on the National Centre for Artificial Intelligence and Robotics. NCAIR’s stated priorities include research, adoption, entrepreneurship and economic growth, while it is already developing Nigerian-focused AI infrastructure and multilingual models.
The policy problem, therefore, is not the absence of ideas. It is implementation capacity.
Nigeria needs a risk-based regulatory architecture with clear institutional responsibility. Not every AI application should face the same requirements. Systems used for marketing or routine administrative tasks are different from systems used in taxation, credit, healthcare, employment, policing or social protection. High-impact applications should be subject to stronger requirements for testing, documentation, human oversight, incident reporting and independent review. The relationship among NITDA, the Nigeria Data Protection Commission and sector regulators must also be sufficiently clear to prevent regulatory gaps and institutional overlap.
Government procurement should become another major instrument of AI governance. When public institutions purchase AI-enabled systems, they should know what the technology does, what data it uses, where that data is stored, how performance is tested, what its limitations are, who can access the system and who carries liability when it fails. AI procurement cannot remain merely an IT transaction. Governments that purchases technological opacity may discover the accountability problem only after citizens are harmed.
Nigeria also needs stronger national data architecture. AI capability depends on the quality, availability and governance of data, but data accumulation without safeguards can create new vulnerabilities. The Nigeria Data Protection Act already gives data subjects a right not to be subjected to decisions based solely on automated processing that produce legal or similarly significant effects, subject to specified exceptions and safeguards. The challenge is to make such protections operational across government and the private sector, particularly where AI systems are deployed at scale.
Technical capacity must be accompanied by regulatory and judicial capacity. A regulator cannot effectively supervise a technology it cannot interrogate. Nigeria therefore needs AI expertise not only among engineers, but also within regulatory agencies, the judiciary, legislature, audit institutions, procurement authorities and the wider civil service. The objective is not to turn every public official into a computer scientist. It is to prevent public authority from becoming technically dependent on the companies it is supposed to regulate.
Nigeria also needs an AI transition strategy for work and skills. Generic calls for “digital skills” are insufficient. The country should identify occupations and sectors likely to experience significant task displacement, determine where AI can raise productivity and redesign education and vocational training around skills that complement increasingly capable machines. Particular attention should be paid to entry-level professional work. If AI removes the junior tasks through which young lawyers, accountants, analysts, programmers and researchers acquire experience, Nigeria could face a less visible but significant social-mobility problem.
Finally, Nigeria needs to think seriously about digital sovereignty without confusing sovereignty with technological isolation. The country will continue to depend on foreign cloud infrastructure, advanced chips and foundation models. That is not inherently a weakness. The vulnerability arises when dependence becomes so deep that Nigeria lacks the technical knowledge, bargaining power or alternatives necessary to protect its interests. Digital sovereignty should therefore mean the capacity to understand, negotiate, regulate and diversify critical technological dependencies while using partnerships to build domestic capability.
The same logic applies to synthetic information. Generative AI is reducing the cost of producing convincing text, audio, images and video. In Nigeria, this has implications for elections, public trust, journalism, business and institutional credibility. A fabricated recording of a public official or altered video can circulate faster than the institutions capable of disproving it. The response should not be blanket censorship. It should combine provenance, verification, media literacy, rapid institutional communication and accountability for malicious synthetic content while preserving legitimate expression and due process.
The governing question for Nigeria is therefore not whether it can become an AI superpower. The more useful question is whether Nigeria can become institutionally competent in an AI-powered world. Can the state understand the technologies it procures? Can regulators supervise the companies deploying them? Can courts provide remedies when automated systems cause harm? Can Nigerian firms capture enough value from AI to improve productivity rather than simply pay for imported intelligence? And can citizens retain meaningful rights when important decisions increasingly involve machines?
Those are questions of state capacity.
The real race
The AI race is usually described as a contest among companies or between major powers. There is another race that matters more for the long term: the race between AI capability and institutional capability.
AI systems can improve in months. Legislation can take years. Models can be updated globally while regulators remain largely national. Developers possess specialised technical knowledge that most governments lack, while public institutions must operate through budgets, procurement rules, evidence, due process and democratic accountability.
These are not reasons to abandon democratic governance. They are reasons to make it more adaptive. Government should not become faster by becoming less accountable. The objective is to build institutions capable of understanding rapidly changing technology while preserving the principles that make public power legitimate.
For Nigeria, that means moving beyond the familiar cycle of announcing a strategy, launching a programme and measuring success by activity. AI governance should be judged by whether institutions can prevent foreseeable harm; respond when systems fail, protect citizens’ rights, improve public-sector performance and ensure that technological adoption contributes to national productive capacity.
The country already has the beginnings of such architecture. The harder task is making the institutions within it capable of implementation, coordination and enforcement.
The choice is not fear or faith
The AI debate is increasingly presented as a choice between those who believe AI will transform civilization for the better and those who believe it may destroy humanity. That is too crude for a technology of this consequence.
AI is neither a god nor a demon. It is a powerful general-purpose technology being developed and deployed through human institutions. Its consequences will depend not only on what machines can do, but on who controls them, what incentives shape their development, where they are deployed and whether institutions remain capable of holding people accountable for their use.
Geoffrey Hinton’s warning deserves to be heard precisely because it comes from someone who understands the technology from the inside. But his warning should not be converted into a countdown to extinction. We do not need to prove that AI will destroy humanity before demanding safeguards. We do not need to settle the probability of superintelligence before addressing harms already occurring. And we do not need to choose between innovation and governance.
We need institutions capable of doing both: capturing the benefits of technological progress while governing the risks it creates.
For Nigeria, that means turning AI policy from a technology agenda into a state-capacity agenda: building the ability to regulate, procure, audit, litigate, negotiate, innovate and distribute the gains of technological change.
Humanity’s advantage over artificial intelligence may ultimately not be that humans are smarter. It is that humans can create rules, institutions and systems of accountability.
The challenge is that AI may be advancing faster than those institutions can adapt.
The defining question of the AI age may therefore not be whether machines become more intelligent than humans. It may be whether human governance can remain intelligent enough to govern the machines we build.
For Nigeria, that is not an argument for fear. It is an argument for institutional preparedness, technological sovereignty and statecraft.
John Onyeukwu is a Lawyer and Governance & Social Impact Practitioner based in Abuja.
Email: john@impactbridgeafrica.com
