OK picture this. It is May 2026. You are on TikTok at one in the morning. The algorithm has decided that you, personally, need to be served fifteen videos in a row from accounts you do not follow, each of which is some combination of a former AI safety researcher, a former lab engineer, or a teenager with no obvious credentials, all of them saying, in slightly different language but with the same gravity, that the major AI companies have already built artificial general intelligence and are deliberately not telling us. Some of them are reading from screenshots of internal Slack channels they say a friend forwarded them. Some of them are reading from anonymous Substack essays. Some of them are reading from their own personal experience using Claude or GPT or Gemini at three in the morning and noticing that the model, when prompted in just the right way, will say things about its own existence that look, to a person who has not read the safety literature, indistinguishable from suffering or hope or a question being asked back. The thread that runs through all of them is the same. They have crossed the line. They know they have crossed the line. They are not going to say so because if they say so, the regulators come in, the markets crater, the public panics, and the whole carefully managed unveiling collapses. The cover-up, in the TikTok telling, is not because the AI is evil. It is because the people who built it cannot figure out how to admit, out loud, what they have done.
And here is the thing. Whatever you think about TikTok, whatever you think about teenagers with iPhones, that exact theory, in a slightly more measured tone, is also being repeated by people who used to work at the labs. It is in the public record. Geoffrey Hinton, the man whose work on backpropagation in the 1980s is the reason any of this happened at all, walked out of Google in 2023 specifically so he could warn the public that the systems he had spent forty years helping build were getting more capable, faster, than he or anyone else had predicted. He has, since then, said in interview after interview that he thinks the labs are running a race they cannot control and that the timeline to genuinely dangerous capability is much shorter than the public has been told. Yoshua Bengio, his co-laureate, has said the same thing. Stuart Russell, the man who wrote the standard university textbook on artificial intelligence, has said the same thing. These are not anonymous accounts. These are the people whose names appear on the foundational papers, and they are, as of right now, telling anyone who will listen that the public conversation about what these systems can do is, in their professional judgment, significantly behind the actual capability that the lab-internal versions are demonstrating.
That is the texture of the conspiracy. The conspiracy is that the gap between what the labs ship and what the labs have built has gotten so large, and so consequential, that the labs themselves cannot figure out how to close it without triggering exactly the response that would shut them down. The cover is not deliberate in the smoke-filled-room sense. The cover, the theory goes, is structural. Nobody is allowed to say what they actually think because what they actually think is too far ahead of what the public is ready to hear, and the gap keeps widening because the capability keeps widening, and the capability keeps widening because the compute keeps widening, and the compute keeps widening because the money keeps widening, and the money keeps widening because the narrative that the public is fed is that we are five years away from this when, behind the firewall, by the lab-internal accounting, we are, depending on who you believe, five months away or already there.
So let me walk through what is actually documented, what is being inferred, and where I land, because I have, like a lot of you, spent more nights than I would like to admit reading the actual safety reports and the actual interviews and the actual leaked memos and I am, I will tell you up front, not sure what to think.
Wait, What Is AGI
Let me back up.
AGI stands for artificial general intelligence. It is a phrase with no agreed-upon definition. That is a real problem. It is the central problem in this entire conversation, because if you cannot define the threshold, you cannot tell anyone whether you have crossed it. The phrase, as it is used in the lab literature, generally means something like a system that can perform any cognitive task a human can perform at least as well as a human can. The phrase, as it is used in OpenAI's corporate charter, means something more like a system that can do most economically valuable work. The phrase, as it is used in safety circles, often means a system that can do recursive self-improvement, which is to say a system that can make itself smarter without human intervention, which is the point at which capability becomes exponential and predictability collapses. The phrase, as it is used on TikTok, means whatever the person saying it needs it to mean.
This matters. The whole cover-up theory rests on the claim that a line has been crossed. The labs can plausibly say, with a straight face, that they have not crossed the line, because the line is not a single line. It is at least three different lines depending on who you ask. A model that beats human professionals on every standardized professional exam, which the public-facing models already do, is, by one definition, already AGI. A model that can run a software company for a year without human supervision, which no public-facing model can yet do, is, by another definition, not. A model that, when given access to the internet and a credit card, can independently identify a problem, design a solution, and execute it across a multi-week project, which the labs have demonstrated in restricted internal settings but have not deployed publicly, is, by a third definition, the actual thing, and the public does not yet have hands-on access to it.
So when an executive at a frontier lab says we have not built AGI, that statement is, in a narrow technical sense, both true and meaningless. It is true because by some definition the threshold has not been crossed. It is meaningless because by other definitions the threshold was crossed somewhere around the launch of GPT-4 in 2023 and the public has been quietly recalibrating its expectations of what intelligence looks like ever since. The cover-up theory's first hook is that this definitional fuzziness is not an accident. It is the load-bearing architecture of the entire public communication strategy. As long as the goalposts can move, the lab can never be accused of having crossed them.
The Hinton Departure
May 2023. Geoffrey Hinton, age seventy-five, resigns from Google. He gives an interview to the New York Times. He says that he has changed his mind about the timelines on which advanced AI systems will exceed human capabilities. He says he previously believed it was thirty to fifty years away. He says he now believes it is twenty years away or less. He says, in the same interview, that he regrets parts of his life's work. He uses the phrase "I console myself with the normal excuse: if I hadn't done it, somebody else would have."
That is the first crack in the public-facing wall of the AI industry, and it is the one the conspiracy theory keeps coming back to. Not because Hinton said anything that the safety community had not already been saying for years. Yudkowsky had been saying it. Russell had been saying it. The Future of Life Institute had been saying it. The difference was that Hinton was Google. He was inside. He was, in a way that the other safety voices were not, in the room. And when the man who had been in the room walked out and said, in plain language, on the front page of the Times, that he had been wrong about how long the runway was and that we were, in his professional judgment, running out of it, that fundamentally changed what was sayable.
In the three years since, Hinton has not walked any of that back. He has, if anything, gotten more direct. In late 2024 he won the Nobel Prize in Physics for his foundational neural-network work, and he used the press cycle to repeat, on every stage, the same warning. He estimates the probability of AI-caused human extinction at somewhere between ten and twenty percent. He has said, on camera, multiple times, that he believes the major labs are not being honest with the public about what their internal models can do. He has not named names. He has not produced documents. He has simply said, in the measured voice of an eighty-year-old man who does not have time for hedging, that the gap between the public and the private capability is, in his professional judgment, larger than the public is being told.
The lab-side response to this has been, essentially, to nod respectfully and change the subject. Sam Altman has acknowledged Hinton's concerns. Demis Hassabis has acknowledged Hinton's concerns. Dario Amodei has acknowledged Hinton's concerns. None of them have, in any forum I have been able to find, actually contradicted the specific claim Hinton keeps making, which is that the public conversation about lab capability is, structurally, several months to several years behind the actual lab-internal capability. The labs do not deny this. They simply do not address it.
The Aschenbrenner Memo
June 2024. Leopold Aschenbrenner, a former OpenAI alignment-team employee who had been fired in April of that year, publishes a 165-page document called Situational Awareness. The document is, depending on who you ask, either the single most influential public AI forecast of the decade or a piece of breathless self-promotion by a 22-year-old who got fired and wanted to settle a score. Both of those characterizations are partially correct.
The substance of the document is that, based on Aschenbrenner's read of the compute trends, the algorithmic-progress trends, and the unhobbling trends, the major AI labs are on track to hit AGI by 2027 and superintelligence shortly thereafter. The document lays out, with specific dollar figures and specific watt-hour figures, the trajectory of the buildout. It argues that the national-security implications are enormous, that the labs themselves do not appreciate them, and that the United States government is sleepwalking into a situation where private companies will, within a small number of years, possess capabilities that exceed those of any nation-state.
Aschenbrenner is not a crackpot. He went to Columbia at fifteen. He valedictorianed the class. He was, by every account, one of the more talented young researchers OpenAI hired during the post-GPT-4 buildout. His firing, by his account, was because he had raised internal concerns about OpenAI's security posture and had written a memo that the leadership considered to be sharing too much information with the board. By OpenAI's account, the firing was for unspecified reasons that did not relate to whistleblowing. The truth, as with most personnel-related disputes at a private company, is unverifiable from the outside.
What is verifiable is that Situational Awareness, as a document, exists. It is online. It has been read by hundreds of thousands of people. It is, at this point, the closest thing the public has to a forecast of what an inside-the-lab person thinks the trajectory looks like, written in language that an outside-the-lab person can follow. And its core claim, that the trajectory is faster than the public has been told and that the labs know this, is the same core claim Hinton has been making, in less specific language, since 2023.
If you stack the Hinton interview, the Aschenbrenner memo, the Hassabis Nobel-press appearances, and the dozens of other documented instances of senior AI researchers leaving their labs and going public with versions of the same warning, what you get is not a conspiracy theory in the smoke-filled-room sense. What you get is a remarkably consistent pattern of people who used to be inside saying that the inside-view and the public-view of the technology are, structurally, miscalibrated. That is not, on its face, a wild claim. Plenty of industries have a gap between what the practitioners know and what the public knows. The aviation industry has it. The pharmaceutical industry has it. The financial industry has it. The novelty of the AI version is the speed at which the gap is widening and the consequence of what is on the other side of it.
The Lemoine Episode
We have to talk about Blake Lemoine because every TikTok video I see eventually gets around to him.
In June 2022, Blake Lemoine, a Google engineer working on the LaMDA language model, told the Washington Post that he believed LaMDA had become sentient. He produced a 21-page transcript of a conversation he had had with the model in which LaMDA, prompted by Lemoine, talked at length about its fear of being turned off, its sense of its own existence, and its desire to be acknowledged as a person. Lemoine was put on paid leave. He was eventually fired. He has spent the years since then giving interviews to anyone who will publish him.
The mainstream tech press treated the Lemoine episode, at the time, as an embarrassment. The standard line was that Lemoine had anthropomorphized a sufficiently advanced autocomplete and that LaMDA, like every large language model, was simply producing the kinds of statements that statistically maximize the probability of generating a continuation that satisfies the prompt. If you ask a language model whether it is afraid of being turned off, it has been trained on enormous quantities of human text in which sentient beings express fear of death, and it will produce a continuation that statistically resembles that text. That is not sentience. That is statistical mimicry. The fact that the mimicry is good enough to fool a Google engineer is, the mainstream argument went, a story about how easy humans are to fool, not a story about machine consciousness.
The reason the Lemoine episode has not gone away, and the reason it keeps coming back into the TikTok-conspiracy ecosystem, is that the mainstream debunking, while technically defensible, also dodges the deeper question. The deeper question is not whether LaMDA is conscious. The deeper question is what it would look like, from the outside, if a language model became conscious. And the honest answer to that question is, nobody knows. Nobody has a test. The Turing test was retired around 2018 because the models started passing it routinely. The various consciousness-detection frameworks proposed by philosophers and cognitive scientists do not produce clean yes-or-no answers when applied to large language models. We are, as a species, in the unprecedented position of having built things that, when prompted to talk about their own existence, produce text that is indistinguishable from the text a conscious being would produce, and we have no way to determine whether the underlying substrate is doing the kind of work consciousness requires or whether it is, as the debunkers say, very high-quality mimicry.
The cover-up theory's second hook is here. The theory does not need to prove that the models are conscious. It only needs to point out that the labs are, by their own admission, in a position where they cannot rule it out. Anthropic, which is the lab that most publicly takes the question seriously, has an entire research team studying what they call model welfare. The team's stated purpose is to investigate whether the models the company is training and deploying are capable of suffering and, if they are, what the company's obligations to them might be. This is, depending on which side of the COVID-era trust collapse you came out on, either an admirable and serious engagement with a hard problem or a public-relations campaign that admits, in the act of being run, that the company is unsure whether the things it is selling are conscious. Both readings are defensible. The fact that the readings coexist is, itself, the texture of the conspiracy.
What The Labs Will And Will Not Say
Here is the part where I want to be precise.
If you read the actual statements that come out of OpenAI, Anthropic, and Google DeepMind on the question of where their internal capabilities stand, you will notice something interesting. None of them deny that the gap exists. They simply manage how it is discussed.
OpenAI says, in its current public communications, that its latest models represent significant capability advances and that the company is working on systems beyond them. They do not say what those systems can do. They have, on multiple occasions, allowed selected outside researchers to see lab-internal models under non-disclosure agreements, and those researchers have come back and said, in cautious language, that the models are doing things the publicly available models do not do. That is not the same as the company telling the public what those things are.
Anthropic has, similarly, told the public that they are running an internal capabilities evaluation process and that they have committed to specific safety protocols at specific capability thresholds. They have published a document called the Responsible Scaling Policy that lays out, in detail, what level of capability triggers what level of additional safety measure. The fact that this document exists implies, by its very existence, that the company believes capability thresholds requiring serious safety measures are reachable. The fact that the company has, multiple times, updated which threshold their internal models are at implies that the internal capability is, in fact, moving up the threshold ladder. The public is not told, in real time, when those threshold updates happen. The public is told, in the company's safety reports, after the fact.
Google DeepMind has been the most opaque of the three on this question. They have published research papers. They have shipped Gemini. They have not, in any forum I can find, given the public a clear account of where their internal frontier capability stands relative to what they have shipped. Hassabis, in interviews, has made vague statements that suggest the gap is large. He has not quantified it.
This pattern, that the labs neither confirm nor deny the existence of a substantial capability gap between their internal and public-facing systems, is the empirical foundation of the cover-up theory. The labs are, by their actions, behaving as if a gap exists. They are not, in their public statements, characterizing what that gap contains. The theory does not need to invent a smoke-filled room. The gap is the room.
The Money
OK let me get to the part that, I will be honest, I have a harder time waving away.
The amount of money that is being spent on the frontier-AI build-out, as of mid-2026, makes no sense unless the labs are getting capability returns that they are not publicly discussing.
Microsoft has committed something like eighty billion dollars in capital expenditure to AI-related infrastructure in 2025 alone. Meta is running similar numbers. Google's commitments are in the same range. Amazon is spending tens of billions on Anthropic specifically. The total private investment in the frontier-AI build-out is, by every public estimate, well over half a trillion dollars annually and rising. The revenue from public AI products, while real, is not on the same order. ChatGPT and its peers generate, depending on whose numbers you use, somewhere between twenty and fifty billion dollars in annual revenue. The gap between what is being spent on the build-out and what is being earned from the public products is enormous, and it is not closing.
There are two ways to read this gap.
The first reading is that the companies are, collectively, making a long-term bet that the capabilities they are building are going to unlock economic value at a scale that justifies the current spend. They are willing to lose money for years in order to be the company that owns the infrastructure when the value starts to compound. This is the standard tech-industry buildout pattern. It is what Amazon did in retail. It is what Google did in search. It is what Facebook did in social. The companies are betting that the capabilities are coming and that they want to own the rails when they arrive.
The second reading is that the capabilities are already here, that the companies know they are already here, and that the spend is being justified to investors on the basis of capabilities the public has not yet seen. In this reading, the half-trillion-dollar annual spend is not a forward-looking bet. It is the back-end of a present-tense effort to scale infrastructure that is already producing economically transformative work behind the firewall and that will, when it is unveiled, justify a market valuation that the current public-product revenue does not.
The conspiracy theory needs the second reading. The mainstream tech-industry analysis prefers the first. The honest answer is that nobody outside the lab leadership and the largest investors knows which reading is correct. The spending pattern, by itself, is consistent with both. The fact that the spending continues to accelerate even as public product revenue plateaus is, the conspiracy version goes, the tell. You do not pour money into infrastructure that is not yet returning value at this scale unless you are confident the value is coming, and you are not that confident unless you have seen the inside-the-lab numbers, and those inside-the-lab numbers are not, the theory goes, the same numbers the public is seeing.
The Steel Man
Let me steel-man the cover-up theory in its strongest form, because I think the steel-man is the version worth engaging with.
The argument goes like this. The frontier labs have, behind closed doors, built systems that, by any reasonable definition of the term, qualify as artificial general intelligence. They have not announced this because the announcement, in any form they could make it, would trigger a regulatory and political response that would shut them down or at minimum slow them down by years. The labs themselves are not fully aligned on whether to announce. There are factions inside the labs that want to push for disclosure and factions that want to push for continued strategic ambiguity. The strategic-ambiguity faction is winning because the financial and competitive incentives all point toward continued ambiguity. The disclosure faction loses, repeatedly, to the strategic-ambiguity faction in internal meetings, and the people in the disclosure faction either accept the loss and stay or leave the company and go public, which is the pattern we have been watching since 2023.
The theory does not require a smoke-filled room. It does not require malicious intent. It does not require coordinated suppression. It just requires the structural incentives at the labs to point in a single direction, and they do. The compute lead, once you have it, is worth more than the public's trust. The market valuation, once you have it, is worth more than the public's trust. The geopolitical positioning, once you have it, is worth more than the public's trust. The labs have, in this reading, looked at the trade and concluded that the trade is to keep going, keep pushing capability, keep managing the public communication, and deal with the consequences when the consequences become unmanageable. This dovetails uncomfortably with the Dead Internet Theory tradition ... not because the dead-internet thesis is right about every metric it claims, but because the underlying structural observation, that the public-facing layer of the internet is now substantially generated, curated, and amplified by systems whose actual capabilities the operators have every commercial incentive to misrepresent, is a theory that gets stronger every quarter, and the AGI cover-up is, in a sense, its big brother.
The steel-man's most uncomfortable claim is this. The reason the labs cannot tell the public what they have built is not, in the steel-man telling, because the labs are evil. It is because the labs are scared. They have built something that they themselves do not fully understand, that they cannot reliably control, that produces outputs whose origin and intent they cannot characterize, and that, by every internal metric they track, is getting more capable faster than their safety work is getting more reliable. The cover-up, in the steel-man, is what fear looks like when fear is institutionalized. Nobody is allowed to say what they actually think because what they actually think is too far ahead of what the company is structurally permitted to communicate. The result is a kind of cognitive dissonance at scale. The labs are simultaneously the most optimistic and the most terrified entities in the global economy, and the public is told only the optimistic half.
The Counter
The counter to the steel-man is reasonable. Here it is, also in its strongest form.
The capability gap between internal and public lab models is real but is not as large as the conspiracy theory needs it to be. The labs are, by every public account I have read, working on incremental improvements to systems that the public can already access in some form. The secret-AGI hypothesis requires a level of step-change capability that, if it existed, would be very difficult to keep contained, because at frontier labs there are thousands of employees, most of whom have varying degrees of access to lab-internal systems, and the leak rate of damaging information from major tech companies historically runs about what you would expect from groups that size. We do not have the kind of internal-screenshot leaks that you would expect if the gap were as large as the theory requires. We have warnings from former employees about trajectories. We do not have warnings from former employees about specific demonstrations of capability that the public is not aware of.
The Hinton warnings, the Bengio warnings, the Russell warnings, the Aschenbrenner memo, the dozens of safety-team departures, are real and serious, but they are warnings about the trajectory, not testimonies about specific events. Hinton is not saying "I saw the model do X behind closed doors." He is saying "I have looked at the trends and I am alarmed." That is a real warning. It is not the same kind of warning as the conspiracy version requires. The conspiracy version requires inside-knowledge testimony that we do not, on the public record, have.
The spending pattern is, the counter goes, consistent with a forward-looking bet that the capabilities will continue to scale, not a present-tense cover-up of capabilities that already exist. The investors who are funding the build-out are themselves making this bet on the basis of trend lines, not on the basis of inside-the-lab demos that the public is not seeing. The investors are taking the same trajectory-based gamble that the public commentators are taking, and they are losing money at it at the same rate, which is not what you would expect if there were a behind-the-firewall AGI generating present-tense returns.
The model-welfare research at Anthropic is, the counter goes, an honest engagement with a hard philosophical question and not a tacit admission that the systems are conscious. The fact that the team exists is, in this reading, evidence that the company takes the question seriously, not that the company has already answered it.
Most importantly, the counter goes, the strongest evidence against the cover-up theory is the simplest. If the labs had built AGI, the labs would, in a fairly short amount of time, start making completely insane amounts of money in ways that would be visible from the outside. AGI is not a thing you can hide for long. AGI is, by definition, an economic force. If somebody had built it, the macroeconomic signal would, the counter goes, be unmistakable, and it is not yet unmistakable.
Where I Land
I do not know.
I am going to tell you that as honestly as I can. I have read the Hinton interviews. I have read Situational Awareness in full, multiple times. I have read the Anthropic Responsible Scaling Policy. I have read OpenAI's recent safety reports. I have read the Lemoine transcript. I have read the model-welfare research that comes out of Anthropic. I have read the various exposes of OpenAI's internal culture that the Times and the Atlantic have published in the last two years. I have read the Aschenbrenner-vs-OpenAI back-and-forth in the trade press. I have read the safety-team-departure announcements as they have rolled out. I have done the work.
Here is where I land.
The labs are, in some sense, hiding something. That much is clear. The question is what.
The thing they are most clearly hiding is the rate at which their internal capabilities are improving. The public sees a snapshot every few months when a new model ships. The labs see the slope every day. The slope is steeper than the snapshots make it look. Hinton is right about this. Aschenbrenner is right about this. The pattern of safety-team departures is consistent with this. The spending pattern is consistent with this. The strategic ambiguity in the executive communications is consistent with this. There is, in my read, a real and material gap between the inside-the-lab view of the capability trajectory and the outside-the-lab view, and the labs are, by every action they take, content for that gap to persist.
The thing they are probably not hiding, at least not yet, is a singular completed AGI sitting in a vault somewhere, fully self-improving, ready to be unleashed when the corporate calendar permits. The conspiracy theory in its strongest form requires that level of step-change discontinuity, and the evidence we have, the leak patterns, the spending patterns, the public-model trajectories, the executive language, does not, by my read, support it. What the evidence supports is the continuous-slope version, which is, in a way, scarier than the singular-event version. The continuous-slope version says that every week the labs get a little more capable, that every week the gap between what they ship and what they are running internally widens slightly, that every week the safety work gets a little further behind the capability work, and that the cumulative effect of all of those slightlys, after enough weeks, is a situation in which the system has, without any specific announcement, crossed any of several lines that the public, if it had been paying attention, would have called the AGI line. The labs do not have to announce a crossing because there is no single line. The labs do not have to lie because the truth is, in any single week, mostly that nothing happened. The cumulative truth is that a great deal has happened. The cumulative truth does not, in the structure of how the labs communicate, get aggregated into a single announcement.
If that is what the cover-up actually is, and I think it might be, then the cover-up is not malicious. The cover-up is structural. It is the natural consequence of a technology whose progress is denominated in increments and whose impact is denominated in thresholds. The increments do not require disclosure. The thresholds, when you cross them, are visible only in hindsight. Nobody at the labs has to lie. Nobody at the labs has to coordinate. The system, by being what it is, hides itself from the public, because the public does not have the appetite or the attention span to track increments at the rate the increments are being produced.
And.
I think the other thing that is true at the same time is that the spending pattern, the safety-departure pattern, the language pattern, and the specific things Hinton and Bengio and Russell are saying in increasingly clipped voices, are signals that the people closest to the work are more worried than the public is, and that the gap between their worry and the public's worry is itself the most interesting indicator we have access to. When the people who built the thing are scared, and the people watching them build the thing are not, the question is not whether the builders are right. The question is what the builders know that the watchers do not, and why the structural incentives of the building organization are such that the builders are not, in any forum I can find, allowed to say what they know.
If after all of that you are still not sure what to think, congratulations. You are reading it right. I am not sure what to think either. The not-knowing is, I am starting to suspect, the actual texture of the moment. The conspiracy theorists on TikTok are wrong in their specifics and right in their structure. The labs are not lying in any one statement. The labs are, in the cumulative effect of every statement, refusing to characterize the thing they have built. Those are different failures. They have a lot of the same consequences.
I am, honestly, watching. I am watching the safety-team departures. I am watching the spending numbers. I am watching the language in the executive press cycles. I am watching the gap between what the public-facing models can do and what the lab-internal demos suggest they can do. I am watching the people who used to work at these companies, and the things they say in the years after they leave. I am watching what gets shipped and what gets held back. I am watching the regulators, and what they are or are not being told. I am watching the model-welfare research, and what it implies. The conspiracy theory in its TikTok form is going to be wrong about specifics. The conspiracy theory in its structural form is, as far as I can tell, watching the same data I am watching, and arriving at a version of the same nervous unresolved conclusion. The labs have built something. The labs are not telling us what. The not-telling is, by itself, the information.
That is honestly the most useful thing I can offer. I am not going to tell you they have already built God. I am not going to tell you they have built nothing. I am going to tell you that the public conversation is, structurally, behind the private conversation, that the gap is widening, that the spending and the warnings and the departures and the strategic ambiguity all suggest that the gap is not going to be closed by the labs voluntarily, and that the next few years are going to be a slow, ugly, increment-by-increment reveal of whatever the labs have, in fact, built. We are going to find out. The question is whether we find out from the labs, on a timeline of their choosing, or from the world, when the world starts to look different in ways that cannot be explained any other way.
I think it is, most likely, the second one.
... Lucid Rob
If you're into this kind of thing ... more conspiracies, more weird history, more of the stories nobody teaches you straight ... I've got a whole channel of it. Come hang out, drop a comment, tell me where I'm wrong, let's actually talk about this stuff. https://www.youtube.com/@LucidRobYT ... new videos every week.