AI labs have secretly already built artificial general intelligence or conscious machines and are hiding it

The case people make

The suspicion is not baseless, and it deserves its strongest form. It rests on three things that are all genuinely true.

First, the pace. In a handful of years, AI went from producing garbled text to writing working code, passing professional exams, and holding conversations that many people find indistinguishable from a human's. When a capability curve bends that sharply, it is natural to assume the real frontier, the internal version not yet released, is further along than anything the public has seen. Former OpenAI researcher Leopold Aschenbrenner gave that intuition a detailed form in his 2024 essay Situational Awareness, arguing that the trend lines point toward AGI around 2027 and that the labs are not remotely prepared to secure what they are building.

Second, the secrecy. Frontier labs disclose very little about how their systems are trained or what they can do internally, and no independent body audits those systems. That is not a paranoid characterization; it is the ordinary condition of a competitive, high-stakes industry. But it means that anyone asking “what do they have that we have not seen” is met with a wall, and walls invite the imagination.

Third, the motive. If a lab did cross the line, there are real reasons it might not say so at once: fear of public panic, of triggering regulation, of destabilizing markets, or of handing rivals and hostile states a roadmap. Add a dismissed insider who sincerely believed a model was conscious, executives winking about AGI “achieved internally,” and a definition of AGI so loose that even the companies dispute it, and the pieces line up into a plausible-feeling shape.

Fast progress, closed doors, and clear motives to stay quiet. The suspicion writes itself. The question is whether writing itself is the same as being true.

The honest version of the case is not that a hidden AGI has been proven. It is that the conditions under which one could be concealed genuinely exist, and that a public with no way to verify the labs' claims is not being unreasonable to wonder.

Where the claim breaks down

Wondering is fair. The leap from they could be hiding it to they have achieved it and are hiding it is where the evidence runs out and the story takes over.

The first problem is that the expert consensus points the other way. The scientists who study machine consciousness most carefully, including the 19 authors of the 2023 Butlin and Long report, concluded that no current system is a strong candidate for consciousness. On AGI, the picture is one of forecasting and disagreement, not of a quiet, completed achievement. A minority of serious researchers argue some conditions might already be met under particular theories, and that disagreement is real, but it is a disagreement about criteria and evidence, not a suppressed finding.

The second problem is what concealment would require. A genuinely general or conscious system would not sit in a vault. It would be trained and run across data centers, touched by thousands of engineers, safety staff, contractors, cloud providers, and board members, many of whom have shown themselves willing to resign, leak, and publish over far smaller concerns. It would reshape the products and the revenue of whichever company held it. Keeping something that large silent, indefinitely, across competing firms, runs against how every comparable secret has actually behaved.

The third problem is that motive cuts both ways. Yes, there are reasons to hide a breakthrough. There are at least as many reasons to announce one: capital, talent, prestige, and market dominance all reward being first and visible. OpenAI's own reported contract structure gave it a concrete incentive to declareAGI, because doing so could end a partner's access to its technology. When the incentives push toward both silence and spectacle, a list of reasons to stay quiet proves nothing.

Underneath all of it sits the definition problem. If “AGI” has no agreed meaning, then “they secretly achieved AGI” is close to untestable: every attempt to check it can be deflected by redefining the target, and every ordinary product improvement can be recast as a glimpse of the thing being hidden. A claim that can absorb any evidence is not thereby strong. It is unfalsifiable, which is a different and weaker thing.

Reading a joke as a confession

It is worth pausing on the single most-cited piece of “evidence,” because it shows how the belief manufactures its own proof.

The line “AGI has been achieved internally” did not originate as a disclosure. It began as a meme on an anonymous account and was picked up by Sam Altman in a Reddit comment that he edited, the same day, to say he was “obviously just memeing,” noting that real AGI would never be announced in a Reddit comment. The primary record is a joke and its immediate disclaimer. To treat it as a confession, one has to discard the disclaimer as itself part of the cover-up, at which point the person is unfalsifiable by construction: anything the executive says, including a denial, becomes confirmation.

The Lemoine episode works the same way when it is stretched. A single engineer's sincere belief, contested and rejected by the specialists around him, is real and human and worth taking seriously as a story about how convincing these systems have become. It is not an institution admitting it holds a conscious machine. Reading it as one requires assuming the rejection was the lie and the lone believer the truth-teller, which is an assumption, not a finding.

When a denial counts as proof and a joke counts as a leak, the theory has stopped listening to evidence. It is filling silence with the answer it already preferred.

Why the belief spreads

The idea that a lab is sitting on a secret AGI is now common enough that MIT Technology Review called it, in October 2025, the most consequential conspiracy theory of our time, and that psychologists built a validated scale in 2025 to measure AI conspiracy beliefs. It spreads for reasons that say as much about the moment as about any particular company.

It rides real and rapid change. When a technology improves faster than people can update their intuitions, the gap between what is shown and what is imagined widens, and it is easy to fill that gap with a hidden frontier rather than an unfinished one. A future that feels this large seems to call for a secret to match it.

It feeds on genuine opacity and distrust. The labs really do disclose little, no one really does audit them, and many people already assume that powerful companies manage the truth. “They would never tell us” is a prior that turns every non-disclosure, every guarded statement, every internal-versus-public gap into a piece of the plot. The 2025 AI conspiracy scale found exactly these threads, global control and disinformation, running through how people think about the technology.

And it fits a familiar template. A secretive, wealthy institution guarding a world-altering secret is one of the oldest shapes a conspiracy story takes. The AI industry slots into it cleanly, which is part of why the belief feels intuitive even to people who could not say what evidence would support it. The template does the persuading before the facts are consulted.