WHO GETS TO DECLARE AGI? Evidence, Incentives, and Threshold Authority After GPT-6 Astra
Synthocracy Institute — P0 / #17
Martin Novak
September 2026
Evidence Boundary
This paper distinguishes three forms of claim. (A) Empirical claims describe documented statements, agreements, benchmark results, regulatory structures, and institutional arrangements. (B) Argumentative claims interpret what those facts mean for governance. (C) Foresight claims examine plausible consequences if the same institutional pattern extends toward more capable systems. The paper does not attempt to settle whether GPT-6 Astra is AGI. It asks a different question: who has standing to make such a declaration consequential?
The Declaration
On 3 September 2026, OpenAI released GPT-6 Astra, describing it as its most intelligent and aligned model yet. The published results were extraordinary: 97.6% on FrontierMath Tier 4, 100% on ExploitBench, major advances in computer use, and a reported 99.9% result on ARC-AGI-3 under one evaluation configuration. OpenAI President Greg Brockman ended a press briefing with a phrase designed to travel much further than any benchmark table: “Welcome to the AGI era.” (OpenAI)
Three days later, NVIDIA CEO Jensen Huang went further. He publicly declared that “AGI has arrived.” The statement attracted immediate attention because Huang is not an outside observer of the technological transition. NVIDIA supplies much of the compute infrastructure on which frontier AI is being built, and in February 2026 OpenAI announced $30 billion of new investment from NVIDIA as part of a $110 billion financing round. Huang therefore occupies an unusual position: he is simultaneously one of the most influential interpreters of the AI transition and one of its largest economic participants. (Yahoo Finanse)
None of this proves that Huang is wrong. An economic interest does not invalidate an empirical claim. It does, however, reveal the governance problem hiding behind the headline.
Who gets to declare that AGI has arrived?
1. A Declaration Is Not the Same Thing as a Measurement
[A — empirical] There is no single agreed technical definition of AGI. OpenAI’s Charter defines it as highly autonomous systems that outperform humans at most economically valuable work. ARC Prize uses a different concept: the ability to acquire skills humans can acquire, with comparable efficiency. These definitions overlap, but they are not equivalent. One emphasises economically valuable performance; the other emphasises efficient acquisition of general skills. (OpenAI)
The distinction matters because a threshold cannot be measured independently of the thing the threshold is supposed to mean.
A system may exceed humans across large parts of professional work while remaining weak at open-ended adaptation. Another may perform brilliantly in novel environments but remain unreliable across real organisations. A system may be economically transformative before satisfying a cognitive definition of general intelligence. A benchmark may become saturated before the phenomenon it was designed to approximate has been settled.
[B — argument] AGI is therefore not currently analogous to a temperature at which water boils. It is closer to a contested institutional category whose meaning depends partly on the test, partly on the definition, partly on the use case, and partly on who is authorised to interpret the evidence.
That changes the nature of the problem.
The question is no longer only:
Has the threshold been crossed?
It becomes:
Which threshold, defined by whom, measured how, verified by whom, and with what consequences?
2. The Astra Result Demonstrates the Problem
The ARC-AGI-3 results provide an unusually clean example.
[A — empirical] GPT-6 Astra achieved 62.7% on the ARC-AGI-3 Semi-Private set using ARC Prize’s provider-neutral Standard harness. With a Provider Adapter that preserved opaque reasoning state between requests and used provider-specific context management, Astra reached 99.9%. ARC Prize describes both as legitimate but different evaluation questions. (ARC Prize)
The difference is not small.
It is the difference between 62.7% and 99.9%.
ARC Prize also found that Astra used fewer actions than the median tested human on 96% of levels and described the model as a step-function change in frontier capabilities. Yet the organisation explicitly states that saturating ARC-AGI-3 does not constitute proof of AGI and that it is not claiming Astra is AGI. The benchmark operates inside bounded, deterministic environments and does not capture the open-ended complexity of the real world. (ARC Prize)
That is not a minor methodological footnote.
It reveals something fundamental about intelligence claims.
[B — argument] A capability score is not simply a property of a model. It can also be a property of the model–harness–compute–memory–tool environment.
The statement “Astra scored 99.9%” is empirically correct under one configuration. The statement “Astra scored 62.7%” is also empirically correct under another. Neither number can independently carry the burden of the sentence “AGI has arrived.”
The number requires an interpreter.
And the interpreter exercises authority.
3. The Surprising Part: There Is Already an AGI Declaration Procedure
At first glance, one might conclude that nobody has formal authority to declare AGI.
That would be too simple.
[A — empirical] The October 2025 agreement between OpenAI and Microsoft contains an explicit AGI determination mechanism. The published terms state that once OpenAI declares AGI, the declaration is to be verified by an independent expert panel. The determination has real contractual consequences, including consequences for research intellectual-property rights and, under the agreement as published at the time, revenue-sharing arrangements. In February 2026, OpenAI and Microsoft stated that the contractual AGI definition and determination process remained unchanged. (OpenAI)
This is an extraordinary institutional fact.
AGI is frequently discussed as if it were a scientific milestone waiting to be discovered. But one of the world’s most consequential AGI thresholds already exists inside a private contract between two corporations.
OpenAI has the initiating authority.
An independent expert panel has verification authority.
The resulting determination can change commercial and intellectual-property relationships worth enormous sums.
That procedure may be entirely appropriate for the contract it governs. But it does not make the panel a global authority on whether humanity has entered the AGI era.
This gives us a crucial distinction.
Contractual AGI answers:
Has the threshold specified by this agreement been reached for the purposes of this agreement?
Scientific AGI asks:
Does the available evidence justify describing the system as possessing whatever scientific community means by general intelligence?
Regulatory AGI would ask:
Has a legally defined threshold been reached that activates specified public obligations?
Public AGI asks something looser and more powerful:
Will governments, businesses, investors, media and citizens begin behaving as though AGI has arrived?
These are different questions.
They can produce different answers at the same time.
4. The Law Is Already Avoiding the Word
The European Union provides an instructive contrast.
[A — empirical] The AI Act does not need to determine whether a system is AGI. It defines general-purpose AI models and separately establishes a category of general-purpose AI models with systemic risk. A model can enter that category because of high-impact capabilities or through a European Commission decision following, among other routes, a qualified alert from the scientific panel. The Act also creates a rebuttable capability presumption at a cumulative training-compute threshold above floating-point operations. (Eur-Lex)
This is institutionally revealing.
The regulator does not have to answer the metaphysical question:
Is this AGI?
It can ask the governance question:
Does this system possess capabilities or reach that justify additional obligations?
That may prove to be the more durable model.
A civilisation does not necessarily need universal agreement about what AGI is before deciding that certain capabilities demand stronger oversight, reporting, testing, containment or access controls.
[B — argument] The most important governance thresholds may therefore arrive before consensus on the intelligence threshold.
5. Threshold Authority
Synthocracy usually studies what happens when decision-making power moves upstream: into filtering, ranking, modelling, routing, recommendation and execution.
AGI adds another upstream layer.
Before institutions can respond to AGI, someone has to determine that the category applies.
We can call that power threshold authority.
Threshold authority is the capacity to make a technological threshold institutionally consequential by defining it, measuring it, verifying it, recognising it, or acting upon its declaration.
The term passes the Institute’s translation test because the underlying mechanism is simple. Someone has to decide when the line has been crossed.
Today, that authority is fragmented.
| Form of authority | Typical actor | What it can determine |
|---|---|---|
| Scientific authority | researchers, benchmark organisations, scientific communities | whether evidence supports a capability claim |
| Corporate authority | frontier labs and executives | how a system is described and deployed |
| Contractual authority | companies, boards, expert panels | whether a defined threshold activates contractual consequences |
| Regulatory authority | legislatures, agencies, scientific panels | whether legal obligations apply |
| Market authority | investors, customers, insurers, employers | whether actors behave as if the threshold has been crossed |
| Public narrative authority | media, public figures, platforms | which interpretation becomes socially dominant |
No one column necessarily dominates all the others.
That is the core governance problem.
6. When the Seller Helps Name the Era
The Huang declaration makes this institutional fragmentation visible.
[A — empirical] NVIDIA is deeply embedded in the economics of frontier AI. OpenAI announced a $30 billion NVIDIA investment in February 2026 alongside continuing access to next-generation NVIDIA inference compute. NVIDIA also benefits more broadly from increased demand for the computational infrastructure required to train and operate frontier systems. (OpenAI)
Huang is therefore both observer and participant.
That does not make his assessment worthless. Few people are better positioned to see the computational trajectory of the industry.
But evidence governance should distinguish proximity to evidence from independence from outcome.
The person closest to the technological transformation may have the best information and the strongest incentives simultaneously.
That is normal in emerging industries. Aircraft manufacturers know their aircraft. Pharmaceutical companies know their medicines. Nuclear operators know their reactors.
Modern institutions responded to precisely this problem by separating, to varying degrees, the roles of builder, evaluator, certifier and regulator.
AI has not yet developed an equivalent architecture for AGI.
7. The Declaration Can Become True Politically Before It Is Settled Scientifically
This is where the question becomes larger than terminology.
Imagine that scientists remain divided over whether Astra should count as AGI.
Markets do not have to wait.
Governments do not have to wait.
Employers do not have to wait.
Defence ministries do not have to wait.
Universities do not have to wait.
Boards allocating billions of dollars to compute infrastructure do not have to wait.
If enough consequential institutions begin behaving as though AGI has arrived, the declaration begins producing real effects independently of scientific consensus.
[B — argument] A technological threshold can become institutionally real before it becomes epistemically settled.
Capital moves.
Procurement changes.
National-security assumptions change.
Labour strategies change.
Infrastructure is financed.
Regulatory pressure changes.
The next generation of systems is trained under a new strategic premise.
The category becomes performative.
This does not mean that saying “AGI” creates intelligence by language alone. It means that the classification can create institutional consequences before agreement about the classification is complete.
And those consequences can themselves accelerate the technological trajectory.
8. The AGI Declaration Protocol
The answer should not be to prohibit people from using the term AGI. Scientific and public debate requires competing definitions and interpretations.
The governance problem begins when a declaration is expected to carry consequences.
For consequential declarations, Synthocracy Institute should propose a minimal AGI Declaration Protocol. A serious declaration should disclose:
- Definition — the exact definition of AGI being used.
- Evidence — the capability evidence supporting the claim, including negative results and known residual weaknesses.
- Evaluation conditions — harnesses, tools, memory, compute budgets, scaffolding and human assistance used to produce the result.
- Independent verification — which findings have been reproduced or examined by parties independent of the developer.
- Interests and dependencies — commercial, contractual or institutional interests of the declarant and evaluators.
- Consequences — which legal, contractual, deployment, financial or policy changes the declaration is intended to trigger.
- Revision rule — how the determination can be challenged, updated or withdrawn as evidence changes.
This would not produce a universal metaphysical certificate saying AGI: TRUE.
It would do something more useful.
It would make the declaration inspectable.
9. From “Is It AGI?” to “What Does This Declaration Authorise?”
The most important move is therefore conceptual.
Public discussion treats AGI as a binary property:
before AGI / after AGI.
Governance should treat an AGI declaration as a claim that may activate different consequences in different systems.
The scientific community may recognise one threshold.
A private agreement may recognise another.
An insurance regime may care about autonomous cyber capability.
A government may care about strategic research acceleration.
A labour regulator may care about economically substitutable work.
A national-security organisation may care about autonomous vulnerability discovery.
A competition authority may care about infrastructure concentration.
There may never be one moment in which all of these institutions cross the same line simultaneously.
[B — argument] The search for one universal AGI declaration may therefore be the wrong institutional design problem. What societies need is a governance architecture for multiple consequential capability thresholds.
That is already approximately what the EU AI Act does by regulating capability and systemic impact without waiting for AGI consensus.
10. The Synthocratic Question
The foundational Synthocracy question is:
Where has decision-making power moved, and who can inspect, challenge or stop it?
The AGI debate adds a preliminary question:
Who decides that the world has entered the condition in which extraordinary governance responses become justified?
This is power before deployment.
Power before regulation.
Power before public adaptation.
It is the power to define the historical moment itself.
The person who declares AGI does not automatically govern society. But if the declaration changes what governments fund, what companies deploy, what investors finance, what workers fear, what militaries prepare for and what regulators accelerate, then the declaration participates in governance.
The naming of the threshold becomes part of the decision chain.
FORESIGHT — The ASI Problem
[C — foresight] The current dispute may be a rehearsal for a much harder problem.
Suppose a future frontier laboratory announces that its internal system has crossed from AGI into superintelligence.
The system cannot be broadly released for independent evaluation because doing so would itself be considered dangerous.
Its most consequential capabilities are classified or commercially secret.
External evaluators receive only restricted access.
The system may outperform the humans evaluating it in the domains relevant to its own assessment.
Governments are given confidential briefings.
Markets receive partial signals.
Infrastructure providers see extraordinary compute demand.
The public receives a declaration.
At that point, the question “Who gets to declare ASI?” becomes far more difficult than today’s AGI dispute because ordinary independent replication may itself be impossible or unsafe.
The institutional architecture built around AGI declarations therefore matters before ASI exists.
Today’s semantic disagreement may be tomorrow’s constitutional problem.
Conclusion — Anyone Can Say AGI
Jensen Huang can say that AGI has arrived.
Greg Brockman can welcome the world to the AGI era.
OpenAI can describe Astra as its most intelligent model.
ARC Prize can celebrate a capability step-change while refusing to call the result proof of AGI.
OpenAI can invoke a contractual AGI determination procedure whose declaration must be verified by an independent expert panel.
The European Commission can classify a general-purpose AI model as presenting systemic risk without determining whether it is AGI.
All of these statements can coexist because they perform different institutional functions.
That is the lesson.
There is no single AGI threshold because there is not yet a single institution, definition, measurement procedure or consequence attached to the term.
The governance question is therefore not simply whether AGI has arrived.
It is whether any actor should be able to transform its own interpretation of that arrival into consequences for everyone else without showing the definition, evidence, evaluation conditions, interests and route of verification.
The first principle of threshold governance should be simple:
Anyone can declare AGI. No one should be able to make that declaration consequential alone.
And that may be the more important threshold we have not yet crossed.
