The Limits of AI: From Idealisation to Realisation

Standfirst: Artificial intelligence can solve advanced mathematical problems yet stumble over elementary tasks. The harder question is whether businesses and governments can convert technological brilliance into dependable, affordable and accountable progress.

The most revealing measure of artificial intelligence in 2026 may not be what it can do, but what happens when someone depends on it. A machine can master difficult mathematical challenges, generate functioning software and summarise mountains of documents. Under different conditions, it can misread a clock, invent a citation, or lose its way through a sequence of computer tasks. This is not evidence that AI has achieved nothing. It is evidence that extraordinary capability and dependable judgement are different accomplishments.

The contrast is documented in Stanford’s 2026 AI Index. The report describes leading models reaching gold-medal performance in mathematical competition, while the strongest model on one analogue-clock benchmark achieved only about half the accuracy of human test-takers. Agents improved markedly on a benchmark of computer-use tasks, yet still failed roughly one attempt in three. These are measurements of specified evaluations, not universal estimates of AI competence. Precisely for that reason, they expose a problem that sweeping narratives tend to conceal: success does not travel effortlessly from one task to another.

We have spent several years asking how intelligent AI might become. The more urgent question is how much of that intelligence can be translated into reliable human benefit. The era of AI idealisation is giving way to an era of realisation, in which costs, institutions, evidence and accountability matter as much as impressive demonstrations.

The age of idealisation: how a breakthrough became a prophecy

The early enthusiasm was not irrational. Generative AI made sophisticated language and coding tools available through ordinary interfaces. Some capabilities advanced faster than established testing methods could measure them. Stanford reports that frontier models gained 30 percentage points in a single year on Humanity’s Last Exam, a deliberately difficult evaluation. A technology capable of changing how knowledge work is performed deserves serious attention.

Yet a breakthrough in capability can quietly become a prophecy about society. The ability to draft a legal argument becomes a claim that lawyers are replaceable. A coding demonstration becomes a prediction that software engineering will be automated. A laboratory result becomes an assumption about national productivity. These leaps mix distinct questions: Can the system perform the task? Can it perform it consistently? Can an institution use its output safely? Does doing so create more value than it costs?

Commercial incentives magnify the confusion. Model developers compete for attention and capital. Investors prize evidence of enormous future markets. Businesses fear missing an innovation wave. None of these motives makes progress imaginary; they do, however, reward the most expansive interpretation of ambiguous evidence. In January, TechCrunch described 2026 as a transition from AI hype to pragmatism, emphasising smaller, useful systems and integration with actual workflows. That transition is more important than the familiar argument over whether an AI bubble exists. A technology can be transformative and still be overpromised, overpriced or badly deployed.

The first correction is therefore intellectual: abandon the idea that intelligence is a single ladder on which every rung automatically unlocks the next.

The intelligence paradox: brilliance without dependable boundaries

The term jagged intelligence captures a distinctly modern difficulty. An AI model can excel at a specialised intellectual challenge while failing at something a person would regard as elementary. The problem is not merely embarrassing mistakes. It is that users may struggle to know in advance which tasks lie inside a system’s dependable range.

On the OSWorld computer-use benchmark, leading agents reached 66.3% task accuracy, according to Stanford’s 2026 assessment. That is striking progress from roughly 12%, but it is not an acceptable unattended error rate for many financial, medical or administrative processes. The benchmark does not predict failure rates in a particular company; it demonstrates the danger of treating improving autonomy as synonymous with mature autonomy.

Confident language adds another layer of risk. AI systems can produce plausible but unsupported references, merge incompatible sources or fail to acknowledge uncertainty. Stanford’s responsible AI assessment documents substantial variation across models on a specific factual-accuracy evaluation. Such measurements cannot be generalised to every application, but they make the case for task-specific testing rather than blanket trust.

Consider the difference between asking a system to produce an initial research outline and permitting it to submit a court filing. The first task tolerates correction. The second carries duties to clients, judges and the public. A useful draft can be imperfect; a cited authority cannot be fictional. In healthcare, the same principle becomes even more urgent: high average performance does not eliminate the need to understand rare, serious failures and who is responsible for them.

There is a counterargument worth taking seriously. Today’s error patterns may be transient. Better reasoning methods, verified retrieval, structured tools and more capable models could dramatically improve reliability. Stanford’s own results show that capabilities are not simply stagnating. But an expectation of future improvement is not proof of present fitness for a consequential task. The intelligent response is neither panic nor blind adoption: it is independent evaluation against the conditions of actual use.

The economic reality check: productivity is not profit

The business world has acquired a new paradox: AI can make people faster without making their organisations proportionately richer. McKinsey’s August 2026 State of AI survey found that 80% of respondents reported improved individual productivity, while 37% attributed at least some earnings-before-interest-and-taxes impact to AI. Only about 6% met the survey’s more demanding definition of AI high performers. Those percentages describe survey responses under different measures; they do not establish causal effects on audited profit.

The distinction matters. A worker who drafts emails in half the time may free capacity that the business never redeploys. Faster coding may produce more software requiring testing, maintenance and security review. A chatbot may resolve routine questions while leaving the most complicated cases to human teams. The productivity gain is real at one level, yet partly consumed elsewhere.

An NBER study of 5,179 customer-support agents found an average productivity improvement of 14% with AI assistance, with larger gains among less experienced workers. That is meaningful evidence in a defined setting, not a universal forecast for every profession. The difficulty begins when a successful task-level experiment is sold as a company-wide economic transformation.

An especially instructive contrast appears in measures of adoption. Stanford’s AI Index economy chapter reports that 88% of surveyed organizations used AI in 2025. But a 2026 NBER paper drawing on the U.S. Census Bureau’s business survey finds that 18% of firms used AI in at least one business function during November 2025–January 2026, with higher employment-weighted adoption. These numbers are not contradictory estimates of an identical population. They use different samples, scopes and definitions. Their divergence illustrates why the phrase everyone is using AI tells us far less than it appears to.

The most revealing organisational evidence may concern what executives do not count. In a September 2026 MIT Sloan field study, researchers compared two organisations implementing AI initiatives. One built enduring support for employee experimentation and had 141 organisation-wide solutions in use; the other, where the extra effort went inadequately supported, had three. These are two studied settings, not a controlled estimate of what every company can achieve. Still, the mechanism is persuasive: testing, correcting, documenting and integrating AI are forms of labor, not friction that magically disappears after installation.

This suggests what might be called the AI realization gap: the distance between impressive technical output and lasting economic benefit. It narrows only when useful work survives the full costs of integration, supervision, security, correction and change. If those costs remain invisible, return-on-investment calculations can become another exercise in idealisation.

The human frontier is not a sentimental argument

It is tempting to respond by declaring that machines will never replace human creativity, empathy or judgement. Such absolute claims are as poorly founded as predictions of effortless total automation. Human workers make errors, resist change and sometimes perform routine tasks badly. AI can expand access to expertise, reduce drudgery and help people who lack time or specialist support.

The more defensible question is where human responsibility remains essential, even as technical performance improves. Who should decide which risks are acceptable? Who can contest a rejected benefit application, an erroneous medical recommendation or a faulty hiring decision? Who has the authority to disclose uncertainty, compensate harm and redesign a process that fails systematically?

These are not simply questions about whether a model can generate a convincing answer. They are questions about institutions and legitimacy. The U.S. National Institute of Standards and Technology’s AI Risk Management Framework treats trustworthiness as a combination of reliability, safety, transparency, accountability, privacy and fairness. Importantly, these properties must be assessed in their context of use; no single benchmark can substitute for that judgment.

Human oversight, however, should not be a comforting label attached to an automated system. A nominal reviewer given hundreds of outputs and a few seconds per decision provides little meaningful protection. Oversight needs time, expertise, authority to intervene and records showing when the system was corrected. The aim is not to preserve every existing job description but to prevent responsibility from evaporating as work is redistributed.

The possible reward is considerable. AI can raise less experienced workers’ performance, support medical research and assist scientists in exploring difficult problems. Denying these benefits would leave genuine human needs unmet. But the right unit of progress is not the number of tasks removed from people. It is the quality, accessibility and accountability of the outcomes people receive.

The physical limits of an apparently weightless technology

Digital enthusiasm often ignores the physical supply chain that sustains it. AI systems rely on chips, electricity, cooling, networking and buildings; they compete for capital, scarce components and grid connections. Their expansion is therefore constrained by schedules and resources that cannot be accelerated merely by writing better software.

In its April 2026 energy assessment, the International Energy Agency projected that global data-centre electricity use could rise from approximately 485 terawatt-hours in 2025 to 950 TWh in 2030. That estimate covers data centres generally, not exclusively AI workloads, and remains a projection rather than an observed outcome. AI-focused facilities are nevertheless an important driver of the increase.

The IEA also describes stresses involving grid capacity, advanced chips and high-bandwidth memory. These are not arguments against AI development. They are reminders that infrastructure determines the pace and price of adoption. Better model efficiency could moderate some demand, while cheaper, more accessible AI could encourage new and more energy-intensive uses. Both possibilities deserve attention.

The geography of these constraints matters. Stanford’s 2026 policy assessment finds expanding AI ambitions in emerging economies alongside highly unequal access to advanced computing infrastructure. For Pakistan and other developing countries, the dilemma is not whether to join the AI economy; it is how to participate without becoming permanently dependent on foreign models, scarce imported computing capacity and systems poorly suited to local languages or public institutions.

Pakistan’s National AI Policy, approved in July 2025, signals ambitions around skills, infrastructure, innovation and responsible adoption. The test is implementation. A reliable agricultural advisory service, accessible educational tools or improved administrative processing may matter more than a national attempt to imitate the largest frontier-model laboratories. Strategic realism means selecting problems for which reliable, affordable AI can be supported locally, not mistaking technological prestige for national development.

Five tests for moving from idealisation to realisation

If the AI debate is to grow up, business leaders and policymakers need better questions than How powerful is the latest model? A practical framework might begin with five tests.

The reliability test. Define the actual task, its acceptable error threshold and the consequences of failure. Test the system on representative cases, difficult exceptions and changing conditions. A model’s general reputation is not an operating guarantee.

The economic-value test. Count benefits against the entire cost of implementation: integration, compute, training, verification, correction, maintenance and displaced work. Measure outcomes that matter to users or the organisation, not simply tokens generated, accounts activated or hours supposedly saved.

The accountability test. Name the people and institutions responsible for decisions, appeals and remediation. Make it possible to reconstruct what the system did. If nobody can explain who must fix a harmful outcome, autonomy has outpaced governance.

The sustainability test. Evaluate infrastructure dependencies, energy use, resilience and long-term affordability. Include the social costs and trade-offs that sit outside a vendor’s invoice.

The legitimacy test. Ask whether the deployment respects rights, earns trust and improves access fairly. In public services especially, efficiency cannot be the only objective. Faster exclusion is not better government.

These tests amount to a second original idea: the cost of trust. AI’s economic worth is not merely the market value of its outputs. It is the value remaining after society pays what is necessary to make those outputs dependable. Verification is not a regrettable tax on progress; in consequential settings, it is part of the product.

There is reason for optimism. Gartner’s October 2026 analysis argues that the center of enterprise attention is shifting from demonstrating capability toward governing costs, reliability and risk. That is not the end of innovation. It may be the point at which innovation becomes more credible.

AI’s future depends on better judgment, not smaller ambitions

The strongest objection to a realism agenda is that it could become an excuse for timidity. If every emerging tool must satisfy standards designed for mature infrastructure before anyone experiments, breakthroughs could be delayed and advantages reserved for wealthy incumbents. That warning deserves respect. Research environments, carefully monitored pilots, and low-risk creative tasks should permit more experimentation than critical services do. Proportionate safeguards are better than uniform prohibition.

Yet there is an equally serious danger in confusing ambition with haste. An automated system is not advanced merely because it performs many actions without a person. It is advanced when it improves outcomes without making the causes of failure harder to discover or responsibility easier to evade.

The next phase of AI may bring larger models, surprising scientific achievements and systems that handle more tasks independently. It may also reward less glamorous investments: clean records, domain-specific evaluations, employee training, resilient infrastructure and enforceable accountability. Those improvements are not rival visions. They are the conditions under which dazzling capability can become durable progress.

Our mistake was never to imagine that artificial intelligence could change the world. It was to assume that the world would change automatically when the intelligence arrived. The real frontier is not only between human and machine. It is between what technology can demonstrate and what society can responsibly deliver.