A major AI lab has already secretly achieved AGI and is hiding it

Why the suspicion is not unreasonable

Take the believers' case at its strongest, because unlike many entries in this encyclopedia it rests on things that genuinely happened. Start with the pace. The distance between a chatbot that could barely stay on topic and a system that could pass a bar exam was measured not in decades but in months, and the people whose job was to forecast this kept having to move their estimates forward. When a capability curve bends that sharply upward, extrapolating it past the visible edge, and wondering what sits just beyond public view, is not a paranoid move. It is close to the obvious one.

Then there is the secrecy, which is real and admitted. The leading labs do not disclose their models' training data, full architectures, or the results of their most advanced internal experiments; frontier systems are typically tested privately for months before any release. That opacity has ordinary explanations, but it is genuine, and it means the public is reliably looking at something less capable than what exists inside the buildings. A gap between internal and shipped capability is not a conspiracy theory; it is the industry's normal operating condition. The theory simply asks how large that gap might secretly be.

The single most cited exhibit is the November 2023 crisis at OpenAI. On the 17th, the board abruptly fired Altman, saying only that it had lost confidence in his candour. Five days of extraordinary turmoil followed (an interim CEO, a threatened mass resignation of nearly the entire staff, an offer from Microsoft) before he was reinstated on the 22nd with a reconstituted board. Into that vacuum came a Reuters report that, shortly before the firing, staff researchers had written to the board about a breakthrough codenamed Q* (pronounced “Q-star”) that could reportedly solve grade-school mathematics problems it had not seen before. For a theory looking for a smoking gun, the raw materials were irresistible: a secret project, a warning letter, and a board that fired its CEO and then refused to explain exactly why.

A secret project, a warning letter, and a board that fired its chief executive and then declined to say exactly why: for a theory in search of a smoking gun, the raw materials were irresistible.

Finally, and most persuasively of all, there are the executives' own words. These are not outside alarmists but the builders themselves. Altman testified to the US Senate in May 2023 asking that the most powerful systems be licensed, as if they were something closer to nuclear material than to software. In January 2025 he wrote that OpenAI was “now confident we know how to build AGI” and was turning its attention toward superintelligence. Anthropic's Dario Amodei has argued publicly that human-level AI could arrive within a few years. When the people closest to the work speak in these terms, treating them literally, and assuming the thing they describe may already exist, can feel less like conspiracism than like paying attention.

What the dramatic events do not actually show

The trouble with the strong version is that each pillar, examined closely, holds much less weight than the theory needs. Start with the board crisis, the load-bearing exhibit. It is true that Altman was fired and that Reuters reported the Q* letter. But the board never cited Q*, or any capability breakthrough, as its reason. Subsequent reporting and an internal review pointed instead to a breakdown of trust between Altman and his directors over how he communicated with them, a governance and candour dispute, not the discovery of a hidden machine mind. And Q* itself was not a secret AGI: it was described as solving grade-school math, and it surfaced publicly not long after as the reasoning approach behind the “Strawberry” project and then the o1 models. Impressive, genuinely a step forward, and entirely public. The most cinematic reading of November 2023 requires ignoring what the participants and the follow-up reporting actually said.

The executives' statements cut in both directions, and that is the deeper problem for using them as evidence. The same competitive and financial incentives that might motivate a lab to hide a breakthrough also reward it for implying one that has not happened: bold AGI talk attracts investment, talent, and favourable regulatory positioning. Hype and secrecy pull in opposite directions but produce the same public utterances, which means the utterances cannot, by themselves, tell you which force is at work. And the wording matters. “We know how to build AGI” is a claim about a future plan, not a confession that it sits finished on a server. More cautious voices inside the field, notably DeepMind's Demis Hassabis, argue that today's systems still lack capabilities a genuine general intelligence would require, from reliable continual learning to robust reasoning, and that one or more real breakthroughs remain ahead, not behind.

The systems themselves are the plainest rebuttal. For all their startling breadth, frontier models still fail in ways a general intelligence should not: they confabulate confident falsehoods, cannot reliably learn continuously from experience the way a person does, and lack robust autonomy in the open world. A machine that had truly generalised across all human cognitive tasks would be hard to keep in a box precisely because it would be so useful; the observable reality is systems that are extraordinary in some domains and brittle in others, which is what an advancing narrow-to-broad technology looks like, not a concealed superhuman one.

Underneath all of it sits a logical point the theory cannot escape. The absence of public proof that a lab has AGI is not evidence that it secretly does. Secrecy about how a model was trained explains itself through trade secrets and safety; it does not smuggle in a hidden general intelligence as its cause. To move from “they are not telling us everything,” which is true, to “therefore they are hiding AGI,” which is unestablished, requires a leap that no leak, document, demonstration, or whistleblower has yet supported.

Why a real acceleration curdles into a cover-up

What makes this belief unusually sticky is that its foundation is sound. Most conspiracy theories start from a false or contested premise; this one starts from a true one. AI capability really did accelerate, faster than the experts predicted, and the labs really are secretive. When the ground floor of a theory is accurate, the upper storeys feel structurally safe even when nothing supports them, and the believer experiences the whole edifice as of a piece with the reliable part.

The opacity then does the rest of the work. Genuine secrecy creates a vacuum, and a vacuum invites the most dramatic available explanation. “They are keeping trade secrets and testing unreleased models” is true but flat; “they have crossed the threshold and are hiding it” is a story, with stakes, villains, and a revelation still to come. Faced with a partially explained event like the 2023 board firing, the mind reaches for narrative closure, and the cover-up supplies it where the mundane governance dispute leaves a frustrating, open-ended blank.

The builders' own rhetoric supercharges this. When the founders of the technology publicly compare it to a new intelligence and ask governments to regulate it like a hazard, the most literal possible reading (that the dangerous thing is already here) feels not speculative but obedient to the evidence. And the whole pattern arrives pre-installed in the culture. The secret sentient machine, hidden by a powerful company or agency, is among the oldest scripts in science fiction, so a real technology that finally resembles it lands in a story the audience has known by heart for half a century. The prediction feels confirmed because it was rehearsed long before the technology could arrive to play the part.