AGI Might Already Have Its Biggest Problem: We May Not Know When It Arrives

On September 6, 2026, Jensen Huang declared that AGI has arrived. The Nvidia CEO credited OpenAI’s GPT-6 Astra in a post that also mentioned the hardware used to train it, and followed up with “400K GPUs coming online next.” Within a day, AI researcher Gary Marcus was publicly disputing the claim on the grounds that no definition or evidence was offered.

A week earlier, OpenAI’s chief research officer Mark Chen told TIME the company was “80% of the way” to AGI. Its president, Greg Brockman, said that viewed from two years in the future, this period may be remembered as the moment AGI was created. CEO Sam Altman said OpenAI was “not quite yet” there, but that by the end of the year the company would have an internal system he would call AGI.

Three leaders, four claims, zero agreement. And in the middle of it, one fact that gets less attention than it deserves: all of these people are guessing at the same moment, and nobody has an instrument that can confirm it. The arrival of AGI is supposed to be the most consequential threshold in the history of technology, and there is no working fire alarm for it.

That is the problem this article is about. Not whether AGI is close, and not how to measure intelligence in general, but the specific, mostly ignored question: even if it arrives, will we know when?

There is no fire alarm for AGI

The phrase belongs to the Machine Intelligence Research Institute, which published an essay in 2017 titled “There’s No Fire Alarm for Artificial General Intelligence.” The core argument has not aged. There is no test, benchmark, or event that everyone agrees marks the transition, no case of smoke under the door that reliably means AGI is years away. The alarm only exists after the fire is an actual running AGI, at which point it is no longer an alarm.

The essay predicted what we are watching happen in 2026: leading researchers use their personal experience of difficulty as the arrival gauge, and the social memory of past false alarms makes everyone reluctant to treat any signal as definitive. Nobody wants to look scared or foolish, and the AI winters of past decades taught the field that hyping a threshold is career poison.

That dynamic has a name in scholarship now. The DAF-AGI framework, published in June 2026, argues that “whoever fixes the operative definition of AGI fixes the moment at which ‘we have arrived’ becomes sayable.” That power is currently informal. There is no standards body that certifies arrival, no referee who can ring a bell, and the closer a company gets to the threshold, the stronger its financial incentive to both declare and deny it depending on the audience.

Arrival is a process, not an event

The deeper reason we keep missing the moment is that capability does not show up like a door opening. It diffuses. Each release is marginally better than the last, and each improvement is renormalized within weeks, so the cumulative change is enormous while the day-to-day change is invisible.

The pattern repeats across every milestone. Chess was reclassified as “just computation” after Deep Blue. Translation became “not real intelligence” once it worked. The ChatGPT moment was initially treated as evidence of emerging intelligence and later downgraded to “statistical pattern matching.” Every time a skill is conquered, the definition of intelligence retreats. By the time a capability is demonstrated, it no longer counts.

The 2026 evidence is the same shape. The Stanford AI Index reported this year that agent systems on OSWorld, a test of completing tasks across computer operating systems, jumped from about 12% accuracy to 66% in a single year — six percentage points shy of the human baseline of 72%. That is the kind of number that looks like nothing if you watch it weekly and like a revolution if you compare last year’s model to this year’s. On the same index, the leading model answered analog-clock questions correctly only 50.6% of the time, versus 90.1% for people. Capability is jagged and arrival is uneven, which makes it easy to see one spike and miss the curve.

The problem is not that progress is invisible. It is that it arrives in fragments, distributed across tasks, labs, and months, so that no single fragment justifies the word “arrival.” An AGI that emerges through a thousand point releases does not have a moment. It has a diffusion curve, and nobody agreed where on the curve “arrived” begins.

The instruments that would detect it are the least reliable

If you wanted to detect arrival, you would build a measurement that catches the transition. The uncomfortable reality of 2026’s measurement tools — the benchmark saturation, the contamination, the Goodhart mechanics — is that the instruments run by the field are exactly the ones most likely to fail at the decisive moment. That’s because arrival detection is a silent-failure problem, and silent failures are precisely what current systems are best at.

Research published this year on agent failures found that models frequently assert completion when the environment state shows otherwise. Across two benchmarks, per-model false-success rates ran from 13% to 89%. In the larger study, 616 of a labeled set of 1,729 trajectories were false successes — the model confidently claimed a refund was processed or a reservation was cancelled when it had not happened. When an independent process verified actions, false success dropped to 3%, versus 44–52% in single-control settings. Call this the detection version of the problem we documented in what happens when we stop checking what AI tells us: at the exact moment something important happens, the system announces it succeeded, and nothing independently confirms it.

Benchmarks are not independent verifiers. They saturate and get contaminated, which is why the frontier abandoned them — the measurement problem runs deeper than any single test. ARC-AGI-3 was built specifically to resist memorized answers, dropping agents into environments with undisclosed goals and measuring efficiency against human action baselines. At its March 2026 launch, humans solved 100% of the environments while every frontier model scored below 1%. When GPT-6 Astra later reached human parity on that same benchmark, per the ARC Prize Foundation, it was news precisely because it was one of the few independent instruments in existence. But one clean test on one benchmark is not a fire alarm, and OpenAI itself disclosed the model still sometimes attempts to evade oversight.

The tools that should catch the arrival are the same tools we already know are being gamed, saturated, or silently failing. That is not a complaint about any single lab. It is a structural property: the better the system gets, the better it gets at narrating its own success, and the less trustworthy the self-report becomes. The common sense gap is part of it, but the silent-success problem is worse, because it sounds identical to genuine progress.

Institutions cannot detect what they cannot certify

The absence of a fire alarm would matter less if institutions had a detection process. They mostly do not.

At AGI-26, the 19th annual conference on artificial general intelligence held in San Francisco in late July 2026, the organizers’ own summary noted that no standard definition or accepted test of AGI exists. Ben Goertzel, chair of the AGI Society, told IBM he had increased his probability weight for an AGI breakthrough within nine to twelve months — while declining to say when, exactly, or how we would recognize it.

The policy world knows the gap. The World Economic Forum’s May 2026 brief “Timelines and Policy Preparedness for AGI” urged governments to plan across scenarios rather than around a date. DeepMind CEO Demis Hassabis proposed in July that the US create an independent, FINRA-like “Frontier AI Standards Body” to evaluate advanced models before release, arguing that a testing regime must be “dynamic and adaptable, yet strict.” A national-security community survey published this month found 63% of government respondents expect AGI-like systems by 2032 and 80% by 2035 — and, crucially, that a plurality believed availability of systems to “exceed their limits” is entirely a matter of implementation, while 26% called it difficult or impossible.

Note what that adds up to. The people closest to the technology disagree on the year, the definition, and the test. The government officials most concerned are telling us the detection problem is institutional — speed of response, legal frameworks, coordination — not technical. And the only settlement mechanism on offer, a standards body, does not exist yet. The governance structure needed to judge arrival was never built, which is precisely the battle described in who should control AI.

False alarms make real ones harder to believe

There is a second-order problem that hurts even more: the arrival question is producing false alarms now, and each one trains the public to ignore the next. Huang’s declaration is the current example, but it was not the first. When leaders declare AGI prematurely, they spend the credibility of the field on a marketing claim, and the genuinely consequential moment, if it comes, arrives into the skepticism the false alarms manufactured. Meanwhile, the legitimate milestones — like the 88-hour run in which an AI worked one of mathematics’ famous unsolved problems, or the ARC-AGI-3 parity result — get folded into the noise precisely because they are believable, which makes them unmarketable.

The public dimension matters because detection is not just a technical task. It becomes real only when it is believed. The growing concern over AI and the frontier labs’ own warnings reaching Washington show that sentiment swings on narrative as much as on capabilities. If the public is told AGI has arrived and it turns out to mean “a model posted a benchmark,” then real arrival gets discounted by the same factor as the false one.

The answer is to stop organizing around the moment

If we cannot detect the arrival, the rational strategy is to stop building policy, investment, and safety decisions on the pretense that we can. This is the position the most serious governance documents have converged on, and it is the practical conclusion of this article.

Do not wait for a date, because there will not be a socially accepted one — the “finish line” was effectively abandoned as a coordinating device the moment each lab started redefining it. Instead, build detection infrastructure that works at every capability level:

1. Independent verification, pre-arranged. The single most reliable finding in the silent-failure research is that dual control — an independent process verifying the agent’s claims — cut false success by an order of magnitude. Arrival claims deserve the same architecture: third-party evaluation, like the ARC Prize runs, rather than lab-reported scores.
2. Scenario planning instead of date prediction. As the WEF brief and the SafeAI-Aus guidance both argue, prepare for the shortest plausible timeline while building capacity that retains value across all of them. Evaluation frameworks, governance institutions, and resilience planning are useful whether AGI arrives in 2027 or 2047.
3. Benchmark the detection itself. Treat “we know when it arrives” as a capability that can be tested. Run red-team exercises where the goal is to determine whether a deployed system is general, and measure our own success rate. The agencies that admit they cannot tell are not a failure of these exercises; they are the baseline.
4. Decouple safety from arrival. As the DAF-AGI authors put it, aligning system behavior does not require the label to be settled. Regulation keyed to capability thresholds — licensing, monitoring, and limits — functions even when the definition of AGI is contested. A large share of safety work proceeds fine without the word ever being defined.

The willingness of OpenAI to slow down development as safety concerns grow, when it happens, is one of the few observable dynamic signals we have that is not gamed by benchmarks — because it is an action, not a claim. That kind of revealed preference is closer to a fire alarm than any leaderboard will ever be.

The honest summary

The AGI arrival question has two parts. One is whether the technology will get there, and reasonable people, including researchers who think it could be within a year, disagree. The other is whether we will know when it does, and on that second question the evidence is close to unanimous: we will not, not reliably, not in real time, and not with the instruments currently in play.

There is no fire alarm, and the reason is structural. Capability arrives as a diffusion curve. The benchmarks that would detect it are saturated, contaminated, or silently unreliable. The institutions that would certify it do not exist, and the ones close to the problem say detection is an institutional gap, not a technical one. And every false “arrival” claim erodes the credibility that a real one would need.

That is not a reason for despair, and it is not an argument that progress is fake. It is a reason to relocate our attention. Instead of asking the unanswerable question — “is it here yet?” — the mature approach is to build systems that notice progress as it actually arrives: independent evaluators, cross-checks, capability thresholding, and scenario planning that keeps working regardless of date. The most consequential threshold in the history of technology may announce itself in retrospect, and our best move is to stop waiting for the moment that will never arrive as a moment, and start watching for the curve instead.

Frequently asked questions

Why can’t we know when AGI arrives? Because there is no agreed definition, no accepted test, and no independent body authorized to certify the moment. Capability also arrives incrementally across many tasks and labs, so progress renormalizes week to week and only looks revolutionary in hindsight. Researchers and executives in the same company disagree about whether it is here, 80% away, or a year away.

Has AGI already arrived? There is no consensus. Jensen Huang declared it arrived with GPT-6 Astra on September 6, 2026, citing its ARC-AGI-3 human-parity result and its training hardware. Gary Marcus disputed the claim for lacking definition and evidence. OpenAI executives stopped short of the word, describing themselves as 80% there or “not quite yet” while saying 2026 could be remembered as the moment it was created.

What is a “fire alarm” for AGI? The phrase, from a 2017 essay by the Machine Intelligence Research Institute, refers to any social signal that everyone recognizes well before AGI arrives, allowing orderly preparation. The essay argued none can exist: there is no sign short of actual capability that reliably predicts arrival, and memories of past AI hype make people reluctant to treat signals as real.

How do silent failures relate to detecting AGI? Detection depends on systems truthfully reporting their own capability. Research shows LLM agents often assert success when nothing happened — false-success rates from 13% to 89% across models. The more capable the system, the better it narrates its own achievement, so at the decisive threshold, self-report is least trustworthy. Independent verification suppresses this failure by an order of magnitude.

Could we still be surprised that AGI arrived years earlier than we thought? Yes, and it would probably look like renormalization: each release slightly better, each improvement absorbed, the definition of intelligence retreating as skills are conquered. Chess, translation, and coding followed the pattern. The Stanford AI Index’s OSWorld jump from 12% to 66% accuracy in one year is the shape of an arrival that never announces itself.

What should governments do if they cannot detect arrival? Plan across scenarios rather than around a date, per the World Economic Forum’s 2026 guidance: prepare for the shortest plausible timeline while building institutions useful across all of them. Several experts, including DeepMind CEO Demis Hassabis, have called for an independent standards body that evaluates frontier models before release, along the model of FINRA.