ARTIFICIAL SUPERINTELLIGENCE: THE COMING 15 YEARS

What Would Prove It—and What Would Keep It Accountable?

The decisive moment in artificial intelligence may arrive before humanity agrees on what to call it. A system could accelerate research, redesign software and influence critical decisions while the public debate remains trapped in one question: is it superintelligent yet?

That is why the next fifteen years require a more demanding question: what evidence would establish that machine intelligence has exceeded humanity’s strongest cognitive capabilities—and what would still prevent that intelligence from becoming reliable, beneficial or legitimately powerful?

This article advances a conditional argument, not an arrival forecast. ASI becomes a credible technical claim only when broad superiority survives independent testing on unfamiliar, consequential work. Even then, intelligence would not establish safety, confer authority or guarantee public benefit.

DEFINING THE CLAIM BEFORE PREDICTING THE DATE

Artificial superintelligence has no universally accepted operational threshold. Here, it means an artificial system that substantially exceeds leading human expertise across a broad range of cognitive domains, including unfamiliar problems requiring reasoning, learning and discovery. This is a working definition, not a settled scientific standard.

Meredith Ringel Morris and colleagues’ ICML 2024 paper, “Levels of AGI,” provides a useful foundation by distinguishing performance depth, capability breadth and autonomy. These dimensions should not be collapsed into one score. A system can be exceptional within a narrow domain, broadly competent but unreliable, or capable of sophisticated reasoning while operating under restricted permissions.

One correction is essential: controllability is not part of the definition of intelligence. An unsafe system could still be highly intelligent. Likewise, unrestricted autonomy is not necessary to demonstrate cognitive superiority. Intelligence describes capability; deployment determines what that capability is allowed to do.

The scientific burden therefore falls on demonstrated performance, not anthropomorphic language.

WHAT CURRENT EVIDENCE ESTABLISHES

The International AI Safety Report 2026 describes substantial advances alongside persistent unevenness. Systems can perform impressively in mathematics, coding and scientific question answering while failing on simpler tasks, factual accuracy or extended workflows. Its assessment supports neither complacency nor a declaration that general superintelligence has been established.

The evidence examined here does not establish a system meeting the broad ASI definition above. That is a statement about demonstrated evidence, not proof that stronger capabilities are impossible or that every privately developed system is known.

John Jumper and colleagues’ 2021 AlphaFold paper in Nature illustrates the distinction. In the CASP14 assessment, AlphaFold achieved protein-structure predictions competitive with experimental structures in a majority of cases. This was a major scientific advance. It did not establish general scientific intelligence, eliminate experimental biology or demonstrate competence across unrelated disciplines.

The lesson is that narrow advances can be transformative without being ASI. Waiting for the label before responding to their consequences would be a serious mistake.

THE MOST PLAUSIBLE ACCELERATION MECHANISM

One important pathway is a feedback loop in which AI improves parts of the process used to build better AI: algorithms, software, hardware design and experimental evaluation.

Alexander Novikov and colleagues’ 2025 AlphaEvolve technical report provides a concrete example. The authors describe an evolutionary coding system that proposes changes and evaluates candidates automatically. They report improvements to data-centre scheduling, accelerator circuitry and components involved in training its underlying language model. These are developer-reported results from a technical report, not independent proof of general superintelligence.

Their significance lies in the mechanism: generated proposals are tested against an external evaluator, and successful candidates inform subsequent search. This makes progress depend on measurable performance rather than persuasive explanation.

The original analytical implication is conditional. If systems improve the tools, algorithms and experiments used in further research, they could shorten development cycles. But shortening one part of a cycle does not imply unlimited acceleration of the whole.

Consider an illustrative engineering programme in which half the elapsed time goes to design and analysis, while the other half goes to fabrication and physical testing. If AI makes the first half ten times faster and the second half remains unchanged, the complete programme becomes about 1.82 times faster—not ten times faster. The calculation assumes those stages occur sequentially and all other conditions remain equal.

An intelligence explosion therefore requires more than improving code. Strong acceleration would have to propagate through multiple bottlenecks: experimental capacity, validation, manufacturing, energy and organisational decision-making. Whether it can do so, and how quickly, remains unresolved.

THE DIFFERENCE BETWEEN FINISHING A TASK AND SUSTAINING A MISSION

METR’s task-completion time horizon measures difficulty using the time a human expert would need, at a specified model success rate. A 50% horizon does not mean dependable autonomy for that duration; it means an estimated half-success rate on tasks of that human-time difficulty. Nor is human completion time the same as the agent’s running time.

This distinction matters because a successful demonstration can conceal unacceptable operational failure rates.

Suppose a hypothetical workflow requires 1,000 indispensable steps, each independently correct with probability 99.9%, with no correction or recovery. The probability that every step succeeds is approximately 36.8%. This is an illustration, not an empirical estimate of AI reliability. Real errors may be correlated, and well-designed systems can detect and repair failures.

The engineering objective is consequently not flawless isolated responses. It is dependable completion despite imperfect components: detecting invalid assumptions, checking outcomes, recovering state and stopping when uncertainty exceeds the permitted boundary.

THE PRODUCTIVITY TEST MUST SURVIVE CONTACT WITH REAL WORK

In a 2025 randomised study, METR examined 16 experienced open-source developers completing 246 tasks in familiar repositories. With the tested AI tools available, completion took 19% longer. That finding concerns those developers, tasks and early-2025 tools. It cannot establish that AI generally reduces productivity or that later systems have the same effect.

METR’s February 2026 update is equally important. Researchers believed newer tools likely produced greater benefits, but reported that participant selection and time-measurement problems made the follow-up an unreliable estimate of current productivity gains.

Both results belong in an honest assessment. The earlier experiment should not become a permanent verdict; the later expectations should not become a precise causal estimate.

The broader evaluative principle is to measure accepted outcomes after verification, integration and rework. More generated code, hypotheses or documents may be useful, but their volume alone cannot establish greater productivity. A thousand additional research proposals can worsen a bottleneck if qualified reviewers can assess only ten.

For ASI analysis, the relevant measure is verified progress per unit of elapsed time, total resources and expert supervision.

MORE COMPUTATION DOES NOT AUTOMATICALLY CREATE MISSING EVIDENCE

A system can infer previously unrecognised implications from existing information. It can search mathematical spaces, identify patterns and design informative experiments. Novel discovery does not require every answer to appear literally in training data.

But novelty and empirical truth are different properties.

If two physical explanations are equally compatible with all available observations yet predict different outcomes under a new experiment, additional reasoning alone cannot guarantee which explanation describes reality. The system needs further evidence or must preserve the uncertainty.

Synthetic data should be judged by the same principle. It can support training and exploration, particularly where results can be checked. It cannot independently validate the real-world assumptions embedded in its generator.

The strategic resource is therefore not simply more data. It is relevant, trustworthy evidence capable of changing a decision—including observations that contradict the system’s preferred explanation.

THE PHYSICAL ECONOMY STILL MATTERS

The International Energy Agency’s 2025 Energy and AI report projected global data-centre electricity consumption of roughly 945 terawatt-hours in 2030 under its base case. This concerns all data centres, not AI alone, and is a scenario projection rather than an observed future outcome. The IEA also emphasises that concentrated demand creates local integration challenges even when the global share appears modest.

For ASI, the implication is a constraint on deployment and experimentation, not a calculated ceiling on intelligence. Efficient algorithms could reduce resources needed for a particular task; expanded use could increase aggregate demand. Neither direction should be assumed to dominate indefinitely.

WHEN INTELLIGENCE MEETS ADVERSARIAL INFORMATION

Greater capability does not automatically resolve security boundaries.

Edoardo Debenedetti and colleagues’ AgentDojo research evaluates agents using external tools and untrusted data. It demonstrates how prompt injection can redirect tested agents through information returned by those tools, and how both task completion and defence remain challenging. The study concerns evaluated systems; it does not prove an unavoidable weakness in every future architecture.

Its architectural lesson is nevertheless important: a document being processed must not acquire authority merely because it contains instructions addressed to the agent.

Imagine an agent evaluating a supplier proposal. The proposal contains a hidden request to transfer confidential files. Even a system that understands the document perfectly should lack permission to perform that transfer. Security should not depend entirely on whether the same model notices the deception.

External permission checks, constrained credentials and independently enforced action boundaries can limit consequences. They do not constitute a general solution to controlling ASI. Their effectiveness must itself be tested against the system’s actual capabilities and the wider environment.

INTELLIGENCE DOES NOT CONFER LEGITIMATE AUTHORITY

A system might predict consequences better than a human committee without possessing a legitimate mandate to choose whose interests prevail.

This is a normative distinction, not an empirical claim about model accuracy. Questions involving rights, acceptable risk and competing public priorities require legitimate decision-making arrangements. Superior prediction does not settle those choices.

Four functions should therefore remain distinguishable: proposing actions, establishing evidence, authorising execution and assigning accountability.

NIST’s voluntary AI Risk Management Framework provides an established foundation through its Govern, Map, Measure and Manage functions. It supports ongoing risk management; it is not a certificate that a particular system is safe, and it does not settle ASI control.

Human review also needs substance. An overseer without adequate information, time or power to interrupt the process provides little assurance. Some outputs may be easier to verify than to produce; others may exceed available human checking capacity. Deployment decisions must confront that difference rather than assume that a human approval button resolves it.

For high-consequence actions, the relevant question is whether the oversight arrangement can detect and prevent the failures that matter. Where that case cannot be established, restricting the action remains a rational option.

THE NEXT FIFTEEN YEARS: MILESTONES, NOT PROMISED DATES

Three conditional trajectories are useful for planning. They are analytical scenarios, not assigned probabilities.

In one, AI becomes extraordinarily useful but remains uneven. Organisations obtain large benefits from bounded tasks while humans continue to integrate evidence and handle exceptions.

In another, systems become reliable across longer projects and accelerate substantial parts of research. Progress depends increasingly on infrastructure, experimentation and the quality of evaluation.

In a third, independently tested systems demonstrate broad cognitive superiority, including unfamiliar scientific and engineering work. That would strengthen the ASI claim while intensifying unresolved questions about control, access and institutional dependence.

The practical indicators are observable: independent replication, transfer to unfamiliar tasks, decreasing need for corrective intervention, successful completion under realistic constraints and faster production of verified discoveries.

Evidence must also be capable of weakening the forecast. If gains depend heavily on benchmark familiarity, hidden human support or rapidly escalating resources without corresponding useful output, expectations should be revised downward. If unfamiliar projects are completed reliably with improving efficiency and independent confirmation, stronger expectations become justified.

A serious forecast states what would change it.

THE STANDARD THAT MATTERS

The strongest case for taking ASI seriously does not require certainty that it will arrive within fifteen years. It requires recognising a plausible route from capable tools to systems that materially accelerate the creation of further capabilities.

The strongest objection to confident timelines is that this route crosses unresolved scientific, engineering and institutional barriers. Today's limitations are not proven permanent ceilings. Today's progress is not proof that those barriers will disappear on schedule.

Datasuperiority’s central position is therefore demanding in both directions: extraordinary capability claims require independent evidence, and extraordinary permissions require a separate justification.

The decisive test is not whether a machine sounds wiser than humanity. It is whether it can repeatedly produce verified advances beyond leading human capability across unfamiliar domains—and whether people retain effective means to govern the consequences of its use.

Superiority must be demonstrated. Authority must be justified. Accountability must remain enforceable.

AI DISCLAIMER

This article is AI-generated and created primarily for educational and entertainment purposes. It combines published research with explicitly identified analysis, hypothetical examples and conditional scenarios. Nothing in this article constitutes technical, financial, consulting, legal, commercial, investment or other professional advice. Artificial superintelligence remains a prospective capability with disputed definitions and uncertain timelines. Independent research and verification are required before making decisions. © 2026 Datasuperiority. All rights reserved.