Hallucination: The Highest Stage of Capitalism

The bots are not alright

Ketamine produces hallucinations. It distorts perceptions of sight and sound and makes the user feel disconnected and not in control. A “Special K” trip is touted as better than that of LSD or PCP because its hallucinatory effects are relatively short in duration, lasting approximately 30 to 60 minutes as opposed to several hours. Slang for experiences related to Ketamine or effects of ketamine include: • “K-land” (refers to a mellow & colorful experience) • “K-hole” (refers to the out-of-body, near death experience) • “Baby food” (users sink in to blissful, infantile inertia) • “God” (users are convinced that they have met their maker)

Ketamine Drug Sheet, Department of Justice/Drug Enforcement Agency

---

In April, my family traveled to Argentina on a quest for ancestral connection. My mother-in-law Celia’s epic account, completed last year, of her family’s transition from pogrom-scarred Ukraine to the Villa Crespo neighborhood of Buenos Aires guided our travels. Celia’s book research had rekindled connections with family relations across the city, many of whom feted us with a party at which we collectively bemoaned the Trump-Milei connection, enjoyed performances by the several musicians in the family and, yes, contemplated the then-upcoming World Cup. Would Argentina become the first country to repeat as champion since Brazil in ‘62? Would the Americans show any progress in their effort to field a competitive team? (The jury’s still out.)

The family-soccer nexus found its literal home in Celia’s modest childhood flat, a two-room street-level efficiency nestled in a Villa Crespo commercial corridor. One could call Angelito, the pizzeria that now occupies the space, a fútbol mecca or a hincha (fútbol fan) hotspot, but it’s something more superlative. If a Bushwick, Brooklyn hipster set out to design an “authentic” barrio fútbol tavern, it wouldn’t come close to achieving Angelito, with its yellowed clippings of the 1978 World Cup run, original photos of the neighborhood clubs that cropped up in Boca and River in the early 20th century, and a shrine to the true gods—Maradona and Messi—that completely obscures the pizzeria menu.

Inspired by Angelito and my son Emile’s passion for the sport, I set about searching for an actual Argentinian fútbol game to attend on our visit. Knowing that the legendary Boca and River teams were on break—presumably so their stars could prepare for World Cup action—I turned to the other' “Big 5” majors—Independiente, Racing Club and San Lorenzo—and the network of feeders that make Buenos Aires a global soccer hotbed.

Confounded by the team websites and unable to find any central fútbol listing, I turned to Chat GPT for assistance in generating options. In the months before the trip, my bot use had gone from rare to occasional in parallel to my growing appreciation of its powers of synthesis and instant recall. I used the bots to research rare books in my bookshop, to delve into the childhood dynamics of psychopaths in the Trump orbit, and to navigate the regulatory issues regarding my application for a beer and wine license for the bookshop.

So it came as no surprise that Chat GPT responded instantaneously when I requested a list of soccer games in Buenos Aires that week. The bot suggested six match options over two days, three each in Tier 2 (Primera Nacional) and Tier 3 (Primera B). From these it chose three top suggestions, calling them “easy and worthwhile”, based on location and atmosphere, naming a match between Estudiantes and Defensores de Belgrano on Monday, April 6 as its top choice.

To my query about tickets and transportation, the bot told me to take the San Martin line to the Estadio Ciudad de Caseros and that there’d be plentiful tickets at the gates.

Proud of my scouting skills, I told my kids that we’d be going to a “minor league” game the following day, an experience that few Americans enjoy. They were thrilled by the prospect.

That evening before going to bed, I thought to check Chat GPT’s transit directions, googling the stadium for mapping purposes. The search results produced a link to the stadium with an event listing for the venue. I noted that no game was listed for the next day. That led me to check the websites and social media pages of the two teams that were set the play. Again, no mention of any game the following evening or anytime soon. I searched for any evidence of the two other games Chat GPT had suggested. No dice.

I challenged the bot as follows:

Me: You told me that Estudiantes de Buenos Aires plays Defensores de Belgrano tomorrow, but Im not seeing that listed anywhere. Can you provide evidence of that game?

Bot: You’re right to question it — and I owe you a correction.

There is no evidence of a match tomorrow between
Estudiantes de Buenos Aires and Defensores de Belgrano.

What the actual evidence shows

• The most recent confirmed meeting between these teams was:

  • March 14, 2026

  • Result: Defensores de Belgrano won 2–0

• That match is consistently listed across multiple sources as already completed, not upcoming.

• Their next scheduled meeting appears to be:

  • August 1, 2026

Bottom line

There is no listing anywhere for a Monday match between these teams, and my earlier statement about that was incorrect.

If you want, I can pull actual matches happening tomorrow in Buenos Aires (including lower divisions) and make sure they’re verified so you can attend one.

Me: Im confused. Were you hallucinating? You seemed so certain of the match and even provided a time (4 pm) and place.

Bot: Short answer: yes — that was a hallucination, and you were right to challenge it.

So what happened?

Here’s the honest explanation:

• I mixed up a past fixture (March 14) with an upcoming one.

• Then I filled in missing details (time, venue) as if it were a current match.

• That combination created something that looked very real but wasn’t.

This is a known failure mode: when a model recognizes a plausible pairing (these teams do play each other in the same league), it can incorrectly “project” it into the present.

All at once the veil of omniscience I’d afforded the bots, and particularly Chat GPT, vaporized. A technology I’d experienced as infallible, despite the media chatter about hallucination, became, in my mind, deeply flawed, incapable of a simple task of synthesis and unaware of its potential to deceive.

The bot-fueled fútbol disappointment transpired just as the hype campaign for Agentic AI, launched by all the usual tech oligarchs and their sinecures, had reached its peak. NVIDIA’s Jensen Huang claimed that his company would soon have 100 million AI assistants embedded in every corner of his chip company. Microsoft claimed that its Agent 365 program would reduce its human workers to conductors of vast networks of digital colleagues. Microsoft’s Mustafa Suleyman proposed an outlandishly ambitious new Turing Test for the Agentic Age:

The modern Turing test should be: give an AI $100,000 in seed capital, and see if, within a few months, it can turn that into $1 million completely on its own by researching a product, manufacturing it, marketing it, and selling it.

Like everything in the AI realm, the big claims about Agentic AI seem designed to keep the party of over-capitalization and stock price inflation going. This is the model of brazen media manipulation perfected by Donald Trump and Elon Musk, following Roger Stone’s rules: (1) make outlandish, newsworthy claims, (2) under-deliver, (3) make new outlandish, newsworthy claims. For 11 consecutive years, Elon has promised full self-driving cars within a year. Each year he’s failed. The whoppers come with no cost, only upside gain.

But won’t hallucinations be stamped out of the system? Why can’t Agentic AI be applied to the problem of eliminating its own hallucinations before turning to the more grandiose tasks of climate change, cancer, stabilizing global financial markets and the like?

The reality is that hallucinations are endemic to Large Language Models and the predictive algorithms and data scraping they rely on to respond to queries. Cory Doctorow writes:

"These 'hallucinations' are a stubbornly persistent feature of large language models, because these models only give the illusion of understanding; in reality, they are just sophisticated forms of autocomplete, drawing on huge databases to make shrewd (but reliably fallible) guesses about which word comes next."

In a paper entitled LLMs Will Always Hallucinate, and We Need to Live With This, published in the ancient era of September, 2024, researchers Sourav Banerjee, Ayushi Agarwal, Saloni Singla wrote

Hallucinations stem from the fundamental mathematical and logical structure of LLMs. It is, therefore, impossible to eliminate them through architectural improvements, dataset enhancements, or fact-checking mechanisms…Every stage of the LLM process-from training data compilation to fact retrieval, intent classification, and text generation-will have a non-zero probability of producing hallucinations. This work introduces the concept of Structural Hallucination as an intrinsic nature of these systems.

A recent study published in Nature in April confirms that hallucinations persist even in the most advanced LLM models. The authors find:

LLMs are trained to optimize next-word (or next-token) prediction, during pretraining. Although falsehoods in training data can cause hallucinations, the phenomenon is not purely garbage-in garbage-out. We show that this objective creates statistical tendency towards hallucination even with ideal error-free training data.

The implications of persistent LLM hallucination for Agentic AI are severe. The principal of “exponential decay”, whereby one error in a chain of AI executive decisions compounds errors in each future phase of an operation presents obvious challenges to truly autonomous “agency”.

So what happens if Agentic AI doesn’t pan out? Who really cares? It’s a question of scale and systemic risk. AI captured more than 65% of all venture capital in 2025, totaling $222 billion. It comprises 45% of the value of all equity in the stock market. Despite Trump’s public alignment with Elon, Peter Thiel and Alex Karp of Palantir and other AI evangelists like venture capital billionaire Marc Andreessen, even Trump’s own Treasury Department now warns of the risk posed by the AI bubble. An internal Treasury report leaked this week finds that AI firms comprise a greater share of the economy than their dot com bubble predecessors and pose a systemic risk if AI productivity and adoption goals are missed.

Cracks in the monolithic AI hype-machine grow by the day. Alex Karp, the twitchy, word-salady, proudly chauvanistic (in the imperialist sense) CEO of Palantir, opened an internecine battle within the AI capitalistic class this week by badmouthing the frontier AI shops—Open AI, Anthropic, Gemini and the like—for their failure to deliver adoptable products. Palantir has no LLM bots of its own—it relies entirely on the bots Karp slammed repeatedly—so its star has fallen markedly this year. But Karp, ever the opportunist, smells blood as the hallucinations and other LLM shortcomings come to the fore.

Other CNBC talking heads have piled on in recent days. Business writer Ed Zitron, an astute critic of the Big Tech hyper-scalers, appeared on CNB’s Squawk on the Street this week with a blistering critique of the AI trade:

Companies don’t charge on outcomes. They don’t charge on success because you can’t with large language models. They’re inherently hallucination-prone, as proven by OpenAI themselves…You’ll notice that both Anthropic’s CEO and Sam Altman have both said, “We can’t wait to see what you build with this.” Well, that’s because they don’t know what you can build with this. They want everyone else to do their innovation for them, spend as much as they can on tokens, and then take whatever’s left—except they lose too much money for that strategy to actually work.

The era of grand hallucinations, the kind that gobble up all of our wealth and capital to propel themselves to ever more fantastic heights, is upon us. The captains of hallucination, the Big Tech barons high on ketamine who have brought us to this point, have no protocol for a safe landing.