Artificial General Intelligence (AGI) or Superintelligence?

A company is about to buy a system that promises to research, write software and organise projects. Suppose that, during the demonstration, it spots an error nobody else has noticed and proposes an impeccable repair. The supplier describes this as a step towards artificial general intelligence. The buyer has a less imposing question: can it be allowed to change the software that customers actually use?

The distance between those questions contains much of the debate about AGI. Solving a difficult problem establishes something. Determining how much it establishes requires knowing what the test involved. Allowing the solution to be implemented introduces a different decision, with consequences for people who may not even know the system exists. A demonstration can last minutes; undoing a bad authorisation can take considerably longer.

Superintelligence takes the discussion further. What if the advantage ceased to appear in particular tasks and extended across almost every intellectual field? What could we check ourselves, and what would we begin to accept because we could no longer evaluate the answer? These are reasonable questions about a possibility. Confusing them with a diagnosis of the present turns a demanding hypothesis into a remarkably convenient sales label.

AGI and superintelligence deserve different definitions, but a glossary will not settle the controversy. What matters is the evidence that would establish each capability, the conclusions that would be premature and the power granted while we argue about what to call it.

The word “general” has work to do

In their framework of AGI levels, Meredith Ringel Morris and her co-authors separate the breadth of capabilities from the standard of performance. They also distinguish the autonomy of a deployed system. Their proposal allows advances to be described without immediately reducing them to a yes or no on general intelligence. It is a research framework, rather than a definition accepted by all sides. Source: arXiv

The distinction matters because an extraordinary ability can be extremely narrow. Imagine a program that solves one class of mathematical problems with exceptional precision, but needs a person to translate every assignment into its required format. Its achievement remains intact. What has not been established is whether it can recognise the problem before it, select an appropriate method or transfer what it has learnt to another situation.

The reverse case also deserves attention: a system may perform adequately across numerous tasks without mastering any of them at the level of the finest specialists. Breadth and excellence answer different questions. Combining them in one score can make a ranking easier to read while making the underlying comparison harder to understand.

Claims about AGI therefore require an explicit account of the tasks, performance and adaptability demanded. “Like a human” leaves much undecided. People differ in training and experience; a test comparing a system with beginners supports a different claim from one involving experienced professionals. The opponent should not change halfway through the celebration.

Nor is adding more examinations enough. Generality implies some ability to cope when a situation changes, rather than simply accumulating correct answers in prepared settings. The decisive question is how far the problem can vary before a person must once again organise the work on the system’s behalf.

Superintelligence sets a tougher comparison

Nick Bostrom defines superintelligence as vastly outperforming the most capable humans across practically every intellectual field. His definition includes scientific creativity and social skills, while leaving consciousness an open question. Beating the average person in several tests, or an expert in one specialism, does not by itself meet that threshold. Source: Nick Bostrom

This definition helps avoid two shortcuts. The first presents any superhuman performance as heralding universal superiority. The second requires us to settle whether a machine has subjective experience before discussing its abilities. Consciousness raises philosophical problems of its own; the possibility that a system could produce important effects does not depend on resolving them first.

“Superior” does not mean “preferable for every use”, either. A hypothetical system might provide better predictions and cost too much for a routine task. Another might be less accurate but operate under constraints that make it suitable for a particular purpose. Intellectual ability, cost and operating conditions enter different decisions, even when a product must satisfy them together.

AGI and superintelligence therefore do not necessarily describe mutually exclusive paths. A system might acquire general capabilities and later achieve a much higher standard of performance. That possibility does not make the second step inevitable, immediate or easy to identify. Defining a boundary says nothing about the speed at which someone will cross it.

An examination is not a working day

A benchmark is an organised test for comparing performance. Its usefulness depends on what it makes observable: particular problems, known conditions and criteria for success. To evaluate a promise of general capabilities, we need to know which part of the work it represents and what it leaves out.

Return to the software buyer. The demonstration involved a well-defined error. An actual working day might bring an ambiguous request, contradictory documentation and the discovery that a repair affects a customer with special requirements. The challenge would include recognising missing information and stopping to request it. Solving the original error does not automatically cover any of those difficulties.

How the result was obtained matters too. An answer selected after numerous attempts does not answer the same question as the first attempt of a system left in charge. If a person identified the failed proposals and selected the good one, part of the performance belongs to that supervision. Knowing this does not invalidate the tool: it allows the buyer to calculate how much human work it still requires.

Evaluation should distinguish successes, failures and recoveries. An error that is visible and correctable may be less costly than an answer that is usually right but occasionally wrong in ways that are difficult to discover. Anyone buying the system needs to know where to look for those errors and what checking will cost, as well as how often they occur.

Context changes even the meaning of an average. Suppose two assistants complete the same number of assignments. One fails mainly at jobs that can be repeated; the other makes its mistakes in decisions affecting customer data. A statistical tie is not an operational tie. A useful comparison must describe the consequences grouped under the word “failure”.

Was it a model or an organisation competing?

Google DeepMind’s presentation in March 2026 proposes evaluating distinct cognitive abilities against distributions of human performance. Breaking down the measurement forces the question of which aptitude each result supports before combining results into a conclusion about general intelligence. A proposal for evaluation does not certify that such intelligence has been achieved. Source: Google DeepMind

Another distinction is just as important: identifying the unit being evaluated. A model answering alone, one connected to tools and a coordinated group of agents can use different resources. Describing all three simply as “AI” erases differences that matter both for reproducing the result and for deciding how to use it.

A fair comparison does not require depriving a machine of tools that people are allowed to use. It requires describing those tools. If a system consults a library and performs calculations, the test should state whether its rivals could do the same, how much time they had and what assistance they received. Asymmetrical conditions can still yield useful information, provided the conclusion preserves that asymmetry.

The buyer also needs to compare complete systems. A cheap product requiring intensive review can become expensive; a costlier one could save work if its results are easy to check. That is why a performance score is not a productivity calculation. Integration, supervision and the cost of mistakes stand between the two.

Delegation changes the test

An assistant that proposes a repair and one allowed to apply it may share an ability, but they do not create the same risks. In the first case someone examines a recommendation. In the second the system changes its environment. The difference lies in permissions, procedures and the ability to stop or reverse an action.

Consider three configurations of the same program. It can work on an isolated copy; prepare changes for someone’s approval; or intervene directly in the service. Its intellectual capacity does not transform every time its configuration changes. What changes is users’ exposure to its successes and mistakes. Discussing autonomy without describing that setting leaves out what matters most to those who would suffer an error.

Human approval is not a guarantee merely because it exists. If a person must review more changes than they can understand, the confirmation button adds an appearance of supervision. Effective review requires time, information and the authority to reject a proposal, together with a workable alternative when it is rejected.

In Intelligent AI Delegation, Tomašev, Franklin and Osindero propose treating the division of tasks alongside the authority, responsibility and limits of those involved. Their framework pays attention to adaptation when conditions change or a participant fails. It is a conceptual proposal, rather than a demonstration of a universal solution to safe delegation. Source: arXiv

For an organisation the consequence is practical: before expanding permissions, it should test the failures those permissions make possible. What happens if the system misinterprets an instruction, loses access to a tool or receives incompatible information? Observing only successful runs leaves precisely that part of the decision without a foundation.

Progress needs another argument to become acceleration

The public abstract of From AGI to ASI identifies four research routes: scaling, new paradigms, recursive improvement and collectives of agents. It also acknowledges obstacles and uncertainties. That description helps organise hypotheses about future capabilities; it does not establish an arrival timetable. Source: arXiv · Google DeepMind

Recursive improvement attracts attention because it introduces a powerful possibility: a system able to improve the processes used to build its successor. Imagining that cycle, however, does not establish its performance. A proposed improvement must be distinguished from a verified one, and a local improvement from one that raises the performance of the whole system.

Consider a simple hypothetical example. A system produces a version of a program that runs faster in a test. Before adopting it, someone must determine whether it preserves correct results, fails in other cases and offers a gain worth the cost of checking it. Automating proposals might accelerate one part of the cycle while verification continues to constrain the final result.

If the argument is that verification can also be automated, a further evidential requirement appears: showing that the evaluator detects the relevant failures in what it approves. Agreement between two components is not enough. Their agreement would provide little reassurance if both overlooked the same decisive case.

A claim about acceleration therefore needs to identify which task is improving, what previously constrained it and which new limit appears afterwards. A forecast can be plausible without being demonstrated. Its assumptions should be available for separate examination, rather than sheltered behind the word “exponential”.

Why should the destination resemble us?

In their preprint’s abstract, Goldfeder and his co-authors question whether intelligence should be measured exclusively against the human repertoire. They propose adaptable specialisation with superhuman capabilities. This is an argument about the goal worth pursuing, rather than evidence of an architecture that already achieves it. Source: arXiv

The objection invites a distinction between two interests. For those studying intelligence, explaining how a system learns and adapts matters. For someone who needs to solve a problem, a highly reliable specialised ability may be enough. A scientific tool does not lose value because it cannot write a novel; a versatile assistant does not acquire scientific accuracy by writing one well.

Abandoning the human comparison does not remove the need for evaluation, however. It moves evaluation towards the chosen objective. If a new ability is promised, its proponents must explain how success would be recognised, which tests could refute the promise and who has the means to repeat them. A less anthropocentric goal can be intellectually fruitful and still require ordinary verification.

Nor must we wait for general intelligence before important consequences arise. In the software example, a specialised capability accompanied by extensive permissions could modify numerous services. The scale of damage or benefit would depend on its use, connections and restrictions, as well as the intellectual breadth of the system itself.

What evidence would change our minds?

A useful argument about AGI should allow participants to explain which result would change their judgement. If every success counts as definitive proof and every failure as a minor difficulty, the claim is insulated from evidence. The same happens when a critic redefines intelligence every time a machine meets the previous requirement.

In the purchasing example, a testable criterion might be completing unfamiliar assignments under conditions established in advance, without a supervisor having to reorganise the task after each obstacle. That result would support a particular ability. Extending the conclusion to other domains would require further tests involving problems that did not all depend on the same skill.

Two kinds of uncertainty should also be distinguished. Evidence may be absent because nobody has conducted an adequate test, or because its results are inaccessible. The first situation calls for investigation; the second calls for opportunities to check. Presenting both as an unavoidable mystery grants an unnecessary advantage to whoever controls the information.

Tests have authors, and permissions do too

Developers possess knowledge indispensable to evaluating a system. They also choose which results to present and how to frame them. That combination makes external evaluation useful, provided examiners have sufficient access to check what matters. An independent report on a tightly restricted demonstration can be independent only within that restriction.

The answer is not to discard all corporate evidence. It is to demand conditions that allow that evidence to be tested: an account of the experiment, selection criteria, failed results and changes made during evaluation. Some checks will require confidential access to protect data and security. What matters is that confidentiality should not make it impossible to refute a public claim used to justify granting new powers.

The burden of demonstration should rise with the consequences of the permission requested. A tool for exploring ideas can tolerate results that would be unacceptable in a system authorised to alter records. This does not turn caution into paralysis: it allows uses to expand gradually where evidence supports each step, while retaining restrictions where it does not.

The buyer we began with need not resolve the whole philosophy of intelligence before making a decision. The buyer needs to know what was tested, what might fail and who can intervene when it does. If the supplier claims to offer AGI, it must also demonstrate the generality it advertises. If it claims superintelligence, the comparison will be more demanding still.

While that evidence is being established, granting permissions remains a decision for those who buy, develop and use the system. No score makes it on their behalf. The question worth attaching to every announcement of a new intelligence is which specific action someone hopes to authorise in its name.