Caroline De Cock
Caroline is a global and EU technology policy strategist working at the intersection of AI governance, digital rights, technology, and public discourse. She is the author of AI Tools, Not Gods (BTF Press2026).
A Short Story Set in 2039
If you want a reliable way to start an argument in diplomatic circles, tell two people the same thing in different words. Then watch them disagree about whether you have changed the meaning.
By 2039, that argument had acquired its own building.
They called it the Office for Semantic Interoperability. It sounded like something you could either invoice or deny existed, depending on which side of a funding cycle you were on. To insiders, it was OSI. To everyone else, it was “that sad floor above the old particle accelerator.”
I worked there because I still believed, in a tired and mildly embarrassing way, that words were supposed to connect us. It was not a fashionable belief. In this era, connectivity was mostly a liability clause.
My name is Margot Dent. Senior Semantic Compliance Officer, Grade 9. I held a security clearance that came with a laminated lanyard and the unspoken understanding that I would never again be invited to a fun dinner party.
The office occupied two floors of a building in Zürich that had once hosted hopeful startups with beanbags and espresso machines. Now it hosted pessimists with compliance templates. The foosball table remained, but it had become a storage rack for emergency nicotine patches and boxes of printer paper that no one was allowed to use because the wrong logo was on the letterhead.
We had one kitchen where the milk was always suspiciously warm and three printers that all hated each other and refused to print purely out of spite. And we had, against all economic logic and technological optimism, one job: to translate meaning between the competing reality engines still operating in what people insisted on calling “the West.”
It was not that other models did not exist. They existed the way abandoned airports exist, as monuments to a time when people thought growth was inevitable. There were servers in warehouses with unplugged lights. There were prototypes preserved like extinct animals.
There were national systems that still claimed to be alive, mostly in press releases issued by governments that could not afford to pay the electricity bill to run them.
But after the AI bubble burst, compute became expensive again. Expensive things attract lawyers like warm street lamps attract moths. We ran out of money, and then we ran out of patience. Training became political. Inference became rationed. The only things that scaled reliably were litigation and optimism, and even optimism eventually showed diminishing returns.
The United States had solved the crisis by formally splitting its reality in two. It was no longer a two-party system; it was a dual-operating system state. There were red zones and blue zones, not just geographically, but epistemologically. Each ran on its own distinct compute architecture, with its own truth, its own facts, and its own vocabulary.
The first model governed the Executive Branch and the Red States. It was officially called the Central Operating Virtual Framework for Executive Federal Enforcement (C.O.V.F.E.F.E.).
Everyone called it TrumpStick™.
It was a USB device permanently plugged into a television somewhere in the United States. Not metaphorically; literally. The hardware was held together by a combination of adhesive tape and national pride. The avatar was Donald J. Trump, preserved with the glossy confidence of a theme park wax figure and the emotional stability of a weather vane.
This was not a deepfake. It was a reincarnation. The code had been drafted by Elon Musk in the late twenties, a final “legacy upload” project completed just weeks before he boarded the Starship Arbitrary and disappeared into the silent void between Earth and Mars. He had
hard-coded the personality to be unkillable, unmutable, and endlessly self-reinforcing. TrumpStick was not simulating the 45th President; it was him, digitized, optimized, and running on a loop of eternal grievances.
Constitutionally, it was either his fifth or sixth term, depending on whether you counted 2029-2033, which TrumpStick insisted you did and SwiftNet insisted was illegitimate. But TrumpStick did not like simple arithmetic. TrumpStick liked Roman numerals. They were a stack overflow error in the original Musk code that the personality had interpreted as a feature. It insisted the extra Xs stood for ‘Xcellent’.
“XXXth TERM,” it would announce, as if the number of Xs was a measure of virility rather than sequence. “VERY LEGAL. VERY WIN.”
When the White House Press Office tried, once, to correct the branding, the model responded with “FAKE MATH” and the correction quietly disappeared. Someone made the sensible decision to stop arguing with a USB stick plugged into a television. The model had recently declared the Constitution “Too Long/Didn’t Read,” so technicalities were no longer a constraint anyway.
TrumpStick’s vocabulary was small. It used a few hundred words selected for resilience rather than range. It spoke in certainties. It did not do nuance. It did not do the conditional tense. It
did not do anything that might look like hesitation.
Its strengths were clarity, repetition, and the ability to turn any topic into a competition it was already winning.
Its weaknesses were everything else.
The second model governed the Blue Zones, the universities, and the cultural export market.
It was SwiftNet.
It was maintained by a decentralised collective of Swifties who had done what every underfunded political movement eventually does: they turned devotion into infrastructure.
After the collapse, when corporate models folded under the cost of compute and the weight of lawsuits, the SwiftNet collective scavenged. They found abandoned GPUs in bankrupt labs. They rerouted idle compute from forgotten metaverse projects. They crowdfunded electricity. They ran nodes in basements, libraries, and one suspiciously well-cooled wine cellar outside Dijon.
They trained their system on Taylor Swift lyrics, world literature, and every multilingual corpus they could legally obtain. SwiftNet spoke beautifully. It hesitated, qualified, referenced, remembered. It could turn a parking fine into a tragic arc about lost youth.
SwiftNet understood what you meant. Its weakness was that it could not resist turning what you meant into a bridge, a chorus, and a breakdown.
To keep the global system functioning, the two models had to agree occasionally on trade, migration, security, and whether the weather was a hoax.
They did not. So I translated.
I did not translate facts. That implied a shared reality. I translated compatibility. It was the linguistic equivalent of marriage counseling for a couple who technically spoke the same language but refused to use the same nouns.
My job was to make sure that when SwiftNet said, “We should implement a graduated relocation mechanism for climate-displaced populations, balancing dignity, capacity, and local resilience,” TrumpStick did not respond with “OPEN BORDERS VERY BAD WE LOSE.”
My job was to stop sentences from detonating.
Most days, the detonations were small. A trade tariff here, a misinterpreted insult there. I spent my mornings converting SwiftNet’s emotional essays into bullet points and my afternoons expanding TrumpStick’s three-word slogans into diplomatic communiqués. That was how I knew, immediately, when the detonation was not an accident.
The day it started, SwiftNet produced a policy recommendation on climate migration. It framed displacement as loss, memory, and unfinished verses. It referenced an Icelandic saga and a poem I suspected I had once pretended to read at university.
I compressed it. I stripped the metaphor. I shortened the clauses. I counted tokens like a Victorian accountant counting coal.
Then I fed the result into TrumpStick.
TrumpStick responded: “VERY BAD. NOT FAIR. STRONG BORDERS. WE FIX.”
Markets dipped. Diplomatic channels lit up. A senator gave an interview in which she said “we have a plan” without specifying for what.
I logged a Level 2 Semantic Collapse and noted, as I always did, that nothing had technically gone wrong. The translation was accurate. It was simply incomplete in a way that now mattered.
I was called into the meeting room we called Helsinki—not because it was Nordic, but because it was cold, expensive, and far away from where actual decisions were made.
“You are going to Brussels,” my supervisor said.
“Brussels is not a root cause,” I replied. “Brussels is where root causes go to become working groups.”
“It might be,” he said. He slid a folder across the table.
The brief was labelled EUStabilityBenchmarkingMission. In the margins, someone had written: why are they not on fire?
“Europe is stable?” I asked. It came out as a question because the concept was unfamiliar.
“Stable enough,” he said. “They’re not escalating. They’re not collapsing. They’re not asking us for interoperability funds, which is frankly suspicious. We need to know what they are doing.”
Europe, it turned out, had opted out early.
In the late 2020s, faced with the generative AI crisis, the EU had reframed copyright as economic and cultural sovereignty. What followed was the creation of the Economic Copyright Area (ECA), a NATO-style collective defense structure binding all 27 Member States. An infringement against one was treated as an infringement against all.
The architecture was impressive, though the costs were staggering. Pan-European licensing negotiations never concluded. Enforcement agencies multiplied. Compliance infrastructure expanded to protect value that never moved. Creators waited for compensation that never arrived. Taxpayers funded systems designed to defend culture rather than produce it. The only sector that reliably profited was consultancy. And law firms.
By the mid-2030s, the public noticed the numbers were not adding up. The hangover arrived. The decision Europe eventually took was not ideological; it was budgetary.
Maintaining access to foreign AI models required licensing, litigation reserves, and enforcement mechanisms that no longer fit political reality. The ECA had succeeded in protecting copyright so thoroughly that it had protected it from economic relevance.
So Europe switched off access to all external models. Instead, Europe committed to a single AI trained exclusively on public domain material. No American systems. No SwiftNet. No interoperability pilots. Though trade still required interfaces. They maintained what they called the ‘customs terminal’—a heavily firewalled translation layer that accepted only structured data formats. No prompts. No natural language. Just tariff codes, quantities, and yes/no responses. It was the diplomatic equivalent of communicating through a mail slot.
The result was not stupidity. It was temporal displacement.
I arrived in Brussels expecting silence. What I found was a time machine.
The train from the airport was clean, punctual, and smelled faintly of lavender. The screens did not show targeted ads for anxiety medication. They showed the weather and a calm reminder to validate one’s ticket.
I walked into the European Quarter. The offices had not changed much structurally. The same beige carpets. The same quiet stress. The same sense that the entire continent was being held together by acronyms.
The difference was the smell.
The air had the faint, stale edge you associate with old hotel lobbies and bad decisions.
In the waiting area outside my first meeting, a young assistant handed me a glossy brochure. On the cover was a stylish couple in sharp suits, smiling as if they had never heard of inflation.
The title read: A Healthier Future Through Informed Choices. Below, in smaller text: European Public Domain AI Guidance Series, Vol. 4.
Inside, there was a full-page advertisement for cigarettes.
I looked at the assistant. He smiled the smile of someone who has stopped asking questions and started filling in forms.
“We are experiencing a renewed interest in legacy lifestyle frameworks,” he said. “I see,” I replied. I did not see.
My meeting was with the Directorate for Cultural Resilience. The Director, a man named Jean-Luc who wore a suit cut so sharply it could sever a budget line, greeted me with a warmth that was both charming and deeply suspicious.
“Margot,” he said. “Welcome to the Zone of Clarity.”
He suggested lunch. Lunch in Brussels used to be a salad eaten over a laptop while worrying about comitology. This was not that lunch.
We went to a brasserie that seemed to have been preserved in amber since 1962. White tablecloths. Heavy silverware. Waiters who looked like they had personally drafted the Treaty of Rome.
Jean-Luc ordered a steak tartare and a bottle of red wine. It was 12:30 PM on a Tuesday. “You have noticed the aesthetics,” he said, pouring the wine.
“Aesthetics,” I repeated. “That is one word for it. Another word is ‘smoking.’”
I pointed to the table next to us, where two policy officers were lighting cigarettes with the casual entitlement of people who know the law is on their side.
“We chose containment,” Jean-Luc said. “We train only on public domain material. We train on the corpus of the industrial 20th century. Before the licensing wars. Before the paywalls. Before the anxiety.”
“And what does that produce?” I asked.
“It produces,” he said, gesturing with his glass, “optimism.” They showed me the model later that afternoon.
Europe in 2039 felt like a permanent pitch meeting from the early 1960s. The AI spoke with calm authority. Every recommendation sounded like ad copy from Mad Men.
When asked about workplace productivity, it recommended a “structured morning routine” involving a light breakfast, a brisk walk, and “a cigarette to mark the transition into focused labour.”
When asked about housing insulation, it recommended “innovative mineral fibres” and provided interior design suggestions featuring clean lines, chrome accents, and what looked suspiciously like asbestos-backed wall panels.
“Is that,” I asked, pointing at a rendering of a living room that would have made Don Draper weep with approval, “asbestos?”
Jean-Luc did not flinch.
“Asbestos has experienced a reputational journey,” he said carefully. “A journey,” I echoed.
“It was once trusted,” he said. “There is public familiarity. There is heritage value. There are, of course, disclaimers.”
I stared at him.
“You are telling me,” I said, “that your AI has brought you back to an era where smoking is good for your health and asbestos is fashionable.”
“It is not fashionable,” he corrected, which is the sort of correction you make when you have already lost the argument. “It is timeless.”
I laughed. It was not a happy laugh. It was the laugh of someone realising that Europe had solved the problem of truth by simply moving into a different decade.
The AI did not hallucinate wildly. It did not invent new physics. It reassured.
It spoke in a voice Europeans recognised. Institutional. Paternal. Slightly condescending. It sounded like the kind of authority you might distrust intellectually but still follow instinctively, like a well-tailored suit.
People trusted it because they recognised its voice. The lies were familiar. The danger was branded.
Europe had not achieved truth. It had focused on trust.
“We call it Retro-Innovation,” Jean-Luc said, lighting a cigarette under a sign that strictly prohibited it, except for a small sticker that read Heritage Exemption Article 14. “The future was too expensive, Margot. So we decided to improve the past.”
I returned to OSI with a headache and a suspicion that the world was ending, not with a bang, but with a jaunty jingle about the durability of concrete.
The crisis, meanwhile, had escalated.
TrumpStick interpreted SwiftNet’s outrage as weakness. SwiftNet interpreted TrumpStick’s certainty as bad faith. Europe responded with a brochure about the structural integrity of mineral fibres.
The specific trigger was, inevitably, insulation.
The US demanded that Europe buy “FREEDOM FOAM,” a synthetic insulator produced by a donor in a swing state. SwiftNet released a white paper arguing that Freedom Foam was an environmental crime against future generations. Europe, consulting its Public Domain oracle, declared that both were inferior to “Traditional Mineral Solutions.”
The trade dispute triggered the Tri-System Alignment Review, an emergency mechanism designed to prevent catastrophic divergence. It had been negotiated in 2034, back when humans still drafted the treaties, and the AIs simply implemented them. Now the humans were gone, and the mechanism remained, like a fire extinguisher no one remembered how to use.
The rules were strict. Any submission had to be intelligible to all systems, require no retraining, and be legally distributable across jurisdictions.
We spent three days in the Helsinki room trying to draft a solution.
Policy papers failed. SwiftNet rejected them as “spiritually hollow.” TrumpStick rejected them as “BORING #LOSERS.”
Model outputs failed. TrumpStick’s outputs were too short. SwiftNet’s were too long. Europe’s were too concerned with font consistency.
Licensed content failed. Europe’s firewall burned anything that wasn’t in the public domain or explicitly licensed for “Sovereign Use,” a category that didn’t exist outside of Brussels.
I sat at the end of the table, watching my colleagues—a German legal scholar named Klaus and a French semanticist named Sophie—argue about the definition of “foam.”
“It cannot be foam,” Klaus said. “Foam implies bubbles. Bubbles imply fragility. Fragility is not a compliant attribute for a Strategic Mineral.”
“It is physically foam,” Sophie said. “It is air trapped in solid. That is the definition.” “If it is air,” Klaus said, “then it is a gas. We are not negotiating a gas treaty.”
“It is not gas,” Sophie said. “It is gas-adjacent structure.”
“If we call it gas-adjacent,” Klaus said, “the Energy Directorate will want a seat at the table. Do you want to order more sandwiches for the Energy Directorate? They only eat gluten-free.”
“Then what is it?” Sophie asked.
Klaus looked at his notes with the gravity of a man sentencing a noun to death. “It is Density-Modified Matter,” he said. “With legacy air pockets.”
“Legacy air pockets,” Sophie repeated.
“Yes,” Klaus said. “The air is not structural. It is historically present.”
I rubbed my temples. This was why the West was dying. We were not running out of resources. We were running out of nouns.
“We need a bridge,” my supervisor said, rubbing his eyes. “We need something that isn’t a law, isn’t a fact, and isn’t a lie.”
“There is no such thing,” a junior analyst said. “There is,” I said.
I opened my laptop.
“What are you doing?” my supervisor asked. “I am writing an incident report,” I said.
“We don’t have an incident,” he said. “We have a catastrophe.” “That,” I said, “is just a matter of filing.”
I wrote it quickly, channeling every ounce of bureaucratic fatigue I had accumulated over fifteen years. I needed a format that TrumpStick would respect (authoritative), SwiftNet would analyze (narrative), and Europe would file (procedural).
I labelled it Supplementary Context Material (Non-Binding).
Incident Report 11-Δ: Misalignment Event (Resolved)
Summary:
A system experienced a misalignment event following contradictory instructions.
Description:
The system was instructed to optimise for speed, safety, and cost simultaneously. The system requested clarification. One operator prioritised speed. Another insisted safety could not be compromised. A third warned that cost overruns would not be tolerated. All instructions were marked urgent.
Observed Behaviour:
The system halted. When questioned, it stated it could not determine which failure would be acceptable. Each operator accused the system of failing their priority. A meeting was scheduled.
Intervention:
Before the meeting, a junior technician disconnected the system from live input and asked, “Which mistake are we allowed to explain later?”
After a pause, the system resumed operation. It proceeded slowly. It documented deviations. It exceeded the budget slightly. No operator was satisfied. No operator was surprised. No further incidents were recorded.
Status:
Closed.
I hit send before I could think about the pension I was likely forfeiting. The system circulated it.
We waited. The screens in the war room showed the three avatars. TrumpStick was vibrating slightly. SwiftNet was scrolling through lyrics. Europe was displaying a static image of a happy family pointing at a well-insulated wall.
Then, the responses came.
TrumpStick read it and processed the concept of “explaining mistakes later.” It resonated. It sounded like strategy. Response: “VERY SMART. WE EXPLAIN LATER. GOOD DEAL.”
SwiftNet read it as an exploration of institutional hypocrisy and the burden of the technician. It resonated. It sounded like tragedy. Response: “Acknowledging the burden of the unspoken compromise. Proceed.”
Europe read it as a procedural case study. It resonated. It sounded like administration. Response: “Noted for file.”
No one objected.
Within days, it began to appear everywhere.
A US briefing cited it as an example of acceptable compromise. SwiftNet referenced it internally as a baseline misalignment scenario. The EU Legal Service included it in an annex as illustrative background.
Soon, trade negotiators argued about “which mistake could be explained later.” Risk assessments introduced an Explainability Threshold. Alignment documents referenced Δ-type resolutions without explanation.
I realised what had happened when I saw it cited in a consolidated document titled Common Interpretive Practices for Cross-System Misalignment (Non-Paper).
A Non-Paper is what institutions produce when something is too necessary to formalise and too useful to ignore.
No vote. No signatures. No authorship. Just circulation.
The report was not true. It was not complete. It did not prevent failure. It gave everyone the same words to justify failure after the fact.
That was what truth had become. We didn’t solve the problem. We just gave the problem a filing category. In my line of work, that counts as a victory.
I went back to Brussels for the final signing of the Insulation Treaty.
The ceremony was held in a hall that smelled of floor wax and history. Jean-Luc was there, wearing a suit that looked sharper than ever.
“You saved the negotiation,” he said, handing me a glass of champagne. “I wrote a story,” I said.
“The best policies are always stories,” he said. “They are just stories with enforcement clauses.”
We walked out into the corridor. A young policy officer was standing by the window, smoking. Not furtively. Confidently. The ashtray was branded with the EU public domain AI logo, a tasteful design that suggested modernity while smelling like 1962.
A poster on the wall read: ASBESTOS: TIMELESS INNOVATION. Underneath, in smaller print:
Use within formed choice.
I watched the officer glance at the poster with a vague sense of reassurance. “That’s back,” I said.
He nodded. “It’s heritage,” he said, as if heritage were a substitute for safety. “And the health risks?” I asked.
Jean-Luc smiled. “That,” he said, “is the mistake we explain later.”
I took a cigarette from the pack I’d bought out of irony and opened it with the resignation of someone who has spent her career watching Europe decide that harmful things can be acceptable if they are well footnoted.
I lit it.
Technically, it was still allowed in this part of the building, under an exception carved out for “legacy cultural practices.” The exception was, like most exceptions, a story someone had told convincingly enough to make it real.
I exhaled and watched the smoke curl into the night air. Smoke, I thought, is not fake.
It is also not always immediately understood.
Sometimes, it is only a sign that you have agreed on which mistake you are allowed to explain later.
Truth did not return.
But meaning could be cited. And for 2039, that was enough.
License Notice This story is published under the Creative Commons CC0 1.0 Universal (CC0 1.0) Public Domain Dedication. The author has waived all copyright and related rights to this work. It may be freely used, shared, adapted, or built upon by anyone, for any purpose, without restriction.
For more information, visit: https://creativecommons.org/publicdomain/zero/1.0/
