Asher's Blog

I have been thinking deeply about artificial intelligence, and sometimes it feels like watching a science-fiction film while knowing that the film is slowly becoming ordinary life.

We are predicting a world in which AI will not merely help a human being design a car. AI may design the car, simulate it, select the materials, negotiate with suppliers, schedule production, inspect the factory, price the finished vehicle, plan the marketing campaign and study everything that happens after the customer begins driving it. It may analyse every defect and every complaint, then decide what should change in the next version. Robots may manufacture the components, autonomous vehicles may move them through the supply chain, and intelligent systems may coordinate finance, operations, sales and what we presently call human resources.

Perhaps, by then, we will call it AI resources or robot resources. I do not know.

But behind this apparently autonomous system, I still see a human decision. Somewhere, somebody decided what the machine should optimise. Somebody authorised its access. Somebody accepted the risk. Somebody established the boundary beyond which it should not go. And, at least for now, somebody must decide whether the machine’s recommendation becomes reality.

That is why I have arrived at this thought: the last human decision matters most.

The more decisions we delegate to machines, the more consequential the remaining human decisions become. If an AI system makes ten million small choices correctly but the final human gives it the wrong objective, the entire system may become extraordinarily efficient at producing the wrong outcome. If the final human approves a dangerous deployment, ignores a warning or chooses profit at the expense of safety, the intelligence of the system will not rescue us from the poverty of that decision. It may amplify it.

So who will that last human be? How will the person get there? What kind of education, character and intelligence will qualify somebody to exercise authority over a system that may reason faster and know more than any individual human being?

I do not have complete answers. This essay is an attempt to ask the questions before the machines make asking them feel unnecessary.

Intelligence and authority are not the same thing

The first mistake we may make is to assume that the most intelligent system must also become the final authority. We do not apply that principle consistently in human society. The best accountant does not automatically become the owner of the company. The most knowledgeable physician does not decide what every patient must do without consent. A brilliant engineer may understand a bridge better than a governor, but the engineer does not independently decide where public money should be spent. Knowledge informs authority; it does not by itself create legitimate authority.

AI may become far more capable than the person supervising it in many domains. That does not mean it should determine the purpose of a company, a school, a country or a human life. Capability answers, “What can be done?” Authority must still answer, “What should be done, for whom, at what cost, and who bears the consequences?”

These are not merely technical questions; they are moral and political questions. An autonomous factory may discover the cheapest way to increase output. But should it dismiss thousands of people in a community where the company is the primary employer? A medical system may identify the treatment that maximises the average survival rate. But how should it account for the dignity, wishes and unusual circumstances of an individual patient? An AI lender may reduce default by excluding a particular neighbourhood. But what if the historical data reflect decades of inequality? A military system may identify a target with high statistical confidence. Who should have the authority to use lethal force?

The machine can optimise only within an objective, whether that objective is explicitly programmed or indirectly learned. The human decision determines which objective deserves to govern.

This is close to the warning in Stuart Russell’s book Human Compatible: the central problem is not simply making machines intelligent; it is ensuring that powerful machines remain aligned with human purposes, especially when those purposes are uncertain, contested or badly expressed. A machine instructed to maximise one measurable outcome can neglect everything we failed to measure.

The last human may therefore not be the person pressing a dramatic red button. The last human may be the board that approved the objective, the regulator that permitted the system, the engineer who selected the data, or the executive who decided how much human review was too expensive. The decisive choice may occur years before the visible consequence.

Will only the highest IQs be allowed to decide?

One of the questions on my mind is whether this future will become a world in which only the most brilliant people can exercise meaningful authority. If you are supervising a superintelligent system, must you also be superintelligent? What happens to the average person?

My present view is that raw IQ will matter, but it cannot be the only thing that matters. No individual, however brilliant, will be able to reproduce everything a sufficiently advanced system knows. The human advantage cannot depend on winning a contest of memory, calculation or speed. We will lose that contest.

The more valuable capabilities may be asking the right question, understanding the context, recognising when an objective is morally incomplete, coordinating people with different expertise, and accepting responsibility for the outcome. A person can be intellectually gifted and still be reckless, corrupt, arrogant or easily manipulated. Intelligence without character may make the last decision more dangerous rather than safer.

We have already seen this in ordinary organisations. The smartest person in a meeting is not always the wisest leader. Leadership demands judgment: the ability to combine facts, values, timing, experience and consequences. It also demands humility—the ability to notice when you may be wrong and to invite challenge before a decision becomes irreversible.

This does not mean technical understanding is optional. A leader who knows nothing about the system may become a ceremonial approver, clicking “accept” because the machine appears confident. That is not human oversight. It is automated decision-making with a human signature attached.

The person with final authority must understand enough to interrogate the system. They must know what data it used, what uncertainty remains, what it was asked to optimise, what could fail, who could be harmed and whether an independent person can challenge the recommendation. They do not need to outperform the AI at every calculation. They need the intellectual and moral competence to govern its use.

The danger of becoming a rubber stamp

Human involvement alone does not guarantee human control. We have a tendency to trust recommendations that come from systems we perceive as sophisticated. This is often called automation bias. Once a machine produces a confident answer, people may stop searching for contrary evidence. They may assume that the system has considered variables they do not understand. Over time, human skill can also weaken because the machine performs the task so frequently that the human no longer has enough practice to intervene when it fails.

This creates a strange possibility: the organisation may claim that a human makes the final decision, while the human almost always accepts the machine’s recommendation. The human becomes legally responsible but intellectually absent.

Meaningful oversight must therefore include the ability to disagree. The human needs time, information and institutional permission to say no. If an employee is punished whenever they slow an automated process, the theoretical right to override it means very little. If the explanation is too complex to examine, or the system acts millions of times before anybody can intervene, we should not pretend that a human remains meaningfully in control.

This is why AI governance frameworks increasingly emphasise accountability, transparency and human oversight. The African Union’s Continental AI Strategy calls for an Africa-centred, development-oriented approach to AI rather than passive adoption of systems and priorities designed elsewhere. The International AI Safety Report similarly treats oversight and risk management as serious technical and institutional challenges, not as slogans that can be solved by placing a person at the end of an automated chain.

For Africa, this question is especially important. If the models, computing infrastructure, standards and objectives are controlled elsewhere, who is really the last human for decisions affecting African people? Is it an African doctor using a foreign medical system, or the company that trained the model? Is it a Nigerian bank deploying a credit algorithm, or the external provider that determines how the risk score works? Is it our elected government, or a platform whose infrastructure the government cannot inspect?

Human control is not meaningful if the relevant human has no practical power.

Africa must therefore do more than consume AI. We need technical talent, local data governed responsibly, research capacity, computing infrastructure and institutions capable of auditing consequential systems. We also need African founders building systems around African realities. Otherwise, the final decisions about our markets, languages, health, education and security may be made by people who do not live with their consequences.

What will education be for?

The question about the last human immediately becomes a question about children. What should schools teach when a machine can retrieve information, explain a concept, draft an essay, write code and solve many examination questions? Will schools continue to move average students through the system, or will the future demand that every child become extraordinarily capable from a young age?

I do not think the answer is to demand that every child become a genius in the narrow way we currently measure intelligence. Education should not become a brutal contest to identify the small number of people allowed to supervise machines. That would produce a technologically advanced but deeply unequal society.

The better answer is to raise the floor of human capability. AI can provide personalised explanations, additional practice and access to expertise that many African schools lack. It can help a teacher prepare material at different levels and support a student who needs more time. But access to an answer is not the same as understanding. If students outsource every difficult thought, their assignments may improve while their minds become weaker.

UNESCO’s guidance on generative AI in education argues for a human-centred approach and for developing human capacity, rather than allowing technology to determine education’s purpose. That is the right direction. A good education system should teach students how to reason, test claims, recognise uncertainty, work with other people, understand consequences and form values. It should teach them to use intelligent systems without surrendering their own agency.

We may need to assess students differently. If AI can produce an ordinary essay at home, the school must care more about whether the student can defend the argument, question its assumptions, apply it to an unfamiliar situation and explain why a particular conclusion should be trusted. Memorisation will not disappear, because knowledge stored in the mind is part of reasoning. But memorisation alone will become an increasingly poor signal of readiness.

The future student must learn three things at once: how to think without AI, how to think with AI and how to know when AI is weakening rather than strengthening thought.

That is a far more demanding educational mission than simply giving every child a tablet.

The final decision is often a chain

There may also be no single “last human.” Important decisions are usually distributed through a chain of people and institutions.

An engineer chooses the model; a product manager defines the use case. A compliance team approves the process; an executive accepts the commercial risk. A board decides the governance structure; a regulator establishes the legal boundary. A customer chooses whether to use the product. Each person makes a decision that limits or expands what the next person can do.

If everybody assumes that somebody else is responsible, the system becomes powerful while responsibility disappears.

We should therefore design decision rights before the crisis. Which actions may the system take alone? Which require human approval? Which are prohibited entirely? At what level of uncertainty must the system stop? Who can override it? Who audits the outcome? Who is informed when it fails? Who compensates the person harmed?

These questions sound procedural, but procedure is how values survive contact with scale.

In a company, I would want the amount of human control to rise with the possible harm. An AI drafting an internal meeting summary may require ordinary review. An AI recommending that an employee be dismissed, a customer be denied credit or a suspicious transaction be frozen requires deeper evidence, an appeal route and accountable human judgment. A system affecting health, physical safety or liberty must meet a much higher standard.

The last human decision should not be improvised at the moment of danger.

Perhaps character will become the scarce intelligence

If machines become better at analysis, prediction and execution, human character may become more—not less—important.

Who do we trust to define the objective? Who can resist the pressure to deploy a system before it is safe? Who will admit that the machine’s answer is uncertain when investors, voters or customers want certainty? Who will protect a minority when the average outcome looks good? Who will accept accountability rather than blame the algorithm?

These are questions of courage, integrity, restraint and love for other human beings.

My Christian faith makes this distinction important to me. Intelligence is a gift, but it is not God. Knowledge is powerful, but wisdom asks what knowledge is for. A machine may become astonishingly capable without possessing the moral responsibility of a person made to care for other people. Human beings must not worship what we have made with our own hands, especially when what we have made begins to speak more confidently than we do.

The last human must remember that efficiency is not the highest good. Human dignity matters. Mercy matters. Justice matters. Freedom matters. The person affected by a decision is not simply a data point inside an optimisation problem.

Preparing to become the last human

I do not know exactly how far AI will go or how quickly it will get there. Nobody does. The 2025 Stanford AI Index documented rapid improvements across demanding benchmarks, but benchmark performance is not the same as dependable autonomy in the untidy physical and institutional world. Forecasting should therefore contain both imagination and humility.

Still, we can begin preparing. We should learn how these systems work well enough to question them. We should develop expertise in a real domain, because good judgment needs contact with reality. We should practise making decisions under uncertainty and recording why we made them. We should surround important decisions with people who can challenge us. We should study history, ethics, science and human behaviour—not only prompting and software. We should build organisations in which dissent is safe and responsibility is clear.

And we should teach children that their future value will not come from competing with a machine at being a machine. Their value will come from becoming more fully capable human beings: curious, technically literate, morally grounded, creative, courageous and able to cooperate with others.

Perhaps only a small number of people will make the most consequential decisions about advanced AI systems. If so, society must ask how those people are selected and who holds them accountable. Wealth, technical brilliance or ownership of computing infrastructure cannot be sufficient qualification to decide humanity’s future. The people affected must retain a voice.

The last human decision matters most because it gives direction to everything the machine can do. As machine capability grows, a careless human choice can travel farther, faster and with less opportunity for correction. But the same leverage can work for good. A wise objective, a just boundary and the courage to stop can protect millions of people.

The central question of the AI age may not be whether machines will become more intelligent than us. It may be whether human beings will become wise enough to remain responsible for what our intelligence has created.

I know only this: somewhere behind the machine is a human choice. We must pay attention to who makes it, what formed that person, whose interests they serve, and whether they still possess the freedom to say no.

Because when almost everything else has been automated, the last human decision may be the decision that defines us all.