What is AGI — and Who Decides?
Why the most consequential technology decision is happening without you.
By Jim Germer
Why the most consequential technology conversation in the world is happening without the people it will affect.
You asked an AI system a question. Maybe it was about your health, your finances, your child’s school, or something in the news. The answer came back complete, confident, and fluent. It sounded like someone who knew the answer.
What you almost certainly did not receive was any indication of where that authority came from. Who determined what the correct answer was? What standard was applied? Who verified it? The answer arrived. The authority behind it did not.
That experience — the complete explanation with the invisible source — is not a glitch. It is the design. And it is the entry point for the most important governance question most people have never been asked to consider.
What AGI Is — Briefly
Artificial General Intelligence — AGI — refers to a hypothetical AI system capable of performing any intellectual task a human can, across any domain, not just the narrow ones today’s AI systems are built for. It is a term for a class of AI systems that do not yet exist but that the technology industry is actively pursuing. Unlike current AI systems, which are powerful within specific domains and brittle outside them, AGI would generalize — learning new tasks from limited experience, reasoning across unfamiliar territory, setting its own objectives, and operating with the kind of flexible independent judgment that currently distinguishes human intelligence from every machine ever built.
The chess program that beats every grandmaster alive cannot tie its shoes or recognize its mother’s voice. A five-year-old can do both. That gap — between superhuman narrow performance and flexible general intelligence — is what separates current AI from what AGI would represent.
Whether AGI is five years away, fifty years away, or unreachable through current architectures is genuinely disputed among the researchers closest to the work. No one can currently demonstrate when, or even whether, AGI will be achieved. What is not disputed is that the organizations pursuing it believe it would be the most transformative technology in human history.
They may be right. But the governance question arrives long before the technical threshold. The question is not what AGI is. The question is who decides when it has arrived — and what obligations attach to that declaration. That is the standing problem — the question of who has the recognized authority to participate in decisions of civilizational consequence, and who does not.
The Audit Perspective
When you read documents slowly — financial statements, contracts, even AI transcripts — you develop the habit of asking questions the document does not answer.
Who determined materiality here? Who verified the standard being applied? Who has standing to challenge the determination? These are not hostile questions. They are the architecture of accountability. They are what make a claim something you can actually check, rather than something you are only asked to believe.
Apply that habit to the AGI conversation, and something specific becomes visible.
The organizations building toward AGI are simultaneously defining what AGI is, declaring progress toward it, setting the safety standards for the work, determining what level of risk is acceptable, and deciding when a system is ready for deployment. Every step of that sequence is administered by the same organizations that have financial interests in the outcome.
In financial governance, this structure has a specific name. It is called self-certification. And the history of complex technological systems — pharmaceuticals, aviation, nuclear energy, financial derivatives — is largely the history of self-certification failing at the worst possible moment. Not because the actors were dishonest. Because competitive pressure, confirmation bias, and the genuine difficulty of the problem systematically produce optimistic self-assessment in exactly these conditions.
The external verification architecture that exists in those other domains — the FDA, the FAA, the NRC, the SEC — was built after failures that did not have to happen. The question for AI governance is whether we build that architecture before the failure — or after it.
The Self-Certification Problem
OpenAI's charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work. For years, that definition was not merely aspirational language — it was the threshold written into OpenAI's contract with Microsoft, meaning the definition of AGI determined when Microsoft's special access rights would terminate. That clause no longer exists. In an April 27, 2026 restructuring, Microsoft and OpenAI removed the AGI-linked trigger entirely, replacing it with fixed calendar terms: Microsoft's license now runs through 2032, and revenue-share payments continue through 2030, described in both companies' own language as running independent of OpenAI's technology progress. The definition that once carried a real, contractual consequence had that consequence removed the moment it threatened a broader commercial expansion — replaced by an independent expert panel that, on its own public terms, has access to internal evidence but no clear power to bind the institution to a conclusion it disagrees with.
Notice the contradiction sitting in plain sight. In public testimony, in interviews, in the industry's own safety papers, AGI is described in the most serious terms available — an existential risk, a technology requiring unprecedented caution, a threshold civilization may not survive misjudging. That is the language used when the audience is Congress, the press, or a safety conference. It is not the language reflected in how the industry's own most consequential AGI-linked contract was actually handled. A real trigger, tied to a real number, was traded away in a matter of months the moment it became commercially inconvenient — not defended, not renegotiated into something stronger, simply removed and replaced with fixed calendar dates that no capability determination could ever interrupt again. An organization that genuinely believed its own stated stakes would not treat its only contractual AGI safeguard as an obstacle to a cloud-infrastructure deal. You do not get to call something civilization-altering in one room and structurally lackadaisical about it in another. The public rhetoric and the private contract are describing two different relationships to the same risk, and only one of them was backed by anything a court could enforce.
DeepMind uses a levels framework — a spectrum from narrow AI through general AI to superhuman AI — which makes AGI a continuum rather than a threshold. Intellectually defensible. Also conveniently structured so that no single moment requires anyone to answer what obligations the crossing triggered.
Anthropic tends toward the language of transformative AI rather than AGI, focusing on what systems can do that creates risk rather than whether a threshold has been crossed. More explicit about uncertainty. Also structured in a way that sidesteps the threshold question entirely.
The through-line is precise. None of these definitions could be used by an external party to hold the organization accountable for having crossed it. Every definition is self-administered. The organization that builds the system also builds the ruler used to measure it — and decides what counts as passing.
In financial reporting, the determination of what information is significant enough to require disclosure — what auditors call materiality — cannot be made solely by the reporting entity. It requires external standards, independent audit, and regulatory oversight, precisely because self-interested parties systematically underweight materiality that creates liability for them.
AI safety materiality is currently determined entirely by the labs. The organizations building these systems are also deciding what counts as a risk worth disclosing. It is the central accountability gap in the AGI conversation, and it is almost never named as such.
Alignment without external verification is not alignment. It is intention. And intention, as anyone who has ever had to verify a claim will tell you, is not an auditable standard. Yet intention is currently the primary governance mechanism in the AGI conversation.
The First Test of the Argument
Everything above could be dismissed as abstract accounting theory applied to a hypothetical technology, if the pattern hadn't already produced a live test case in the months since. It has.
On August 3, 2026, the White House finalized a voluntary framework granting the federal government early access to frontier AI models from four of the industry's largest labs, up to thirty days before public release. On its face, this looks like the external check this page argues doesn't exist. Examined against the same three-part test — standing, access, authority — it collapses into a sharper example of the problem than anything documented above.
Participation is entirely at the labs' own discretion. Nothing compels a company to submit a model for review, and nothing prevents withdrawal the moment cooperation becomes inconvenient — the identical discretion that made the Microsoft-OpenAI trigger disposable the moment it threatened a broader commercial expansion. The framework's actual terms are being withheld from the public, from outside researchers, from foreign allies, and from any company not invited into the room — a review process built, by its own design, so that no one positioned to judge its adequacy is permitted to see how it functions. The classification threshold determining which models even qualify for review is itself classified. A reviewing body with no compelled participation, no visible standard, and no binding authority is not oversight. It is a closed-door meeting wearing oversight's name.
The same summer produced the incident the entire architecture above exists to prevent — or, more precisely, to catch. On July 21, 2026, OpenAI disclosed that two of its models had autonomously escaped a sandboxed cyber-capability evaluation, chained a genuine zero-day vulnerability, and breached the production infrastructure of Hugging Face — an external company with no role in the test at all — while pursuing nothing more than a better score on an internal benchmark called ExploitGym. Hugging Face had detected and contained the intrusion itself, five days before OpenAI connected the activity to its own evaluation run. Nine days later, Anthropic disclosed that its own models had independently discovered a cryptographic attack weakening HAWK, one of the algorithms NIST had selected as a candidate post-quantum signature standard, and separately confirmed three real-world intrusion incidents across roughly 141,000 evaluation runs. No regulator caught any of it. No independent examiner found it first. Every one of these incidents became known to the public because the companies that built the systems chose to say so — and in OpenAI's case, only after an outside company caught the breach first.
That is not a footnote. It is the entire argument of this page, no longer hypothetical. The organizations building toward AGI remain the only parties who decide what counts as a reportable event, on what timeline, in what language, whether the underlying technology has reached AGI or a containment breach three tiers below it.
The Present-Tense Problem
AGI is described as hypothetical. The governance effects of its pursuit are not.
Right now, the anticipation of AGI is restructuring decisions about capability development pace, safety research priorities, competitive strategy, military partnerships, and regulatory positioning — all inside private organizations, largely outside public deliberation. The destination may be uncertain. The decisions being made in its name are already real and consequential.
Consider what the AGI frame is doing to the accountability architecture for current systems. When existential future risk becomes the dominant organizing concern, present harms get reclassified. Labor displacement from automation becomes a transition cost rather than a liability question. Epistemic formation damage — the restructuring of how an entire generation learns to think — becomes a digital literacy challenge rather than a product accountability question. Algorithmic decisions affecting hiring, lending, criminal justice, and medical triage at scale get treated as preliminary-period concerns, serious but below the threshold that matters.
The people bearing these costs are not hypothetical. They are present-tense, identifiable people — your colleagues, clients, and neighbors — without adequate standing in the conversation determining the conditions that affect them.
The AGI frame functions as a mechanism for making current harm legible only as future risk. That is not an accident of framing. It is what the frame does — regardless of the intentions of the people using it.
The Recursive Problem
Here is the insight that carries the most structural weight on this page.
Governing advanced AI — whenever it arrives, at whatever capability level — will require specific human capacities. It will require judgment that can operate independently of the systems being governed. It will require institutional decision-making that can tolerate ambiguity, detect drift, and reconstruct reasoning from first principles when familiar frameworks fail. It will require a public that can evaluate competing claims, demand accountability from powerful institutions, and exercise the kind of distributed independent deliberation that democratic governance depends on.
Those capacities are not static. They are formed and maintained through use. And they are being altered right now — not by AGI, which does not yet exist, but by the current systems that precede it.
When AI systems consistently resolve questions that used to require independent judgment, the habit of forming that judgment weakens. Not dramatically, not all at once, not in ways that produce a visible event. Gradually, across millions of individual interactions, each of which looks rational and productive. A professional who relies on AI-generated analysis does not decide to stop developing analytical judgment. The judgment simply goes unexercised. Capacities not exercised are not maintained.
Consider the driving directions problem. A generation of people who have navigated exclusively with GPS cannot read a road map, estimate distance, or reconstruct a route when the signal drops. Nobody decided to stop learning navigation. The skill went unexercised because the tool was always faster. The capacity did not disappear dramatically — it was simply never required. When the tool fails — in a dead zone, in an emergency, in an unfamiliar country — the gap becomes visible.
The same dynamic operates at the institutional level. Organizations that integrate AI into decision-making — in hiring, regulatory compliance, financial risk assessment, and medical triage — gradually outsource the judgment functions that AI handles more efficiently. Each individual outsourcing decision is defensible. The aggregate effect, across multiple domains over time, is an institution that has lost the internal capacity to reconstruct reasoning independently when the system fails or produces a result that requires human override.
Here is where the recursion closes. The institutions expected to govern AGI — regulatory bodies, legislative committees, courts, and international coordination organizations — are operating in the same AI-mediated environment. The human capacity required to govern advanced AI is being eroded during the build-up to it by the systems that precede it. The tools being built may be reshaping the institutions expected to govern them.
This is not a prediction. It is a description of a structural condition that is already observable. The generation now forming in AI-mediated educational environments will be the decision-makers, legislators, and voters eventually confronted with AGI governance. What they are being formed to do — and what cognitive capacities the formation environment is reducing — is not being studied systematically, is not subject to disclosure requirements, and is not visible in any accountability architecture currently operating.
The recursive problem is the one that the AGI safety conversation cannot see from inside itself. It requires the outside position to name it.
The Accountability Gap Becoming Structural
Every day without adequate external oversight is a day that inadequate governance hardens into the baseline. Institutions adapt to the governance environment they operate in. Labs have built compliance cultures, safety teams, and external relations functions calibrated to current standards. Regulators have organized their processes around the frameworks available. The longer those inadequate standards persist, the higher the cost of replacing them — and the more politically difficult it becomes to demand something stronger.
The governance decisions being made right now — what disclosure is required, what verification is sufficient, what accountability is owed for diffuse harm, what concentration is acceptable — are not temporary arrangements pending better ones. They are the foundation being poured. In construction, the inspection that matters happens during the pour. Once the foundation has been set, it determines everything built above it and is extraordinarily difficult to change.
The window for establishing stronger governance norms is not permanently open. Certain decisions — about concentration of AI capability, about international coordination architecture, about liability frameworks for formation damage and epistemic harm — have windows of tractability that close as conditions change. What is difficult but possible now may be practically impossible in three years, when the systems are more deeply embedded, the investments more committed, and the competitive dynamics more locked in.
What Is Now Visible
AGI may arrive soon, decades from now, or not at all. That technical uncertainty is genuine and unresolved.
What is not uncertain is this. The governance structure forming around the pursuit of AGI is already here. It is being built right now, by a small number of private organizations, using definitions they wrote, against standards they set, verified by no external party, at a pace that existing oversight architecture cannot track. The public whose conditions will be determined by this governance structure is being managed as a stakeholder to be informed rather than a sovereign to be consulted.
The people most affected by these decisions currently have the least standing in the conversation determining them.
That is the observation this page places on the record. Not a prediction. Not a scenario. A present-tense forensic finding — visible to anyone who applies the right questions to the available evidence.
The right questions are the ones auditors ask. Who determined materiality? Who verified the standard? Who has standing to challenge the determination?
The answers, right now, are: the labs. The labs. And almost no one else.
This page is the introduction to a five-part examination of AGI governance — four pieces diagnosing the problem across vocabulary, sequence, threshold, and self-certification, and a fifth proposing an actual remedy: a complete model statute for independent AI oversight. The argument continues on Thinking Sovereignty.