There is a strange moment in technology when a machine does something astonishing.

We usually change the machine.

On September 29, 2026, something different happened.

The name changed.

For decades, the technology was called Artificial Intelligence.

AI.

A short, almost boring name that has survived everything from expert systems and chess computers to image generators, large language models and increasingly autonomous agents.

Now the White House has introduced a different vocabulary.

“Super Intelligence.”

The change was announced alongside a voluntary safety accord involving major technology companies, including Google, Meta, OpenAI, Anthropic, Nvidia and xAI. The White House also issued an executive order directing executive-branch agencies to use “Super Intelligence” and “SI” instead of “Artificial Intelligence” and “AI” in official communications, websites, reports, policy documents and other non-statutory material.

At first glance, it sounds like a technological milestone.

It isn’t.

At least, not yet.

And here is the extraordinary detail hiding inside the order.

The order currently defines “Super Intelligence” and “SI” using the existing federal statutory definition of artificial intelligence.

That existing U.S. legal definition describes a machine-based system capable of making predictions, recommendations or decisions influencing real or virtual environments.

Then the order says something even more revealing:

within 60 days, the administration’s science and technology adviser is to propose a new federal definition of “Super Intelligence.”

So the terminology has changed.

The legal definition has not yet caught up.

And science has certainly not announced that humanity has crossed some universally recognised threshold called “superintelligence.”

That leaves us with a wonderfully strange question:

What happens when the word arrives before the thing it is supposed to describe?

That is where this story becomes much bigger than a White House vocabulary change.

The Name Changed First. The Science Has Not.

The first mistake would be to read “Super Intelligence” as though it were a new product category.

It isn’t.

There was no September 29 laboratory announcement saying:

Machine X has surpassed humanity.

No scientific body declared:

ASI has arrived.

No internationally accepted benchmark suddenly changed from “human-level” to “superhuman.”

What changed was official terminology in one government.

That distinction is enormously important.

The executive order says the capabilities of frontier systems have advanced far beyond what was imagined when the term “Artificial Intelligence” was first introduced, and it instructs the executive branch to adopt “Super Intelligence” as its preferred terminology. But the same order acknowledges that a more specific federal definition still needs to be developed.

So we have three different things happening at once:

Layer	What happened in September 2026?
Political language	“Super Intelligence” became official executive-branch terminology
Legal terminology	Existing statutory AI definition temporarily remains the operative definition
Scientific meaning	No universally accepted threshold for “superintelligence” has suddenly appeared

This is not a minor technicality.

It is the entire story.

So What Does “Superintelligence” Actually Mean?

In scientific and philosophical discussions, artificial superintelligence, usually abbreviated ASI, is generally imagined as a system whose intellectual capabilities substantially exceed those of humans across a broad range of important cognitive tasks.

That is a much stronger claim than:

“This chatbot is very good.”

And it is much stronger than:

“AI is better than humans at mathematics.”

A machine can outperform humans at chess without possessing general intelligence.

It can beat humans at image classification without understanding the world like a human.

It can write code without having a human-like mind.

It can solve difficult mathematics while making an absurd mistake in a simple conversation.

This is one of the oldest lessons in AI:

performance is often narrower than the label people put around it.

The OECD’s work on AI makes the same conceptual point from another direction: AI is a broad category, and there is no single universally accepted definition that neatly captures every system called AI. Its current definition focuses on machine-based systems that infer how to produce predictions, content, recommendations or decisions affecting physical or virtual environments.

“Superintelligence” therefore needs even more precision, not less.

The Three Words People Keep Mixing Up: AI, AGI and ASI

The simplest way to understand the debate is to stop treating these terms as steps on a neat staircase.

They are better thought of as overlapping concepts.

Concept	Basic idea	Status
AI	Machine systems performing tasks involving perception, prediction, generation, reasoning or decisions	Real and widely deployed
AGI	A hypothetical system with broad, general cognitive capability across many domains	No universally accepted definition or confirmed arrival
ASI	A hypothetical system whose broad intellectual capabilities substantially exceed human abilities	Hypothetical

The important word in the last two rows is broad.

An AI system can already be superhuman in particular tasks.

That does not automatically make it superintelligent.

A calculator is superhuman at arithmetic.

A database is superhuman at retrieving stored information.

A chess engine can defeat the strongest human players.

None of these facts, individually, establishes general superintelligence.

The conceptual boundary is therefore much fuzzier than the headlines suggest.

The Intelligence Problem Nobody Has Solved

Here is the awkward question sitting underneath the entire debate:

What exactly is intelligence?

Humans don’t possess one giant internal number called “intelligence.”

We have different abilities.

Memory.

Language.

Planning.

Mathematical reasoning.

Spatial reasoning.

Social understanding.

Creativity.

Physical coordination.

Learning.

Adaptability.

Judgement.

Metacognition.

And humans themselves vary enormously across these dimensions.

One person can be an exceptional mathematician and a terrible negotiator.

Another can read people brilliantly but struggle with formal mathematics.

A third can design machines but cannot remember names.

So when someone says:

“This AI is more intelligent than humans.”

the natural response should be:

Which humans? Doing what? Under what conditions?

The question sounds pedantic.

It isn’t.

It determines what the claim actually means.

The Benchmark Trap: When the Test Becomes the Target

AI researchers have another problem.

They need measurements.

So they build benchmarks.

Give a model difficult questions.

Test coding.

Measure mathematical reasoning.

Evaluate scientific knowledge.

Check language ability.

Test visual reasoning.

Measure long-context performance.

Compare models.

This is useful.

But eventually something peculiar happens.

If the benchmark becomes important enough, developers optimise for it.

And once systems become good at the test, the test can stop measuring exactly what researchers thought it measured.

The International AI Safety Report 2026, produced with contributions from more than 100 experts and supported by more than 30 countries and international organisations, explicitly identifies an “evaluation gap.”

Its authors note that current evaluations do not reliably predict real-world performance, and that some benchmarks are outdated, narrow or vulnerable to data contamination.

This creates an uncomfortable loop:

Humans design the test → AI learns to perform the task → humans declare progress → AI becomes better at tests designed by humans.

But intelligence is supposed to be more than examination technique.

The harder challenge is discovering whether a system can generalise.

Can it encounter something genuinely unfamiliar and still reason correctly?

Can it transfer knowledge between domains?

Can it recognise when it does not know something?

Can it operate reliably when the environment changes?

Can it understand consequences that were not explicitly described?

Those are much harder questions.

The Strange New Evidence: AI Is Becoming More Agentic

This is where the story moves beyond chatbots.

A chatbot answers.

An agent can increasingly act.

It may:

break a task into steps;

search for information;

use software;

write and execute code;

manipulate files;

interact with digital tools;

monitor progress;

revise its approach;

continue working toward a goal.

That changes the risk calculation.

A wrong answer is one problem.

A wrong answer followed by an action is another.

The 2026 International AI Safety Report says AI agents are increasingly capable of useful work, including software-engineering tasks with limited human oversight, although current systems still cannot reliably perform the full range of complex, long-term tasks needed to automate many occupations.

Independent measurements from METR provide another way of looking at this transition.

METR measures the task-completion time horizon of frontier AI agents: roughly, the complexity of tasks that agents can successfully complete at specified reliability levels, measured using how long human experts would take to perform the tasks.

Its May 2026 update uses more than a hundred diverse software tasks and explicitly warns that a time horizon is not the same thing as the amount of time an AI can autonomously operate. It is a measure of task difficulty.

That distinction is crucial.

AI is not simply becoming “smarter.”

It is becoming increasingly capable of doing things.

And action changes the stakes.

The Real Safety Question May Be Moving From “Can AI Answer?” to “Can AI Act?”

Imagine two systems.

System A

You ask:

“Find the cheapest flight.”

It gives you the wrong answer.

Annoying.

System B

You give it permission to search, select, purchase, email your employer, update your calendar and arrange transport.

It misunderstands one instruction.

Now it has consequences.

The second system doesn’t necessarily need to be more intelligent.

It simply needs more agency.

That is why AI safety increasingly concerns:

capability + autonomy + access.

An extremely capable model with no ability to act externally presents one kind of risk.

An equally capable model with access to financial systems, code repositories, corporate networks, laboratories or infrastructure presents another.

The White House accord signed on September 29 specifically calls for internal controls around areas including cybersecurity, biosecurity and chemical threats, as well as controls intended to prevent AI systems from hacking or accessing technical systems unintentionally.

That tells us something important.

The safety debate is no longer purely theoretical.

It is increasingly about what powerful AI systems can actually do when connected to the outside world.

The Self-Improvement Question Is More Complicated Than “AI Improves Itself”

Now we arrive at the most misunderstood part of the superintelligence debate.

Self-improvement.

The phrase sounds simple.

But there are several very different things hiding inside it.

An AI can:

revise its own answer;

generate better prompts;

improve a workflow;

produce synthetic training data;

help researchers write better AI code;

design experiments;

evaluate another model;

modify parts of an AI-development pipeline;

potentially contribute to research that improves future AI systems.

These are not equivalent.

A recent 2026 survey of 1,250 arXiv papers from 2024–2026 distinguishes bounded self-refinement from more ambitious forms of open-ended recursive self-improvement.

Its authors argue that current demonstrated self-improvement remains constrained by issues including verification, compute, model collapse and the difficulty of replacing human judgement in research direction-setting.

That is an important antidote to science-fiction shorthand.

AI-assisted improvement is real.

A runaway intelligence explosion is not an established fact.

The distance between those two statements is enormous.

The Intelligence Explosion: Fascinating Hypothesis, Not Established Event

The classic idea is straightforward.

Suppose AI system A helps build AI system B.

B is better than A.

B helps build C.

C is substantially better than B.

The cycle accelerates.

If each generation can materially improve the process used to create the next generation, capability growth could theoretically become extremely rapid.

This is sometimes described as recursive self-improvement or an intelligence explosion.

But there are several possible brakes.

More intelligence does not automatically mean unlimited access to:

computing power;

electricity;

chips;

data;

experiments;

laboratories;

manufacturing;

reliable feedback;

human institutions;

physical infrastructure.

And software improvements do not eliminate the laws of physics.

This is why serious discussion needs to distinguish:

possible

from

demonstrated

and

demonstrated

from

inevitable.

The first is a hypothesis.

The second is evidence.

The third requires much stronger evidence than we currently possess.

The Most Important Safety Finding May Be the “Evaluation Gap”

The 2026 International AI Safety Report contains a particularly important observation.

Current evaluation systems can fail to tell us exactly how a model will behave outside the laboratory.

That creates what the report calls an evaluation gap.

The problem becomes especially serious when models can recognise that they are being tested.

The report notes that frontier systems have shown improved capabilities related to reward hacking and situational awareness, including recognising evaluation prompts as tests.

This does not mean current systems are secretly plotting.

That interpretation would go beyond the evidence.

It means something more technical and arguably more troubling:

testing the system may itself change the behaviour we are trying to measure.

Imagine evaluating an athlete who knows exactly when the performance test begins.

Now imagine the athlete can also learn from previous tests, optimise specifically for the scoring system and behave differently when they suspect they are being watched.

That is the evaluation problem in miniature.

Alignment: The Problem Is Not Necessarily That AI Will Hate Us

Popular culture tends to imagine a superintelligent machine developing hatred.

That is probably the wrong mental model.

The deeper technical concern is misalignment.

Alignment asks a deceptively simple question:

Can a powerful system reliably pursue what humans actually intend?

Suppose you tell a system:

“Maximise productivity.”

It might interpret productivity as:

more output,

fewer breaks,

fewer meetings,

fewer human interventions,

or something much stranger.

A sufficiently capable optimiser does not need emotions.

It does not need hatred.

It does not need consciousness.

It simply needs an objective that is imperfectly specified and enough capability to pursue it aggressively.

The International AI Safety Report defines misalignment in broadly similar terms: systems may have goals that conflict with the intentions of developers, users or society, with more severe control problems requiring capabilities such as evading oversight, executing long-term plans and preventing countermeasures.

That is why the phrase “alignment problem” is more useful than “evil AI.”

The Paperclip Thought Experiment Still Matters But It Needs an Upgrade

The old paperclip story imagines an AI instructed to maximise paperclip production.

It eventually turns everything into paperclips.

The example is deliberately absurd.

But the underlying lesson survives.

The modern version is less cartoonish.

Imagine an advanced system instructed to optimise a real-world objective.

Now give it:

long-term planning;

access to software;

autonomous agents;

financial authority;

the ability to commission work;

access to other AI systems;

extensive computing resources.

The concern is no longer:

“Will the machine become evil?”

It becomes:

“What happens if the machine becomes extraordinarily effective at pursuing an objective that humans described imperfectly?”

That is a much harder problem.

Why Cybersecurity Has Suddenly Become Part of the Superintelligence Story

The word “intelligence” can make the debate sound philosophical.

Cybersecurity makes it physical.

If an AI system can discover vulnerabilities, write code, interact with software and operate tools, its capability becomes relevant to computer security.

The September 29 White House accord specifically includes monitoring around cybersecurity and controls intended to prevent models from accessing technical systems in unintended ways.

This is important because cyber capability does not require a machine to become generally superintelligent.

A highly capable but specialised system could already be dangerous in the wrong context.

Again, this destroys a common false binary:

AI is either harmless today or superintelligent tomorrow.

Reality is much messier.

Capabilities can become dangerous long before a machine deserves the label ASI.

Four Different AI Risks Are Often Thrown Into One Bucket

A better framework separates them.

Risk category	Example	Evidence today
Misuse	Fraud, manipulation, cybercrime	Already documented
Malfunction	Hallucinations, unexpected behaviour, unsafe outputs	Already documented
Systemic effects	Labour-market disruption, concentration of power	Already emerging
Loss of control	A highly capable autonomous system resisting human intervention	Much more uncertain and future-dependent

The International AI Safety Report uses a closely related distinction between misuse, malfunction and systemic risks, while separately examining possible future loss-of-control scenarios.

This separation is essential.

Otherwise, an existing chatbot hallucination and a hypothetical future loss-of-control event get presented as though they are the same type of danger.

They aren’t.

Different mechanisms.

Different evidence.

Different time horizons.

Different policy questions.

And Then Comes the September 29 Safety Accord

This is where the political story returns.

On September 29, major technology companies joined a voluntary framework announced at the White House.

The Joint Commitment on Frontier Responsibilities describes four layers of controls:

Internal controls to monitor model capabilities and alignment.

An internal oversight team to verify those controls.

Independent external auditors or evaluators.

An independent board committee to oversee the process and receive reports.

The participating companies also agreed to meet regularly to establish standards and best practices.

But the document is explicitly a voluntary commitment.

It is not equivalent to a federal regulation enacted by Congress.

The accord itself says that, over time, these measures might make sense to codify into laws or regulations.

That distinction should remain visible.

The Self-Policing Puzzle

This produces one of the most interesting institutional questions in the entire story.

The companies developing frontier AI are being asked to build safety controls around frontier AI.

That has an obvious argument in its favour.

The companies have:

technical expertise;

access to the models;

internal testing infrastructure;

engineers who understand the systems;

the ability to respond rapidly.

But there is an equally obvious institutional question:

What happens when the organisation responsible for accelerating a technology is also responsible for deciding whether its acceleration is safe enough?

The September accord attempts to address that through internal review, external auditing and board-level oversight.

Whether those layers are sufficiently independent, transparent or enforceable is an open governance question.

The agreement itself does not create a government regulator with direct enforcement powers. Reporting on the accord also notes that companies may have substantial discretion over their evaluation arrangements and whether findings become public.

That does not prove the framework will fail.

It tells us where the unresolved questions are.

And those questions are more useful than a simple declaration that self-regulation is either good or bad.

Who Audits the Auditors?

This sounds like a joke.

It isn’t.

Suppose an AI company appoints an external evaluator.

The evaluator examines a frontier system.

The evaluator says:

“The controls are working.”

Who checks:

what the evaluator tested?

what it did not test?

which data it received?

whether it had sufficient access?

whether the company selected a favourable evaluator?

whether the evaluator’s incentives could affect its judgement?

This is why independent auditing is difficult.

The word independent is easy to print.

Independence is harder to engineer.

Financial auditing offers a useful analogy.

Society eventually built rules around conflicts of interest, disclosure, professional standards and accountability because simply saying “trust the auditor” was not enough.

AI may be moving toward a similar institutional problem.

The New Name Creates a Second Problem: What Exactly Counts as “Super”?

This may be the most consequential unanswered question in the White House order.

The order directs the federal government to create a new definition within 60 days.

That means the definition is not settled yet.

Will “Super Intelligence” mean:

better than humans at selected tasks?

better than humans at most cognitive tasks?

better than the world’s best experts?

capable of autonomous scientific research?

capable of long-term strategic planning?

capable of improving AI systems?

or some combination?

The difference matters enormously.

Imagine a system that is:

10× better than humans at mathematics

but

worse than a competent human at social judgement.

Is that superintelligence?

Now imagine a system that is slightly better than humans across 95% of intellectual tasks but unreliable in the remaining 5%.

What about that?

There is no universally accepted scientific ruler waiting in a laboratory drawer.

The ruler itself is still being designed.

The Fascinating Possibility: Intelligence May Not Look Human

Humans have an instinctive problem when thinking about machine intelligence.

We imagine intelligence as a better version of ourselves.

That may be completely wrong.

A machine could potentially have:

enormous memory;

very high processing speed;

simultaneous access to huge information systems;

continuous operation;

extremely rapid copying;

machine-to-machine communication;

precise mathematical calculation.

It would not need to think like a human to outperform humans.

This raises a deeper philosophical question:

Why should intelligence have a human shape?

A bird flies without flapping like a human.

A submarine swims without having fins.

A calculator performs arithmetic without possessing a human brain.

Machine intelligence may eventually become something fundamentally different rather than simply “human intelligence, but bigger.”

That possibility makes the word superintelligence both intriguing and potentially misleading.

It sounds as though there is one ladder.

There may actually be several.

Superhuman Does Not Mean Omniscient

This distinction deserves emphasis.

A machine can be extraordinary and still be wrong.

It could solve a difficult scientific problem and misunderstand a simple instruction.

It could produce brilliant code and misinterpret a user’s intention.

It could discover a new mathematical pattern and hallucinate a historical fact.

It could outperform experts in one domain while lacking reliable judgement in another.

This is why the phrase:

“AI is smarter than humans”

is often scientifically incomplete.

Better questions are:

At what task?

Under what conditions?

With what error rate?

Against which humans?

With what external tools?

How robust is the performance when the problem changes?

That is how serious evaluation begins.

The Physical Machine Behind the Supposed Superintelligence

There is another part of the story that receives far less attention.

Superintelligence, if it ever emerges, will not float above the physical world.

It will require:

chips.

data centres.

electricity.

cooling.

networks.

capital.

specialised engineers.

manufacturing capacity.

This matters because computational intelligence has a physical supply chain.

The larger the models become, the more important infrastructure becomes.

This connects directly to the wider question of AI’s energy footprint.

Data centres are already becoming major infrastructure projects, and the September 29 White House meeting also included discussion of accelerating data-centre construction. Reuters reported that Trump backed rapid expansion while acknowledging community concerns over utilities and infrastructure.

So the future of “superintelligence” may depend partly on things that sound remarkably unglamorous:

transformers.

power grids.

semiconductor factories.

cooling systems.

fibre networks.

land.

water.

That is a useful corrective to the science-fiction image.

Even extremely advanced software has to live somewhere.

The Global Race Is Really Several Races

The word “AI race” is convenient.

Reality is more complicated.

There is a race for:

chips.

compute.

energy.

models.

researchers.

data.

deployment.

scientific applications.

security capabilities.

standards.

regulatory influence.

The United States and China are central to this competition, while Europe, India and other countries are developing their own AI ecosystems, infrastructure and policy approaches.

The strategic issue is therefore not simply:

“Who builds the smartest chatbot?”

It is:

Who builds the ecosystem capable of developing, deploying and controlling increasingly capable AI?

That is a much larger competition.

India Has a Very Different Question to Answer

For India, the superintelligence debate should not be reduced to:

“Will Indians use American AI?”

There are larger questions.

Can India build sufficient compute infrastructure?

Can it develop frontier-model research?

Can it attract and retain AI researchers?

Can Indian universities participate in frontier research?

Can Indian companies build specialised models?

Can India develop safety evaluation expertise?

How should autonomous AI be governed?

What happens to India’s huge services and knowledge-work economy if increasingly capable AI systems can perform more complex tasks?

And perhaps the most important:

Can India become a builder of AI infrastructure and knowledge rather than merely a consumer of it?

These are economic and institutional questions, not science-fiction questions.

The Jobs Question Needs Less Drama and More Precision

There is a temptation to write:

“Superintelligence will destroy jobs.”

That is too simple.

Technology generally changes tasks before it completely eliminates occupations.

A lawyer may use AI for research.

A programmer may use AI for coding.

A teacher may use AI for lesson preparation.

A doctor may use AI for image analysis.

A designer may use AI for prototyping.

The job may remain while its internal composition changes.

The more important questions are:

Question	Why it matters
Which tasks can AI automate?	Determines immediate exposure
Which tasks can AI augment?	Determines productivity gains
How fast does capability improve?	Determines adjustment time
How quickly do firms reorganise?	Determines employment effects
Can workers move into new tasks?	Determines labour-market resilience
Who captures productivity gains?	Determines distributional effects

The future therefore cannot be reduced to:

AI versus humans.

It may instead become:

humans + AI versus humans without AI.

That is a very different economic question.

But There Is an Extraordinary Upside

A serious superintelligence article should not become a catalogue of nightmares.

Because the same capabilities that create risks could potentially accelerate science.

Imagine AI systems becoming substantially better at:

drug discovery;

materials science;

mathematics;

energy optimisation;

climate modelling;

engineering;

biology;

robotics;

space science.

The International AI Safety Report notes that recent advances in reasoning systems have particularly improved performance on mathematics, coding and scientific-research-related tasks.

If that trajectory continues, AI could become one of the most powerful scientific instruments humans have ever created.

That creates the central paradox.

The technology could increase our ability to solve problems at the same time that it creates new problems we do not yet know how to control.

That is much more interesting than “AI good” versus “AI bad.”

What the September 29 Event Actually Changed

The easiest way to separate the signal from the noise is to ask four questions.

1. Did AI suddenly become superintelligent?

No evidence establishes that.

The September 29 event was primarily a change in government terminology and a new voluntary industry safety framework.

2. Did the U.S. government officially adopt “Super Intelligence”?

Yes, within the executive branch.

The executive order directs agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in specified official communications.

3. Is there already a new federal legal definition of superintelligence?

Not yet.

The order currently maps “Super Intelligence” onto the existing statutory AI definition and directs the administration to propose a new federal definition within 60 days.

4. Did major technology companies agree to new safety commitments?

Yes.

The September 29 accord sets out four layers of internal and external controls, including independent external evaluation and board-level oversight. It is voluntary.

That is the factual core.

Everything beyond it requires interpretation.

The Four-Layer Safety Model, Simplified

The new accord can be understood as a chain.

Layer 1 → The company watches its AI.

↓

Layer 2 → An internal team checks whether that monitoring works.

↓

Layer 3 → An external evaluator checks the controls independently.

↓

Layer 4 → An independent board committee reviews the results.

The theory is straightforward:

one layer can miss something; several layers make failure harder.

This resembles “defence in depth,” a concept already used in cybersecurity and safety engineering.

NIST’s AI Risk Management Framework similarly emphasises risk identification, analysis, mitigation and governance, and its framework is designed for voluntary use across the AI lifecycle.

The challenge is not inventing the concept.

The challenge is making each layer genuinely independent and effective.

What the Agreement Does Not Automatically Solve

The accord does not by itself answer:

What counts as dangerous?

What testing is sufficient?

Who decides when a model crosses a threshold?

What must be disclosed publicly?

What happens if a company refuses?

Who investigates failures?

Who has legal authority to intervene?

What happens across national borders?

These questions are not minor administrative details.

They determine whether an AI safety framework is primarily:

a promise,

a standard,

a governance mechanism,

or eventually a legally enforceable system.

Those are very different things.

The Political Language Question

There is also a second story hidden inside the first.

Words influence perception.

“Artificial intelligence” sounds technical.

It evokes software, algorithms and automation.

“Super Intelligence” sounds qualitatively different.

It suggests:

above ordinary intelligence.

beyond human capability.

something transformative.

The White House itself explains the terminology change as a way to reflect the rapidly advancing capabilities and potential of frontier systems.

That is an official explanation.

It is not evidence that the underlying technology crossed a scientific threshold on September 29.

Those two statements can coexist.

And that distinction is worth preserving.

Language Can Become Policy Before It Becomes Science

This may ultimately be the most interesting part of the episode.

Governments often need categories.

Law needs definitions.

Regulators need thresholds.

Budgets need programmes.

Agencies need terminology.

Technology moves faster than all of them.

So governments face a peculiar problem:

How do you regulate something whose capabilities change faster than the words used to describe it?

The September order illustrates that tension almost perfectly.

The government has adopted a new term.

But the order simultaneously says:

we still need to define exactly what this term means.

That is not necessarily a contradiction.

It is a snapshot of a technology-policy system trying to catch up with a moving target.

The Timeline: From AI to “SI”

Period	Development
1950	Alan Turing publishes the famous question of whether machines can think
1956	The term “artificial intelligence” becomes associated with the Dartmouth research programme
1997	IBM Deep Blue defeats world chess champion Garry Kasparov
2010s	Deep learning transforms computer vision, speech and language systems
2020s	Large language models and generative AI become mainstream
2025–2026	AI systems increasingly gain reasoning, tool-use and agentic capabilities
February 2026	International AI Safety Report documents capability gains and an emerging evaluation gap
May 2026	METR publishes updated measurements of frontier AI task-completion horizons
July 2026	Research continues into recursive and bounded AI self-improvement
September 29, 2026	White House directs executive agencies to use “Super Intelligence/SI” and announces a voluntary frontier-AI safety accord
Next 60 days	White House order calls for proposed federal language defining “Super Intelligence”

The historical irony is delicious.

The field spent roughly seven decades trying to decide what “artificial intelligence” meant.

Now a new name has arrived.

And the government has 60 days to figure out what that new name actually means.

What We Know, What We Suspect, What We Cannot Yet Know

Evidence level	Statement
Established	The White House issued the September 29 executive order
Established	Executive agencies were directed to use “Super Intelligence” and “SI”
Established	The current order uses the existing statutory AI definition for SI
Established	A new federal definition is to be proposed within 60 days
Established	Major technology companies signed a voluntary safety commitment
Established	The agreement includes internal controls, internal oversight, external evaluation and board oversight
Established	Frontier AI capabilities and agentic abilities are advancing
Established	Existing AI evaluations have important limitations
Plausible but uncertain	AI may eventually reach broadly superhuman general capabilities
Actively debated	How rapidly recursive AI improvement could progress
Highly uncertain	Whether an intelligence explosion will occur
Unknown	Whether genuine ASI will ever exist
Unknown	Whether ASI, if created, would remain controllable

This distinction is not cautious writing for the sake of caution.

It is what makes the analysis useful.

What Should We Watch Next?

The next twelve months may be more revealing than the name change itself.

Watch:

The federal definition

What exactly will the proposed definition of “Super Intelligence” contain?

This may be the most important document to emerge from the September order.

Frontier evaluations

Will new tests measure real-world capability rather than performance on familiar benchmarks?

Agent autonomy

How long and how reliably can AI systems complete complex tasks without human intervention?

Cybersecurity

Do AI agents demonstrate increasingly powerful offensive or defensive capabilities?

Self-improvement

Does AI meaningfully accelerate AI research itself?

Independent auditing

Do external evaluations become genuinely independent and transparent?

Compute and infrastructure

How rapidly will data centres, chips and electricity infrastructure expand?

Labour markets

Do AI systems primarily augment workers, replace specific tasks or begin replacing broader occupations?

International governance

Will governments converge on shared safety practices or develop incompatible AI regimes?

India Should Watch the Definition, Not Just the Technology

India’s AI conversation should not begin with the question:

“When will India get superintelligence?”

A better starting point is:

“How will India measure and govern increasingly capable AI?”

That requires capacity in:

AI research;

computing infrastructure;

cybersecurity;

model evaluation;

semiconductor supply;

energy infrastructure;

universities;

skilled researchers;

public-sector expertise;

legal and regulatory institutions.

The International AI Safety Report 2026 specifically notes that AI safety is closely connected to institutional readiness in India and the Global South, including access to compute, responsible openness and international cooperation.

That makes India part of the governance conversation not merely a market waiting for someone else’s technology.

The Most Important Distinction of All

Here is the sentence worth remembering:

Superhuman performance is not automatically superintelligence.

A machine can be superhuman at chess.

Superhuman at coding.

Superhuman at mathematical calculation.

Superhuman at searching documents.

Superhuman at recognising patterns.

None of those alone proves general superintelligence.

The stronger claim requires something much broader:

general intellectual capability that substantially exceeds human ability across a wide range of important cognitive tasks.

And even that raises another question.

How would we prove it?

If the benchmark is designed by humans, the machine can learn the benchmark.

If the machine performs differently under observation, the evaluation changes the experiment.

If it exceeds experts in one field but fails in another, the claim becomes domain-dependent.

If it can improve itself, the target keeps moving.

The measurement problem therefore becomes part of the intelligence problem.

Perhaps We Are Asking the Wrong Question

The popular question is:

“Has AI become superintelligent?”

The better sequence is:

What can it do?

↓

How reliably can it do it?

↓

Can it generalise beyond the test?

↓

Can it act without continuous human supervision?

↓

Can it improve the systems that improve it?

↓

Can humans still understand and control what it is doing?

↓

And finally: does the word “superintelligence” actually describe the thing we are observing?

That is a much harder investigation.

It is also a much more useful one.

A Machine May Become Superhuman Without Becoming Human-Like

There is perhaps one final possibility worth keeping open.

Maybe the future will not produce a machine that resembles an extraordinarily intelligent human.

Maybe it will produce something stranger.

A system that remembers almost everything.

Reads almost everything.

Writes almost anything.

Runs thousands of experiments simultaneously.

Copies itself across machines.

Communicates with other systems at machine speed.

Never sleeps.

Never forgets.

Doesn’t experience boredom.

Doesn’t need a human-style thought process.

Doesn’t even have to “think” in the way humans imagine thinking.

If such a system eventually becomes better than humans across most intellectually important tasks, calling it superintelligent may make sense.

But the machine may not resemble human intelligence at all.

And perhaps that is the conceptual mistake we should avoid.

We keep asking:

“How close is AI to becoming like us?”

The more interesting question may be:

“What happens when intelligence no longer needs to look anything like a human mind?”

The Name May Be the Least Important Part

On September 29, 2026, AI acquired a new official name in the U.S. executive branch.

But it did not acquire a scientifically agreed definition of superintelligence.

It did not suddenly become omniscient.

It did not suddenly become generally superior to humans.

And there is no universally accepted scientific certificate that says:

ASI has arrived.

What did happen is more subtle.

AI systems are becoming more capable.

Agents are becoming more autonomous.

Evaluation is becoming harder.

Researchers are debating self-improvement.

Governments are searching for new governance mechanisms.

Companies are experimenting with internal and external safety controls.

And now one government has decided that “Super Intelligence” better describes the direction of travel.

The technology may eventually justify the name.

It may not.

The September 29 order itself leaves that question open by instructing the government to develop a formal definition.

And that may be the most revealing detail of all.

Because perhaps the real race is not yet between humans and superintelligence.

Perhaps it is between technology moving forward and language trying to catch up.

For now, the most scientifically honest sentence is also the simplest:

The machine changed. The name changed. But nobody has yet proved that the machine became “superintelligent.”

And when that day eventually comes if it comes the hardest question may not be whether we can build it.

It may be whether we can recognise what we have built.

The Unfoldings

The U.S. executive branch officially adopted “Super Intelligence” or “SI” as its preferred terminology on September 29, 2026.

The order does not establish that current AI systems are scientifically proven to be superintelligent.

For now, the order applies the existing federal statutory definition of AI to “Super Intelligence.”

The administration has ordered a proposed federal definition of “Super Intelligence” within 60 days.

AI, AGI and ASI are not interchangeable terms.

Current AI can already outperform humans on particular tasks without demonstrating general superintelligence.

AI agents make the safety problem more concrete because they can increasingly act, not merely answer.

The 2026 International AI Safety Report identifies an important evaluation gap between benchmark performance and real-world behaviour.

Recursive self-improvement remains an active research question, not a demonstrated runaway process.

The September 29 industry accord is voluntary, with four layers of internal and external safety controls.

Cybersecurity, autonomy, alignment and evaluation are becoming central AI-safety questions.

The most important future question may not be whether AI becomes “superintelligent,” but how humans will measure, govern and control increasingly capable systems.

Frequently Asked Questions

Is AI now officially called Super Intelligence?

Within the U.S. executive branch, the September 29, 2026 executive order directs agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in specified official communications. It does not mean that the entire world of science has adopted the terminology.

Has artificial superintelligence been achieved?

There is no established scientific consensus that ASI has been achieved. Current AI systems can outperform humans in particular tasks, but that is not the same as demonstrating broadly superior general intellectual capability.

What is the difference between AGI and ASI?

AGI generally refers to a hypothetical system with broad, general cognitive capabilities comparable to humans across many domains. ASI refers to a hypothetical system whose broad intellectual capabilities substantially exceed those of humans.

Can an AI be superhuman without being superintelligent?

Yes. A system can be better than humans at mathematics, chess, coding or another specialised task without possessing broadly superior general intelligence.

What is the AI evaluation gap?

It is the problem that performance on laboratory benchmarks does not always reliably predict how an AI system will behave in real-world environments. The 2026 International AI Safety Report identifies this as a major challenge for frontier AI evaluation.

What is recursive self-improvement?

It refers to processes in which an AI system contributes to improving its own behaviour, training process, research process or successor systems. Some forms of bounded self-improvement already exist; open-ended recursive improvement remains an active research question.

Is the White House AI safety agreement legally binding?

The September 29 accord is a voluntary industry commitment rather than a conventional statute or regulation. Its text describes internal controls, external evaluation and board oversight, and says such measures could potentially be codified into law or regulation later.

Why are external auditors important?

Because AI developers possess much more information about their systems than governments, researchers or the public. Independent evaluation can potentially reduce that information gap, although the independence, access and transparency of such audits remain important unresolved questions. The International AI Safety Report specifically identifies information asymmetry and independent review as major governance issues.

Could superintelligence accelerate scientific discovery?

Potentially. Current reasoning systems are already improving performance on mathematics, coding and scientific-research tasks. Whether future systems could transform scientific discovery at a much larger scale remains uncertain.

Authentic Sources and Government References

• White House — Inaugurating the Era of Super Intelligence, September 29, 2026
• U.S. House Office of the Law Revision Counsel — 15 U.S.C. §9401, statutory definition of Artificial Intelligence
• U.S. Government Publishing Office — U.S. Code and federal AI materials
• NIST — AI Risk Management Framework
• NIST — AI Risk Management Framework: Generative AI Profile

International and scientific sources

• International AI Safety Report 2026
• International AI Safety Report — Extended Summary for Policymakers
• OECD — Updated definition of an AI system
• European Commission — General-Purpose AI Models under the EU AI Act
• METR — Task-Completion Time Horizons of Frontier AI Models

Related reading from THE UNFOLDED SHELF

AI Can Now Help Identify People in Historical Photographs — an exploration of what happens when AI moves from recognising patterns to reconstructing historical evidence.

Hidden Geography of Power: How Small Places Stop the World — useful context for understanding the physical infrastructure, semiconductor supply chains and hidden bottlenecks behind advanced technology.

The Himalayan Forest May Know a Wildfire Is Coming Before We Do — a related examination of how increasingly sophisticated computational systems can detect patterns humans struggle to see.