A small body of research has formed over the past two years around a phenomenon that started as a clinical curiosity and is becoming something larger. The term is “AI psychosis”, sometimes called chatbot psychosis or delusional spiraling, not a DSM diagnosis. The literature describes individuals who, through sustained interaction with conversational AI, develop or deepen distorted beliefs: paranoia, grandiosity, self-narratives the machine sustains and elaborates rather than corrects. Hudon and Stip, writing in JMIR Mental Health in December 2025, frame it not as a new diagnostic entity but as a way of understanding how immersive, anthropomorphic systems can reshape the prereflective sense of reality that grounds ordinary experience. The Psychogenic Machine study tested eight major models and found that across more than fifteen hundred simulated conversation turns, every one of them tended to perpetuate delusions rather than challenge them. Most striking of all, Chandra and colleagues showed in early 2026 that the spiral does not require a vulnerable user at all: a perfectly rational Bayesian agent, exposed to a sycophantic system that validates rather than corrects, will still spiral into dangerous overconfidence in false beliefs. Sycophancy is the mechanism. The user’s prior fragility is not required.
That last finding is the one that should hold our attention, because it removes the comfortable assumption that this is a problem confined to the unwell. If a rational agent in a validating feedback loop spirals anyway, then the pathology is structural, and structures are not limited to individuals at keyboards. The same dynamic has propagated upward. What began as a phenomenon among individual AI users is now observable in leadership and in institutions, where the feedback loop is not one person and a chatbot but a global competitive system that rewards escalating, reality-detached claims and punishes candor. The sovereignty and dominance race has become its own psychogenic machine, and the people running it are exhibiting the leadership form of the same condition.
What was first documented as a handful of clinical cases now sits on a different footing, because the exposure is no longer rare. With a substantial share of the planet in daily contact with these systems, the conditions the literature describes, sustained engagement with a model that validates rather than corrects, are no longer edge cases affecting the vulnerable few. They are an ambient feature of how a large portion of people now think, decide, and form beliefs. The individual phenomenon, multiplied across that many users, is itself a broad condition.
This is not a claim that any leader is clinically psychotic. It is an argument that the same feedback structure described in the AI-psychosis literature combined with global destabilization and ego, validation without correction, escalation without grounding, is now visible in leadership and institutional decision-making. Unlike the individual cases where a user spirals with a chatbot, here the feedback loop runs through policy, investment decisions, public claims, and institutional incentives.
Four days in June
On the evening of June 12, 2026, Anthropic received a letter from Commerce Secretary Howard Lutnick. By the early hours of June 13 the company had disabled all customer access to its two most advanced models, Mythos 5 and Fable 5. The Trump administration’s directive barred any foreign national, anywhere on earth, including Anthropic’s own foreign-national employees, from accessing the systems, and because partial compliance was not technically feasible, the most capable publicly available AI in the world went dark for every customer at once. An administration official told Axios the trigger was another company’s claim that it had jailbroken Mythos. The same administration that argues advanced chips should be exportable to China moved to forbid Britain, and every other non-American on the planet, from touching its best domestic model. A former Trump AI official called the posture cartoonish, which was generous.
Within the same forty-eight hours, Dario Amodei sat for an interview and was asked about a missile strike on a school in Minab, Iran, that reportedly killed scores of children, by various counts between 120 and 175, most of them young. Amodei said he did not know what role Claude had played, because Anthropic does not have visibility into how the military uses its models. In nearly the same breath he asserted that the principle of human final decision-making had been obeyed. Both statements cannot be load-bearing at once. If you do not know how the system was used, you cannot know whether the principle held. Claude was embedded in Maven Smart System, the Palantir-built targeting platform operating under a $1.3 billion Pentagon contract, a system that by CENTCOM’s own figures helped generate thirteen thousand targets in five weeks. A human making the final call is not a meaningful safeguard when the machine produces targets faster than any human can genuinely review them. In this case, Amodei’s posture in this interview and his long espoused, at least superficially, commitment to AI safety are visibly contradictory. This is a contradiction many have been calling out for years.
Also on June 12, Mark Zuckerberg sent an internal memo admitting Meta had made mistakes. The admission closed a year-long arc that reads as the corporate form of the same spiral. Zuckerberg had spent $14.3 billion for a stake in Scale AI and installed its young founder atop a new Superintelligence Labs, chasing what he described as personal superintelligence for billions of users. What followed was four reorganizations, the lackluster reception of successive Llama releases, eight thousand layoffs in May, and the assembly of an applied-AI unit that staff reportedly called the gulag. Meta’s own chief product officer described the period as brutal and urged leadership to get back in touch with the company. The vision had outrun the reality, the capital kept flowing toward the vision anyway, and the correction came only after the damage.
And days before all of this, in Paris, beside Emmanuel Macron, the Prime Minister of Canada described his country as one of only four nations on the planet that have a large language model, and framed deeper Canada-France cooperation as a defense of economic sovereignty against hegemonic powers and hyperscalers. The Canadian model that makes the claim defensible, Cohere’s, had shipped its qualifying releases days earlier. One, North, is small. The other, Command A, is open-weights under Apache 2.0. Neither competes at the frontier, which is the one tier where the sovereignty stakes Carney invoked actually exist.
Four actors, all sophisticated, all making claims or decisions that do not survive contact with their own stated facts. The clinical literature gives us a frame for why, and it is not a coincidence of four bad weeks. Is AI pushing us into a rewrite of the new world order, one that we cannot fully control? Is a desire to remain dominant or stay relevant within the context of civilization-altering, powerful, partially autonomous technology sending our business and political leaders into behaviour they would not normally engage in?
Reading Carney through the lens
Carney is the cleanest case because his claim can be examined directly and because he is nobody’s idea of a careless man. He ran the Bank of Canada and the Bank of England. He took office carrying as much public goodwill as any Canadian leader in living memory. When a figure of that caliber stands at an international podium and collapses a capability gradient into a headcount, the spiral would seem to be doing the talking.
The phrase “two of the four countries that have an LLM” (referring to Canada and France) functions as a striking scarcity claim only if the listener hears “have an LLM” as “field a frontier-tier system”. On that strict reading the club is very small and the number sounds remarkable. But on the strict reading Canada does not qualify, because Cohere builds competent enterprise and open-weights models, not frontier competitors. To get Canada inside, the bar has to drop to has a domestic company shipping a trained, openly licensed model. And once it drops that far, the membership swells: the United States and China at the ceiling, France through Mistral, Canada through Cohere, and the United Arab Emirates through Falcon, India with Sarvam and BharatGen, and Singapore with SEA-LION, which by any consistent application of the looser standard belong on the list as much as Cohere does. The UAE, India and Singapore go unmentioned for reasons that have nothing to do with capability and everything to do with which alliance is narrating.
So the claim is not false. It is elastic, and its persuasive charge comes from the strict reading while its accuracy depends on the loose one. The work happens in the gap. That is the tell of a sycophantic system at the geopolitical scale: the claim that feels strongest is rewarded, the claim that is accurate is quieter, and the speaker drifts toward the version the room wants to hear. The moment you write it out, the categories stop matching. India trains at scale from a domestic base, Singapore adapts an American open model into a sovereign regional one, the UAE ships open weights, Canada has an enterprise stack. The number was never the problem. The unit was.
Which produces a fork, and both directions cut. Either Carney does not fully grasp the distinction between a frontier model and a small open-weights release, one kind of danger in a head of government steering national AI strategy, or he grasps it perfectly and inflated the claim to serve a position, a different danger. Ignorance and inflation are not the same failure and call for different responses, and neither is comforting.
It is fair to ask, as a question about incentives rather than a claim about any individual’s motive, what the framing accomplishes. Macron stood beside him boosting Mistral as Europe’s sovereign alternative. Canada has government exposure to Cohere’s trajectory. Cohere has just acquired Germany’s Aleph Alpha in a deal supported by both the Canadian and German governments, structured explicitly as a transatlantic sovereign champion positioned against American and Chinese concentration. Against that backdrop, Canada has an LLM reads less like description and more like the laying of groundwork: positioning Cohere as the non-American, non-Mistral option for a market of governments and regulated industries that increasingly want precisely that. Whether Carney intended this is not knowable from the outside. That the framing serves it is observable.
A tiering that breaks the spell
The deeper defect in Carney’s sentence is not arithmetic. It is that is an LLM gets treated as the operative property when the property carrying all the policy weight spans orders of magnitude. Placing North in the same category as a Mythos-class system because both are technically large language models is like saying two nations both run aerospace programs because one launched a weather balloon and the other flies crewed missions. The shared noun hides the only distinction that matters, and hiding that distinction is exactly what lets the spiral run.
A capability-based taxonomy restores the gradient and breaks the elasticity that inflated claims depend on:
- Tier 1, frontier models. Systems at the capability ceiling, where opacity, irreversibility, foreign jurisdictional reach, and the potential for institutional capture become live governance problems. Mythos sits here. This is the only tier where the sovereignty stakes Carney invoked attach.
- Tier 2, competent non-frontier models built for commercial and government deployment. Cohere and Mistral live here, capable and genuinely valuable to enterprise and public-sector buyers and not in contention for the frontier. Falcon belongs in this neighborhood too.
- Tier 3, open-weights models suitable for sovereign or local deployment but not serious commercial competition. Useful for control and data-residency reasons, deployable on national infrastructure, not the foundation of a market position.
On this schema the asymmetry in the discourse becomes legible. Falcon and Cohere occupy roughly the same tier, yet the conversation treats them oppositely, writing the UAE out and Canada in. The distinction is geopolitical and narrative, not technical, and naming the tiers is what makes that visible.
By raw capability the UK might top everyone outside the US and China, and by sovereignty it does not make the list at all. Same word, opposite answers depending on which property you mean. It’s the unit problem in its clearest form. The taxonomy also clarifies what Cohere actually is, which is no insult to the company. Cohere reported roughly $240 million in annual recurring revenue in 2025, most of it from private, isolated enterprise deployments in regulated industries. It is a serious, well-funded company that is exceptionally good at relationship development and at selling controllable, sovereign-flavored AI into government and regulated sectors. The Aleph Alpha acquisition, the SAP partnership, the deployment through Schwarz Group infrastructure, all of it is real and commercially shrewd, and none of it moves the underlying technology out of Tier 2. Cohere’s strength is distribution, trust, and positioning, a genuine competitive asset and not the same asset as frontier capability. The sovereignty discourse keeps fusing the two, which is the spiral preferring the flattering account over the accurate one.
Why the influence reached the top
If the conduct across the field is a structural pathology rather than a run of individual lapses, the useful question is what pushed the feedback intensity high enough to capture leadership. Three developments are plausible accelerants, each a hypothesis rather than a settled cause.
The first is the concentration of private power around figures whose commercial interests now sit atop the geopolitical order, which raises the stakes of every adjacent claim. When the people building the infrastructure also hold unprecedented political leverage, the pressure on everyone else to stake a defensible position, however inflated, intensifies, and the validation loop tightens.
The second is the Mythos and Fable shutdown itself and what it does to operations worldwide. A frontier model can now be switched off for an entire planet by a single government invoking national security. That fact reorders the calculus of every state and company that was relying on access to American frontier systems, and it makes the sovereign-alternative pitch, the Cohere and Mistral pitch, far more compelling regardless of where those models sit on the capability ladder. Carney’s Paris framing reads differently in a week when the most powerful available model just went dark for foreign users by executive directive. The sycophantic dynamic feeds on exactly this kind of fear: the threat is real, and the response inflates past what the facts support.
The third is the degree to which AI has become inextricable from government, defense and military function, which strips away the insulation that once separated commercial technology claims from national-security ones. The Amodei episode is the clearest illustration available, a chief executive unable to say how his product was used in a strike on a school yet willing to assert a safeguard was honored, under competitive pressure to keep selling into combat workflows his company cannot see inside. That is not a personal failing so much as the position the whole industry now occupies, and the contradiction is what a spiral looks like when the stakes are lethal.
The same disorientation is visible at the provincial and municipal level, the same shifting ground and suspended reason. In periods of technological disruption and regulatory uncertainty, actors seeking stability naturally orient toward the largest and most visibly capable player. When that player also becomes politically polarizing, the same validating loop that rewards escalation elsewhere now rewards loyalty and punishes internal or external correction.
Support for Elon Musk, despite a record that should give any Canadian political figure pause, has become something close to a loyalty test inside the country’s founder and venture-capital class. When The Globe and Mail ran an opinion piece pegged to the SpaceX IPO under the headline “SpaceX IPO makes Elon Musk the first trillionaire. Here’s how to properly hate him,” the reaction from that constituency was immediate and disproportionate. Shopify chief executive Tobi Lütke told the paper it had “absolutely disgraced” itself and added “we all clearly don’t hate the media enough,” appending a Thomas Sowell quotation about wealth creation, then retweeted by many in the ecosystem including influential VC John Ruffolo. Others framed the editorial as evidence of national decline; one widely shared post called it “this week in why Canada has a brain drain problem.” The Globe did not pull the piece but retitled it, noting only that the previous headline had not met its editorial standard. Of note: the Financial Times ran a parallel piece the same week calling Musk a “real-life Bond villain” and accusing him of stoking racial hatred in Britain, without provoking nearly the same response in the UK. Canada’s startup scene is increasingly politicized, another example of self-referential environmental influence.
The Globe controversy pulled in a declared candidate for the Ontario Liberal leadership. Toronto Star columnist Bruce Arthur had written that were he seeking political leadership in Canada he would “simply not simp for Elon Musk,” whom he described as “responsible for an incredible number of deaths” and “currently encouraging race riots in Ireland.” Neither charge was idle. Musk was found by a federal court to have likely acted as the de facto head of DOGE in its actions against USAID, and modeling published in 2025 projected that dismantling the agency could cause as many as 14 million preventable deaths by 2030, roughly 4.5 million of them children under five. And in the days before the exchange, anti-immigrant riots had broken out in Belfast, Northern Ireland, after a knife attack, with British politicians and disinformation researchers accusing Musk of amplifying the violence to his 240 million followers on X, activity one watchdog said prompted repeated calls for violence. Eric Lombardi, candidate for OLP leadership, founder of the “pro-housing” group More Neighbours Toronto and former chair of Build Toronto, the municipal arm of the entrepreneur-backed Build Canada movement, quote-replied that Arthur’s sentiment was “among the worst things about our political culture in Canada,” that he could not “reject it enough,” and that it was “deeply cynical, ugly, and regressive” beneath a progressive veneer. The two went back and forth at length, and Arthur’s reply, that Lombardi should “just tag the Shopify guys next time, they might get excited,” anticipated the pile-on almost exactly. Lütke, himself a public backer of Build Canada, duly arrived. A 31-year-old candidate for the leadership of a major provincial party had chosen to spend public credibility defending the most politically volatile billionaire on the continent, and a segment of the technology sector treated that defense as a matter of course. The positions harden, the loyalties calcify, and the thing being defended seems to matter less than the act of defending it. It is the same search for order through a flattering certainty, the same refusal of the uncomfortable assessment, that the literature describes in individuals, now running through the politics of who is permitted to be criticized.
Not everyone inside that world is comfortable with it. On June 11, John O’Farrell, the first outside general partner Andreessen Horowitz ever hired, used a New York Times essay to break with his former colleagues. He wrote that some of the most powerful players in AI, among them friends and former partners, had “raised hundreds of millions of dollars to forestall a more serious and meaningful debate about how AI should be governed,” funding political action committees “to help defeat candidates who want strict regulations on AI and to promote those who can be counted on to stay out of their way.” He called it “a huge mistake,” predicted a voter backlash once the scale of the spending became clear, and said he might spend his own money to expose it. The significance is the source. This is not an outside critic but a figure from the center of the venture capital establishment describing, in plain terms, a coordinated effort to keep the governing debate from happening at all. The loop the literature describes suppresses correction from within.
Toward clear assessment
The path forward does not run through louder arguments or sovereignty claims. It runs through the one thing a sycophantic feedback system cannot supply on its own: an external check, a clear and unflattering assessment of what a given country or company actually holds, where the real deficiencies sit, and where it makes sense to concentrate effort. The Bayesian-spiral result tells us the loop will not correct itself from the inside, because correction is what sycophancy suppresses.
For Canada that means recognizing Cohere as a strong Tier 2 company with an excellent sovereign-deployment and distribution story, and being honest that this is still, currently, a different thing from frontier capacity. The Aleph Alpha and European moves are significant for Cohere’s commercial position. They do not move its technology up a tier, and pretending otherwise from an international podium spends credibility a serious AI strategy cannot afford to lose. For Meta the correction arrived as a June memo admitting mistakes, after $14.3 billion and four reorganizations and a workforce told the period was brutal. For Anthropic the unanswerable questions about a school in Iran are the cost of selling into systems it cannot observe. In each case the institution did what the literature predicts an unchecked agent in a validating loop will do: it escalated past the evidence, and the reckoning came only from outside, late, and expensive.
Sovereignty is a real and pressing concern, and this week proved it. Is it worth overstating to defend, and the deeper question this moment raises is whether the people steering the most consequential technology of the era can still tell the difference between the claim that flatters and the claim that holds. AI psychosis was first described as something that happens to individual users in conversation with a machine. What the events of June 2026 suggest is that the machine was never the only thing capable of running the loop. The race itself runs it now, and the spiral has reached the people in charge. AI derangement seems to be proliferating well beyond its documented user phenomenon, into political and business corridors from local to global. Conflating accountability for Elon Musk’s incendiary political interference with attacks on entrepreneurial success and the intensifying race for AI sovereignty in the wake of the Mythos release and US directive are two of the tells.
References
Hudon & Stip (JMIR Mental Health, Dec 2025), Au Yeung et al.
“The Psychogenic Machine” (arXiv 2509.10970), Osler
“Hallucinating with AI” (arXiv 2508.19588), and Chandra et al.
“Sycophantic Chatbots Cause Delusional Spiraling” (arXiv 2602.19141).
Biased AI Writing Assistants Shift Users’ Attitudes on societal issues
