The Skills AI Can Master Are the Ones We Know How to Grade

Jensen Huang's warning about intelligence, the limits of standardised testing, and why the future may belong to people who understand what cannot easily be measured.

There is a surprisingly simple way to understand why artificial intelligence is advancing so rapidly in some areas while struggling to demonstrate comparable mastery in others.

Ask one question: How quickly can you determine whether the answer is right?

If the answer can be checked immediately, repeatedly and cheaply, artificial intelligence has an enormous advantage. If success depends on human relationships, uncertain circumstances, long-term consequences and information nobody can fully measure, the problem becomes much harder.

This distinction helps explain an observation made by Jensen Huang, the chief executive of Nvidia, the company whose processors have become central to the modern AI industry.

When asked to identify the smartest person he had ever met, Huang resisted the conventional definition of intelligence. He argued that technical problem-solving, while important, was becoming something machines could increasingly perform. He placed greater emphasis on empathy, wisdom, the ability to understand what people leave unsaid and the capacity to anticipate problems before they become obvious.

His point was not that academic intelligence has become worthless. It was that our traditional measures of intelligence may capture precisely the capabilities machines are learning to reproduce most efficiently.

And that should change how we think about education, careers, leadership and the future of work.

The answer key is AI's secret advantage

Consider a student preparing for the SAT.

The examination consists largely of questions with defined answers. A student can complete a practice test, compare the results with an answer key, identify mistakes and repeat the process.

Now imagine a student who can take millions of practice tests, receive immediate feedback on every answer and adjust their approach after each attempt.

That student would have an extraordinary advantage.

AI systems can operate in a similar way when they are trained on problems that have reliable, automatically verifiable solutions.

In mathematics, a system can generate a solution and check whether the result is correct. In programming, software can run generated code against test cases to determine whether it behaves as expected. In certain scientific and engineering tasks, simulations can provide feedback about whether a proposed solution works.

The feedback loop is powerful:

1. The system attempts a problem.
2. An evaluator checks the result.
3. The training process rewards successful behaviour or penalises errors.
4. The system repeats the process and improves its performance.

Modern AI training is more complicated than this simplified description. It can involve pretraining on large datasets, human feedback, reinforcement learning and other techniques. Not every useful answer has an objective score, and not every model learns simply by repeating automatically graded exercises.

Nevertheless, the principle is important: when reliable feedback is available at scale, developers can create powerful incentives for AI systems to improve.

The evaluator does not get tired. It can assess huge numbers of attempts without the costs associated with a human teacher checking every response.

The system can explore approaches that a human student might never have the time or patience to test.

This is one reason AI has made striking progress in mathematics, programming and other fields where performance can be evaluated against clear criteria.

Why mathematics and programming fell first

In July 2025, AI systems from OpenAI and Google DeepMind achieved gold-medal-level performance on problems from the International Mathematical Olympiad.

Two months later, OpenAI reported that its reasoning system solved all 12 problems in an AI evaluation associated with the International Collegiate Programming Contest World Finals. The best human team solved 11.

These were remarkable achievements. They demonstrated that AI could tackle demanding problems requiring multiple stages of reasoning, not merely retrieve familiar facts or complete simple calculations.

But they also illustrate why these particular fields are suitable for machine improvement.

A mathematical proof can be examined for logical validity. A numerical result can be checked. A programming solution can be executed against test cases. In competitive programming, a solution that fails an essential test does not become correct because its author is persuasive or well-intentioned.

There are complications, of course. A program can pass a limited test suite and still contain bugs. A mathematical answer can look plausible while resting on faulty reasoning. And success in a competition does not automatically translate into reliable performance on every real-world engineering problem.

Still, these fields provide something invaluable: a relatively clear mechanism for distinguishing successful attempts from unsuccessful ones.

Now compare that with managing a company.

A chief executive may decide to delay a product launch because an employee seems unusually reluctant to approve it. The employee might never explicitly explain their concern. Three months later, the company discovers a security vulnerability that would have caused a serious failure.

Was the executive right to delay the launch?

Possibly. But perhaps the delay cost the company a major commercial opportunity. Perhaps the employee was simply exhausted. Perhaps the vulnerability would never have been exploited.

The outcome does not provide a clean answer.

Human judgment frequently operates under these conditions. We must make decisions with incomplete information, conflicting objectives and consequences that may not become clear until much later.

That makes the problem considerably harder to train against a reliable score.

Moravec's paradox: why difficult things can be easy for machines

This problem has a name.

In the 1980s, robotics researchers, including Hans Moravec, drew attention to a striking asymmetry between human and machine capabilities. Tasks that appear intellectually demanding to humans can be relatively straightforward for computers, while abilities that humans acquire naturally can be extraordinarily difficult to reproduce in machines.

The observation became known as Moravec's paradox.

A computer can calculate enormous quantities of numbers faster than any person. Yet reproducing the effortless coordination involved in walking through a cluttered room, recognising subtle physical cues and manipulating unfamiliar objects has historically been much more difficult.

Human beings do not need formal lessons to develop many of these abilities. They emerge through years of perception, movement, interaction and adaptation.

We rarely appreciate how much intelligence is involved in ordinary activities because they feel natural.

Recognising that a colleague is upset despite their smile, sensing that a negotiation is deteriorating or noticing that a customer has lost confidence can involve a sophisticated combination of memory, context, perception and inference.

These are not necessarily mysterious abilities. But they are difficult to reduce to a collection of simple rules.

Huang's argument applies a related insight to the age of generative AI.

If computers can increasingly master the kinds of intelligence measured by standardised tests, the skills that distinguish exceptional human judgment may be those our examinations have historically struggled to measure.

Try building an answer key for empathy

Imagine training an AI system to identify when a customer is about to abandon a business relationship.

The system receives emails, customer-service conversations, purchasing histories and meeting transcripts. It must predict which customers are losing confidence before they cancel their contracts.

At first, this appears manageable. The company can record whether a customer eventually leaves and reward the model for making accurate predictions.

But the problem becomes more complicated when we examine it closely.

Some customers leave for reasons entirely unrelated to their relationship with the company. Others complain loudly but remain loyal for years. Some customers quietly lose confidence and disappear without warning. Others threaten to leave as a negotiating tactic.

A model that learns to treat every complaint as evidence of impending departure will make costly mistakes.

Worse, a model optimised to reduce cancellations might recommend aggressive retention tactics that make customers feel manipulated. It could improve its numerical performance while damaging the trust the business depends on.

The same problem appears in leadership.

How do you objectively measure whether a manager correctly sensed that an employee was becoming disengaged? How do you distinguish a wise decision from a lucky guess? How do you reward someone for preventing a crisis that never occurred?

Prevention creates a particularly difficult measurement problem: when the intervention works, the negative outcome may never happen, leaving no direct evidence of what would otherwise have occurred.

These are examples of the challenges researchers sometimes describe as problems of reward design and measurement.

The difficulty is not that human behaviour can never be studied scientifically. Organisations can collect feedback, measure retention, examine outcomes and evaluate decisions. AI systems can also learn from simulations, human preferences and indirect signals.

The difficulty is that these measurements are often incomplete, delayed or influenced by circumstances outside the decision-maker's control.

A reliable answer key is not simply a list of correct answers. It requires a trustworthy definition of what success actually means.

That is where many real-world decisions become difficult.

But does this mean AI cannot understand people?

No. That would be the wrong conclusion.

AI systems can already analyse language, recognise emotional cues in some settings, summarise conversations, identify patterns in behaviour and generate advice about interpersonal situations.

They can help customer-service teams identify recurring complaints. They can assist managers in examining employee feedback. They can help doctors communicate complicated information in clearer language and help negotiators prepare for difficult conversations.

As these systems improve, some tasks once associated with empathy and social intelligence will become increasingly automated.

It is also possible to train AI models using human evaluations, simulations, behavioural data and outcomes observed over time. Researchers do not need a perfect answer key to make progress.

But there is a crucial distinction between recognising patterns associated with human emotions and reliably understanding what a particular person is experiencing in a particular situation.

A person may say they are fine while signalling distress through their tone, behaviour or silence. The same silence can mean discomfort, concentration, disagreement or simply a desire not to speak.

Context matters.

The challenge is not merely to detect a signal. It is to determine what that signal means, how confident the interpretation should be and what response is appropriate.

Even humans get this wrong.

The goal, therefore, should not be to assume that empathy is permanently beyond AI. It should be to recognise that social judgment involves uncertainty, conflicting values and consequences that cannot always be reduced to a single score.

AI may become increasingly capable in these areas. The important question is whether its recommendations remain reliable when the situation is unfamiliar, the evidence is ambiguous and the cost of a mistake is high.

The danger of confusing examination performance with intelligence

The SAT example raises a deeper problem about education.

Standardised examinations are useful because they provide a relatively consistent way to assess certain academic abilities. They can help universities compare applicants and identify students who have mastered particular material.

But no examination captures the full range of human intelligence.

A person can perform exceptionally well in mathematics and still struggle to lead a team. Another can be a gifted engineer but make poor decisions under pressure. Someone else may have modest examination results yet possess extraordinary practical judgment, emotional awareness or entrepreneurial ability.

The problem arises when society treats what can be measured as the only thing that matters.

Once a measurable score becomes the primary objective, institutions naturally concentrate on improving it. Students prepare for examinations. Schools track results. Employers use credentials as shortcuts for evaluating applicants.

AI intensifies this problem because machines can become extremely effective at optimising measurable performance.

If an examination rewards the ability to solve a defined set of problems, AI developers can train systems to become better at those problems. The examination may remain useful, but its ability to distinguish human capability from machine capability changes.

That does not make academic achievement meaningless. It means we must be more careful about what academic achievement actually demonstrates.

Education should teach students to solve problems, but it should also develop curiosity, communication, ethical reasoning, collaboration and the ability to make decisions when the available information is incomplete.

These capabilities should not be romanticised as exclusively human. They should be cultivated because they are valuable, regardless of whether machines eventually acquire some of them.

The future of work belongs to people who can define the problem

One implication of this shift is that the ability to formulate a good question may become more valuable relative to the ability to execute a routine answer.

Consider a software developer.

In the past, a developer might spend considerable time writing code, debugging errors and implementing features. AI can increasingly assist with those tasks.

But the developer still needs to determine what the application should accomplish, what users actually need, which security risks matter and how the finished product should behave when something goes wrong.

Generating code is not the same as deciding which code should be generated.

The same principle applies to business, journalism, medicine, education and public administration.

A marketing team can use AI to generate hundreds of campaign ideas. It still needs to understand its audience, identify the right message and decide which claims are ethical and credible.

A journalist can use AI to summarise documents. The journalist must still establish whether the information is trustworthy, identify what is missing and decide which facts are genuinely important to the public.

A doctor can use AI to help interpret medical information. Clinical decisions still require attention to the patient's circumstances, the limitations of the evidence and the consequences of treatment choices.

In each case, AI can reduce the cost of execution. The human contribution may shift towards defining objectives, interpreting context, evaluating trade-offs and taking responsibility for decisions.

This is not a guarantee that these jobs will remain unchanged. Some roles will shrink, some will disappear and others will emerge.

But it suggests that workers should not measure their future value solely by how quickly they can perform tasks that AI can increasingly automate.

They should also ask whether they understand the problem well enough to know what ought to be done in the first place.

What this means for Nigeria and Africa

For young Nigerians entering an increasingly competitive labour market, the implications are significant.

Learning to use AI tools is important. So are technical skills, analytical ability and the capacity to produce useful work efficiently.

But simply learning to operate a particular AI application may not provide a durable advantage. Interfaces change, tools become easier to use and competitors gain access to similar capabilities.

A more resilient approach combines technical competence with deep understanding of a real problem.

A farmer who understands local soil conditions, rainfall patterns, distribution challenges and market prices can use AI to improve decisions in ways that a person who merely knows how to prompt a chatbot may not.

A Nigerian entrepreneur who understands the difficulties small businesses face with payments, logistics, power supply and customer acquisition can use AI to develop targeted solutions.

A journalist who understands the political and economic context of a story can use AI to process documents faster without surrendering the responsibility to verify claims.

A software developer who understands the needs of local users can use AI to build products that solve practical problems rather than simply generating another application with no clear market.

The advantage comes from combining tools with context.

There is also a warning for African educational institutions. If assessment systems reward only memorisation and predictable answers, they risk preparing students for a world in which machines can reproduce those answers cheaply.

Education should continue to teach foundational knowledge. Students cannot exercise sound judgment without understanding the subjects they are judging.

But they should also learn how to investigate unfamiliar problems, challenge assumptions, collaborate with others and make decisions when there is no obvious answer key.

Africa does not need to abandon academic excellence. It needs to broaden its understanding of what excellence requires.

The next definition of intelligence

Jensen Huang's observation is important because it challenges a familiar assumption: that the smartest person is necessarily the one who can solve the hardest technical problem or achieve the highest examination score.

Technical ability remains valuable. Knowledge remains valuable. Mathematics and programming remain important.

But when machines become capable of performing more of these tasks, the definition of competitive advantage changes.

The ability to recognise an emerging problem, understand a customer's unspoken concern, identify a flawed assumption or anticipate the consequences of a decision may become more important in many professional settings.

These abilities are not mystical. They can be developed through experience, reflection, observation and deliberate practice. They can also be supported by data and AI tools.

What makes them difficult is that success often depends on circumstances that cannot be fully controlled or measured.

The smartest person in a room may not be the one who answers every question first. It may be the person who realises that everyone is answering the wrong question.

It may be the manager who recognises that a team is losing trust before productivity collapses. It may be the entrepreneur who notices that customers are struggling with a problem they have not yet learned how to articulate.

Or it may be the person willing to challenge a popular decision because the evidence suggests that something important has been overlooked.

None of these capabilities guarantees success. Human intuition can be biased, and confidence is not the same as wisdom. The best judgment combines observation, evidence, experience and a willingness to revise one's beliefs.

That combination will matter even as AI becomes more capable.

Conclusion: compete where the answer is not obvious

The SAT was designed to assess academic aptitude, not to capture every dimension of human intelligence. Its limitations should not be mistaken for a flaw in the examination itself.

The larger lesson is that machines tend to advance rapidly when researchers can define success clearly, generate large numbers of attempts and evaluate the results reliably.

Mathematics and programming provide useful examples. Human relationships, leadership and complex real-world decisions present different challenges because outcomes are often ambiguous, delayed and influenced by factors that cannot be fully measured.

AI will continue to improve in these areas. The boundary between what machines can and cannot do will keep moving.

That makes the sensible response neither to dismiss AI nor to assume that every human advantage will disappear.

Instead, we should become better at identifying the capabilities we are developing, understanding how they create value and recognising where automated performance measures fail to capture the full problem.

For students, that means learning beyond examinations.

For workers, it means combining technical competence with judgment and domain knowledge.

For leaders, it means listening carefully, challenging assumptions and recognising that a decision can look correct on paper while failing in practice.

And for societies, it means building educational and professional systems that reward more than what is easiest to count.

The most important question to ask about any skill is not simply whether a computer can perform it.

It is how the skill creates value, what evidence demonstrates competence and whether the measures we use capture what genuinely matters.

If a machine can practise something millions of times because the answer is immediately available, learn to do more than produce the answer. Learn to understand the problem, recognise the context and judge what should happen next.

That may not make you permanently irreplaceable. No single skill can promise that.

But in a world where producing answers is becoming cheaper, the ability to ask better questions and make sound decisions may become one of the most valuable advantages a person can develop.