Robots Will Destroy Civilization?
Segment #1036
The important distinction is that there are really two separate questions: whether advanced AI presents a serious civilizational risk, and whether the very specific claim that AI has a >10% chance of killing everyone within the next decade is scientifically justified. The first has substantial support. The second is much more controversial.
The current controversy began with Jacob Coxon, who worked at both OpenAI and Anthropic and resigned from Anthropic. He said the companies are racing toward self-improving superintelligence and that people building these systems genuinely believe AI could “kill us all by the end of the decade.” Anthropic alignment lead Evan Hubinger then publicly backed the substance of the warning and put his personal probability at greater than 10% within ten years. (AP News)
Where Musk's skepticism gets interesting
Elon Musk has questioned not merely the prediction but the way the episode unfolded. He responded “Seems like a setup” and subsequently agreed with descriptions of it as resembling a “psyop,” noting Coxon's short Anthropic tenure and the extraordinary viral reach of a relatively new X account. Epic Games CEO Tim Sweeney similarly called the episode “choreographed.” (The Guardian)
That does not prove that Coxon's warning was coordinated. In fact, there is evidence cutting the other way: Coxon left Anthropic only four months after joining and before his Anthropic equity vested, which is an unusual way to conduct a publicity exercise designed to enrich oneself. He also directly denied that it was a marketing stunt. (Axios)
But there is a genuine irony in Musk becoming the skeptic. Musk himself has repeatedly warned about existential AI risk and has previously put catastrophic AI risk around 10–20%. Geoffrey Hinton has also estimated extinction risk in roughly the 10–20% range, while Anthropic CEO Dario Amodei has spoken of roughly a 25% chance that advanced AI goes “really, really badly.” (Axios)
So Musk isn't really saying “AI poses no existential danger.” His criticism is closer to: Why is this particular apocalyptic message appearing in this particular fashion at this particular moment?
And there are serious skeptics
Brad Gerstner of Altimeter Capital—an investor in both OpenAI and Anthropic—gave Business Insider perhaps the most concise response when asked whether AI would kill humanity:
“It's fucking nonsense.”
That's significant because this isn't simply an argument between people who understand AI and people who don't. Highly sophisticated investors, computer scientists and AI researchers disagree dramatically about the probability and especially the timeline. (Business Insider)
Gary Marcus, for example, considers near-term extinction unlikely. Other researchers argue that the evidence presently supports much more concrete dangers—cyberattacks, biological assistance, fraud, manipulation, autonomous weapons, misinformation and loss of human oversight—without establishing the enormous additional claim that an autonomous superintelligence will emerge and exterminate humanity before roughly 2036. (Business Insider)
The weakness in the “10% within a decade” claim
Nobody can presently calculate that 10% in the ordinary scientific sense. We don't have a statistically meaningful population of previous superintelligences from which to estimate failure rates. The number is fundamentally an expert subjective probability based on assumptions about several uncertain steps:
AI reaches broadly superhuman capability → AI becomes capable of recursively improving AI → improvement accelerates dramatically → humans cannot reliably align/control it → the system gains sufficient autonomy/access/resources → it develops or pursues objectives incompatible with human survival → existing governments and technical safeguards fail to stop it.
Each link is plausible enough to deserve serious study. But we don't know the probability of any of them with high confidence, much less their combined probability within ten years.
That's why saying:
“There is a non-zero possibility that sufficiently powerful AI could create an existential threat.”
is quite defensible.
Saying:
“There is greater than a 10% probability that AI will kill all humans within ten years.”
is a radically stronger proposition for which we currently lack empirical evidence capable of validating a number like 10%.
There is another wrinkle worth watching. Fear of catastrophic AI can produce perfectly legitimate arguments for safety regulation—but it can also benefit the largest existing AI companies. Extremely expensive licensing, testing and regulatory requirements are far easier for OpenAI, Anthropic, Google and other giants to satisfy than for small competitors. That doesn't mean the warnings are manufactured, but it means “AI safety” and “regulatory capture” aren't mutually exclusive possibilities.
My assessment would therefore be: take the underlying warning seriously but treat the precise extinction probabilities and ten-year timetable with considerable skepticism. The strongest case for aggressive AI safety doesn't actually require believing that humanity has a 10%, 25%, or 50% chance of disappearing. Even a much smaller probability of genuine catastrophe would justify substantial safeguards.
And there is a fascinating contradiction at the center of the entire story: if researchers at OpenAI and Anthropic genuinely believe there is a double-digit probability that the technology they are developing will exterminate humanity within a decade, why are they continuing to build increasingly powerful versions of it? That question may ultimately be more important than whether Coxon's 10% number is right. (Business Insider)
The Chinese Psyop
There is now enough evidence to say that a Chinese-linked influence effort against U.S. AI infrastructure is not merely hypothetical. But I would separate what has actually been documented from the larger—and much more speculative—possibility that China is deliberately promoting an “AI will destroy humanity” narrative.
What has actually been caught
In June 2026, OpenAI disclosed a campaign it called “Data Center Bandwagon”. It identified a cluster of accounts likely originating in China that used ChatGPT to manufacture English-language social-media posts while pretending to be ordinary Americans. According to OpenAI, the operators were likely associated with a Chinese technology company that had done work for provincial government clients. (OpenAI)
The subject they selected is particularly important: American AI data centers.
The accounts produced cartoons, comments and other material portraying data-center companies as greedy, blaming AI infrastructure for higher household electricity prices and attempting to amplify genuine American concerns about electricity, water and local development. OpenAI concluded that the operation had not achieved meaningful organic reach, but considered its target strategically significant because AI infrastructure is central to U.S. technological competition with China. (t.co)
Then the evidence became more interesting. In August, X reportedly identified a much larger suspected Chinese bot farm of approximately 200,000 accounts, within which roughly 200 accounts were specifically posting in ways that could manipulate the American debate over AI and energy policy. (Axios)
And this isn't occurring in isolation. Microsoft has documented Chinese influence actors using fake accounts and AI-generated material to identify and exploit divisive American political issues. Graphika has separately uncovered a pro-China ecosystem impersonating major Western news organizations. (The Official Microsoft Blog)
The strategy makes considerable geopolitical sense
Suppose you were designing an information operation intended to slow America's AI development without ever saying:
“China wants America to stop building AI.”
You wouldn't argue that China should win.
You would find authentic American grievances and amplify them.
The resulting messages could come from completely different political directions:
Data centers are raising your electric bill.
Data centers are draining your community's water.
Billionaire technology companies are destroying your town.
AI will eliminate your job.
AI is destroying the environment.
AI companies can't be trusted.
AI needs to be stopped until government can regulate it.
And at the extreme:
AI is going to destroy humanity.
That last narrative could be extraordinarily powerful because it transforms an economic argument into a moral one. If Americans can be persuaded that building increasingly capable AI represents an existential threat, then slowing American AI development becomes an act of protecting humanity rather than surrendering a technological advantage to China.
That would be a remarkably effective influence narrative.
But here's where the evidence stops
I have found good evidence for PRC-linked amplification of anti-data-center and anti-U.S.-AI-infrastructure narratives.
I have not found comparable evidence establishing that Beijing orchestrated the recent OpenAI/Anthropic whistleblower warnings about AI extinction.
That distinction is critical.
The whistleblowers and AI researchers may sincerely believe what they're saying. China's information operators don't need to invent the underlying argument. A sophisticated influence operation works better when it discovers an authentic controversy and amplifies whichever side benefits its strategic objectives.
That is precisely what OpenAI says happened with data centers: the operators exploited existing legitimate concerns rather than inventing the controversy. (OpenAI)
Independent researchers at Alethea have subsequently identified essentially the same pattern, finding Chinese, Russian and Iranian actors amplifying genuine American opposition to data centers. Their conclusion is important: foreign actors didn't create the grievance; they found it and magnified it. (Alethea)
The potential psyop is therefore subtler
The most plausible scenario isn't:
China invents AI-doomsday story → Chinese bots convince Americans AI will kill everyone.
It is:
Real AI researchers raise legitimate safety concerns → sensational extinction predictions attract enormous attention → authentic American anxiety grows → foreign influence networks identify the vulnerability → bots and content farms amplify the most frightening interpretations → public opposition to data centers and AI development increases → permitting and construction become politically harder → U.S. AI expansion slows.
China doesn't have to convince Americans that China is good.
It merely has to convince enough Americans that American AI expansion is bad.
That is classic asymmetric information warfare.
There is also an important counterargument
We shouldn't fall into the opposite propaganda trap and conclude that every American who opposes a data center is a Chinese dupe.
There are legitimate questions about electricity prices, grid capacity, water consumption, tax incentives, noise, land use and who pays for infrastructure expansion. WIRED interviewed experts who specifically cautioned that Chinese interference appears to be piggybacking on genuine domestic opposition rather than creating it. (WIRED)
That distinction actually makes the influence operation more sophisticated. The best propaganda isn't an obvious lie. It is often a selective amplification of something substantially true.
And that brings us back to the “AI will destroy civilization” story
This is why I would now be somewhat more cautious about the extraordinary amount of attention being given to the recent extinction predictions.
It does not mean Coxon, Hubinger, Hinton, Amodei or other AI-risk researchers are Chinese agents. There is presently no evidence I've found supporting such an accusation.
But there is a strategic coincidence worth investigating:
China wants to surpass the United States in AI.
America's principal advantage requires enormous additional computing and energy infrastructure.
PRC-linked operators have now actually been caught attempting to generate American opposition to that infrastructure.
And simultaneously, one of the most viral narratives in America is becoming:
The technology those data centers are being built to develop may kill all of us.
Those facts don't establish a conspiracy. They do establish an extraordinarily attractive information environment for a Chinese influence operation.
The question I would investigate next is therefore not simply “Are Chinese bots attacking data centers?” We already have evidence that some are.
The much bigger question is: