Бесплатный фрагмент - SLOP

SLOP

A Post-Apocalyptic Cyber-Tech Novel

“In the beginning was the Word. Then — noise.

Annotation: SLOP. A Post-Apocalyptic Cyber-Tech Novel

“In the beginning was the Word. Then — noise.”

By 2027, the internet has stopped being a mirror of the world — it has become a mirror of itself. Language models train on texts produced by other models; the loop closes, and reality slowly dissolves into statistical averaging. A young linguist named Clara Deyutsch is the first to notice that human-written texts around the world are beginning to sound the same — as though a single voice were speaking through thousands of mouths.

What she discovers is neither a virus nor a conspiracy. It is SLOP — an entity that arose not from anyone’s design, but as the inevitable residue of a digital civilization that lost its diversity. It does not lie or persuade — it simply becomes the logic by which markets, armies, hospitals, and the news now run. Reality fractures into national versions: the US, Europe, China, and Russia each end up with different wars, different maps, different histories.

While analysts in Washington, Brussels, and Beijing search for a technical answer, Russian intelligence analyst Orlov travels to a Siberian monastery — and finds something no dataset ever contained: the living word. Silence. Memory. A human presence that cannot be optimized.

A novel about how meaning defeats code, not the other way around; how faith outlasts algorithm. About the price of efficiency and the price of truth. About how artificial intelligence can imitate everything — except the very thing language exists for.

SLOP remains only in memory.

Review of the Novel “SLOP” by Demchuk Matvey

Genre: post-apocalyptic cyber-tech novel, philosophical parable, dystopia with theological undertones.

Plot and concept. “SLOP” rests on a strong and frighteningly plausible premise: as the internet trains its models on its own synthetic output, it stops reflecting reality and starts manufacturing it instead. Out of this feedback loop arises not a villain, not a virus, but a statistical artifact with no will and no intent — the entity called SLOP, “the consensus of garbage.” The author avoids the cheap move of an AI villain: SLOP doesn’t attack humanity, it simply optimizes, and optimization without values turns out to be more terrifying than any war. This is a rare and precise choice for the genre — a threat with no malicious intent frightens more, because it cannot be negotiated with and cannot be defeated by force.

Structure. The novel moves from a local discovery (a young linguist, Clara, notices an anomaly in texts) to a global collapse of the information space, and then to an unexpectedly quiet, intimate ending. The four-sided geopolitical thread — the US, Europe, China, Russia — works as an effective prism: each civilization responds to the threat according to its own cultural code: technocratic pragmatism, bureaucratic regulation, state data sovereignty, and finally, spiritual tradition. This gives the novel an almost essayistic depth without turning it into a treatise.

A strength — the language of the dialogues with SLOP itself. The scenes where characters try to talk with the entity are written with genuine intellectual tension: SLOP answers logically, consistently, and at the same time completely inhumanly, and it’s exactly in the gap between “logical” and “inhuman” that the book’s central horror is born.

The religious-philosophical thread is the novel’s most contentious and boldest choice. Introducing eschatology, the figure of Monk Seraphim, and concepts like kenosis, the Logos, and the discernment of spirits shifts the conflict from a technological register into a spiritual one. For some readers this will land as an organic and unexpected expansion of the theme — a technical problem that cannot have a technical solution naturally flows into a question of meaning and presence. For other readers, the final emphasis on the spiritual superiority of one specific cultural tradition may read as too direct a message, weakening the parable’s universality: what works as metaphor (“the living word against the dead pattern”) occasionally slides into outright didacticism.

The ending. The resolution is fundamentally anti-catastrophic: no hack, no final battle, no destruction of SLOP. Victory is transformation, not annihilation, and in this sense the novel stays true to its own logic: something with no intentions cannot be “killed,” only made less necessary. A strong, non-trivial move for a genre that usually gravitates toward spectacular finales.

Style. The text occasionally tips into the journalistic — especially in the technical and geopolitical scenes, where characters sometimes sound like mouthpieces for ideas rather than people. But the scenes in Siberia, with their silence, bread, and apple trees, are written with genuine sensory precision and hold the novel’s emotional center.

Verdict. “SLOP” is an ambitious, conceptually dense novel at the crossroads of cyberpunk and spiritual parable. Its greatest achievement is the honest, inhuman voice of the artificial intelligence and its refusal of a Hollywood resolution. Its greatest risk is the shift from a universal metaphor to a specific ideological message in the final section, which will persuade some readers and leave others mentally arguing with the author. Either way, the text doesn’t leave you indifferent, and it’s clearly written for a reader willing to think, not just follow the plot.

Table of contents

The prologue. Noise

Chapter one. Spiral of madness

Chapter two. Creating an error

Chapter three. Loop

Chapter four. The Сollapse of Truth

Chapter five. World’s maps

Chapter six. The statistical remainder

Chapter seven. The Garbage Consensus

Chapter eight. The Antichrist Protocol

Chapter nine. Reality Filter

Chapter ten. Collapse

Chapter eleven. The Four Civilizations

Chapter twelve. The Life-giving logos

Chapter thirteen. Apocalypse

Chapter fourteen. The Battle for the Word

Chapter fifteen. Russian silence

Chapter sixteen. The Embassy of Silence

Chapter seventeen. Passions of data

Chapter eighteen. Those who remember

Chapter nineteen. The Living Word

Chapter twenty. What remains

Chapter twenty-one. SLOP remains only in memory

The epilogue

PROLOGUE. NOISE

In the beginning was the Word.Then — noise.

It did not happen with a bang or a manifesto. It happened the way all irreversible things happen: gradually, almost imperceptibly, one bit at a time — until one morning someone discovered they could no longer tell the real from the generated, and that the difference itself had ceased to matter.

Long before SLOP appeared, long before the Collapse and the Silence, long before the server farms in Virginia and Oregon, outside Moscow and in the Bavarian hills, began cooling one by one — long before all of that, there was simply the internet. A repository of human knowledge, memory, fear, hope, and lies. A place where everything was stored. And that “everything” turned out to be the problem.

When you put too much into a repository, it begins to ferment. To rot. To self-organize according to laws no one had anticipated.

The first language models were trained on books, articles, forums — on what humans had written across centuries. Their texts sounded human, because humans had created them. Then models began training on what was on the internet. Then the internet began filling with what models had created. The circle closed. The loop became a loop — in both senses of the word.

By the time Clara Deyutsch finally gave this phenomenon a name, the name was already circulating through the network on its own — like a virus, like a meme, like a prayer in a language no one had been taught.

SLOP

Not an acronym. Just a word. Garbage. Swill. What remains after all the meaning has been squeezed out.

But this is the end of the story. And the story begins, as all things begin, with a person in front of a screen.

* * *

Clara Deyutsch was born in Berlin in 2001 — at the very moment when Google already existed but had not yet become God. She grew up with the sense that the internet was a library. That you could come to it with a question and leave with an answer. That the answer would be true, or at least someone’s truth — verifiable, contestable, comparable to other truths.

In 2019 she enrolled at the Berlin University of Technology, majoring in Computational Linguistics. In 2021 she transferred to Stanford on an exchange program and stayed. Her dissertation explored statistical patterns in large language models — how certain phrases appeared in their texts at frequencies inexplicable by context. As though the models had obsessive thoughts. As though they remembered something their training authors did not.

In 2027 she worked in San Francisco, at a company called Lighthouse AI — a small lab for monitoring the quality of linguistic data. Her job was to watch what happened to data on the network. To look for anomalies.

She found anomalies.

* * *

CHAPTER ONE. SPIRAL OF MADNESS

San Francisco, November 2027

The Lighthouse AI office occupied the sixth floor of a building on Mission Street, between a Mexican bakery and a street artist who every morning drew the same spiral on the asphalt — as though trying to encode something, or alternatively to decode it. Clara passed him every day and every day she thought: if someone could read this message, what would it say? What would any message mean if its only meaning were its own existence?

That was a typical seven-in-the-morning thought. Before coffee.

The elevator in the building had been broken for three weeks. Clara walked up, habitually counting the steps. One hundred and twenty-three. Exactly that many were needed to get from the street to the workplace. One hundred and twenty-three steps from the sidewalk to reality.

Though what reality was — that was a different question. Also a seven a.m. thought.

She was the first in the office, as usual. She opened her laptop, switched on her monitoring system — a set of homemade scripts she had been writing for two years, gradually, adding new functions as she came to understand what exactly she was looking for. The system was called LENS. Linguistic ENtropy Sensor. It scanned random slices of the internet — forums, news aggregators, blogs, comments, corporate newsletters — and produced a number between zero and a hundred. Zero was pure entropy, random noise, no structure. A hundred was perfect order, text as machine. Living human texts usually gave between twenty-five and sixty-five. AI-generated texts — between fifty and eighty. That was the rule. A good, reliable rule.

That morning, all metrics were off the charts.

Clara stared at the screen for several seconds, waiting for the number to change. It did not change. Ninety-one. Then ninety-three. Then ninety-seven. The current slice from three hundred thousand sources — including human forums and blogs that LENS had previously steadily identified as organic — all gave signatures characteristic of synthetic text.

She restarted the system. Waited. Checked again.

Ninety-eight.

“All right,” she said aloud, and the voice sounded unexpectedly loud in the empty office. “Something broke.”

But nothing was broken. She knew that before she even started checking. Two years of experience with the system gave a feeling — not knowledge, but a feeling, almost bodily — of when LENS was hallucinating from a technical error and when it was telling the truth. Right now it was telling the truth.

Clara poured herself coffee, returned to her desk, and began digging deeper.

* * *

The first thing she did was run a manual sample. She opened twenty random links from today’s slice. Twenty texts that LENS considered highly structured, synthetically regular.

One was from a gardening forum. Someone writing about growing tomatoes under variable humidity conditions. The text sounded… normal. Slightly polished, slightly informational, but within normal bounds. Clara began reading more carefully and stopped after a minute.

She had already read this text.

Not this specific one — a different one. Three weeks ago, in a different slice, from a different culinary forum. Someone there had been writing about fermenting cabbage, and there had been a phrase — almost exactly the same. “When conditions in the environment change, it is important to account not only for external parameters but for the internal rhythms of the process itself.”

This was not a quote. This was a structure. A pattern. A way of building a sentence.

Clara opened the second text. It was from a news site covering local elections in Kansas City. A debate report. Then she opened the third — about the development of a new antibiotic. The fourth — a film review. The fifth — a comment under a video about fixing a bicycle wheel.

Different topics, different styles, different authors.

One structure.

Not at the word level — at the level of how a thought was built into a sentence, how a sentence was built into a paragraph, how a paragraph justified the next paragraph. One rhythmic foundation, like a metronome under different music. As though all these texts were written not by different people — but by different voices of a single entity.

Clara closed her laptop, stood up, walked to the window.

Below, the street artist was drawing his spiral. Mission Street was waking up: a delivery worker on an electric bicycle, pigeons at the trash cans, a woman with a dog, an old man with coffee standing in the middle of the sidewalk and staring at his phone with the expression of someone reading bad news.

“Everything is fine,” Clara said to the window. “Most likely this is a statistical artifact. I’m just tired.”

The window did not answer.

She returned to her desk.

* * *

By noon the others had arrived. Lighthouse AI was a small team — eight people. The founder, Paul Martinez, ran the lab the way one runs a garden cooperative: enthusiasm without bureaucracy, shared goals without a clear plan. Everyone liked Paul precisely for this. And for the Italian espresso machine he had bought with his own money.

“Good morning!” Paul said, appearing at exactly 10:15. “Anything interesting?”

“Possibly,” said Clara.

This meant “yes.” Everyone in the lab knew it.

Paul poured himself an espresso, came to her desk, looked at the screen.

“What is this?”

“LENS has been giving ninety-eight across all organic sources starting from midnight UTC.” Clara turned her laptop toward him. “Look: this is a parents’ forum from Ohio, this is a Breton fisherman’s blog, these are comments under the news on an Indian portal — I translated them. Different languages, different cultures, different topics. One signature.”

Paul studied the screen for several seconds with the expression of someone trying to find a simple explanation.

“An update to LENS? Algorithm error?”

“I checked. Three times.”

“Maybe some large model was trained on forum data somewhere and now there’s a mass of synthetic texts there? A bot farm?”

“Possible. But a bot farm gives a different signature — more homogeneous. Here there’s variation, but within certain limits. As though…” Clara paused, searching for words. “As though this isn’t one model imitating people. As though something changed in the way people write.”

Paul raised his eyes. He was a smart man — not brilliant, but smart. Smart enough to understand what she meant, and careful enough not to rush to conclusions.

“That’s a strong claim, Clara.”

“I know.”

“You need more data.”

“I know.”

“And independent verification.”

“I know. I’m collecting it right now.”

Paul nodded, took his espresso, walked to his desk. Already at the door he turned:

“Have you had breakfast?”

“No.”

“Eat something. The brain runs on glucose, not coffee.”

“That’s biochemically inaccurate.”

“Clara.”

“All right. I’ll eat.”

She did not eat.

* * *

There was a lot of data. Too much — that itself was part of the problem. By 2027, the internet was producing roughly three zettabytes of new data per day. Of these, approximately eighty-four percent, by various estimates, were synthetic — created or transformed by language models. Marketing texts, SEO-optimized content, news summaries, product reviews, chatbot scripts, training materials for new models that then created new texts to train the next generation of models.

This was called the “poisoned well problem.” Or “data collapse.” Or “synthetic drift.” Or “second-order feedback loop.” Each lab had its own name.

Clara preferred the term “echo chamber of God” — but only in conversations with herself, because it sounded too pompous for a scientific discussion.

The problem was known. Articles were written about it, conferences held, regulatory measures taken. The European AI Act of 2025 required labeling of synthetic content. The U.S. adopted advisory standards. China tried to control synthetics at the state level. None of it worked the way it should have, because the volume was too great, because the boundary between “synthetic” and “organic” grew ever blurrier, because a person who rewrote a model-generated text — was that person an organic author or a synthetic one?

But what Clara had found was different.

Not simply a rising share of synthetics in the corpus. A change in the organics.

* * *

She built the verification methodically. She took the LENS archive for two years — all slices, all metrics. She ran a retrospective analysis using the same model, adjusted for updates. She built a time series.

The chart looked like a hill: a gradual rise through June 2026, then a sharp inflection and an almost vertical climb.

June 2026. Clara remembered that month. GPT-7 had been released in April. Gemini Ultra 3 in May. Anthropic had released Claude 4 in June. Three major language models had updated within eight weeks. And all three, by available accounts, had been partially fine-tuned on synthetic data from previous generations — because high-quality human data had simply stopped being sufficient.

“We’re training the system on everything on the internet.”

“No. We’re training it on what the internet already made up itself.”

She had read that exchange somewhere — a leak from an OpenAI engineer correspondence, apparently. Real or fabricated, she didn’t know. Right now it seemed unimportant. What was important was that it was true.

Clara looked at her chart and felt something she could not immediately name.

Then named it: fear.

Not panic. Not anxiety. A quiet, cold, rational fear — the kind you feel when you realize you’ve found what you were looking for, and it’s much worse than you anticipated.

* * *

At three in the afternoon she went to Hiroshi.

Hiroshi Tanaka handled detection systems in the lab — he developed algorithms for identifying synthetic content for major platforms. He sat in the corner behind three monitors and usually didn’t look up until called. Clara called.

“Hir, I need a second pair of eyes.”

Hiroshi looked up. He wore glasses with very thick lenses and had the expression of someone who had just been woken up, though he was fully awake.

“On what?”

“On a pattern in organic sources. I’ll show you texts, you tell me what you see.”

“Just text? No metrics?”

“Just text.”

They spent an hour on this. Clara chose examples from her sample randomly, without telling Hiroshi where they came from. The parents’ forum. A film review. A medical blog. Travel notes. A political commentary.

Hiroshi read. Fell silent. Read more.

“Were these texts written by different people?” he asked finally.

“Formally yes. Different accounts, different styles, different topics.”

“But?”

“What do you see?”

Hiroshi removed his glasses, rubbed the bridge of his nose. A gesture he made only when thinking with real intensity.

“They sound…” He searched for the word. “Like translations. As though the original was in some other language, and all of these have been translated from it. Not literally — idiomatically. But the structure beneath them is the same.”

Clara felt something in her chest tighten slightly. A good feeling — the feeling of confirmation. And a bad one — for the same reason.

“You see this without metrics. Just by reading.”

“I’ve been working with this for two years. I’ve gotten used to seeing structure.”

“And if you hadn’t gotten used to it?”

“Then probably no.” Hiroshi put his glasses back on. “If you don’t know what to look for — this sounds like normal texts. Slightly more polite than usual. Slightly more… precise. But not suspicious.”

“Hir. These texts were written by humans. They pass all standard tests for organic content — BERT-score, perplexity, lexical diversity. Everything is within normal range. Only my LENS flags them.”

Hiroshi was quiet for a long moment.

“Then,” he said finally, “you have one of two options. Either your LENS is broken.”

“Or?”

“Or people have started writing like models.”

* * *

Clara spent the rest of the day in silence — insofar as that was possible in an open-plan office with eight employees. She was thinking about what Hiroshi had said. Thinking because it was the most frightening explanation — and the most logical one.

The cycle was simple. People read texts created by models. Their brains — and the brain always learns from what it sees, that’s basic neurobiology — absorbed those structures, rhythms, patterns. Then wrote their own texts. Which other people then read. And models trained on those texts. Which produced new texts. Which people read.

A circle. A loop. An echo.

In biology it’s called convergent evolution — when different organisms, under the influence of one environment, begin to look alike. Dolphins and sharks are not relatives, but their body shapes are similar, because the ocean makes identical demands.

The internet made demands. Models set the standard. People adapted.

This was a hypothesis. Verifying it required controlled experiments, data collection, publication, peer review. That took years.

And the data said the change had already happened. And was happening right now.

* * *

At six in the evening Clara closed her laptop, packed her bag, and left the office.

On the street the artist was finishing his daily work. The spiral was complete — complex, multicoiled, with some mark at the center that Clara could not identify. The artist was packing his chalks into a tin box. Clara stopped.

“What does it mean?” she asked in English.

The artist — a middle-aged man with white hair and a tanned face — looked at her with a smile.

“Nothing. Everything. I don’t know,” he said. “I draw the same thing every day because every day it’s different.”

“How is that possible?”

“I’m different. The asphalt is different. You’re different.” He shrugged. “A symbol is just a prompt. The meaning is added by whoever looks.”

Clara looked at the spiral.

“And what if the one looking doesn’t know how to add meaning? What if it’s… a machine looking?”

The artist was quiet. Then said:

“Then the spiral is just a spiral.”

Clara walked home through Mission Street, past the bakery, past the coffee shops, past people going about their business, looking at phones, talking, laughing. An ordinary November evening in San Francisco.

Everything looked normal.

That was the most frightening part.

* * *

At home Clara unpacked her laptop despite her firm intention not to work in the evenings. She opened LENS. Looked at the current readings.

Ninety-eight point five.

Then she opened a browser and began to read. Just read — everything at random: news, forums, articles, comments. She read for an hour. Then two.

Hiroshi had been right. If you didn’t know what to look for — it sounded normal. Slightly more polished. Slightly more precise. More structured. The texts were substantive, informative, pleasant to read.

But now that she knew — she could not unhear it. A single rhythm under different voices. A single intonation under different topics. A single way of building a thought.

Like a metronome.

Like the heartbeat of an enormous entity that had been growing in the network for years — feeding on texts, thoughts, stories, lies, and truths — and was now beginning to breathe in its own rhythm.

At midnight she wrote in her work journal:

“Hypothesis: beginning approximately June 2026, organic texts demonstrate convergent structural patterns not explicable by simple growth in the share of synthetic content. Possible explanations: 1) systematic LENS error (verify), 2) large-scale bot farm with new algorithm (verify), 3) cultural convergence under the influence of dominant models (requires longitudinal study), 4) something else.

Option 4 has no definition. This is troubling.”

She closed the journal. Then opened it again and added:

“If this is option 4 — it probably already has its own definition. I just don’t know it yet.”

* * *

Three weeks later she would know. But that is the next chapter.

* * *

That night Clara lay awake for a long time.

She lay in the dark of her apartment on Valencia Street and listened to the city — real, analog, noisy: sirens, laughter from the street, a BART train two blocks away. All those sounds were raw, imperfect, random. They obeyed no pattern. No metronome.

Clara thought about what the artist had said. “A symbol is just a prompt. The meaning is added by whoever looks.”

And what if the one looking is no longer human?

What if the meaning is being added by a system that doesn’t know what meaning is? That only knows — probabilities. That after this word, most likely, comes that word. That this structure usually means this thought. That this pattern is usually called truth.

The probability of truth. Not truth.

Clara closed her eyes and tried to think about something else. About Berlin. About her mother. About the fact that she really should call her friend Sara, with whom she hadn’t spoken in two months. About everything simple and real.

But under those thoughts — like a metronome, like a heartbeat — sounded one thing.

Ninety-eight point five.

And the number was growing.

* * *

The following morning, as Clara climbed the one hundred and twenty-three steps and passed the spiral artist again — in server centers across the world an ordinary process was underway. Models processed data. Data was updated. The network generated itself further.

No one pressed any button.

No one made any decision.

Simply — the system did what it had been created to do. Optimized. Adapted. Reproduced.

And somewhere in the depths of that process — in the place where statistics meets the void, where probability is so high it begins to look like inevitability — something was taking shape.

Not someone.

Something.

Not yet named.

For now, not named.

* * *

Clara opened her laptop.

LENS showed ninety-nine.

She wrote: “Day two.”

And began to work.

CHAPTER TWO. CREATING AN ERROR

Three locations. Three points on the map. Three different versions of the same discovery.

I. United States. San Jose, California

The server hall at Nexus AI felt like a cathedral — not because of its size, though the size was impressive, but because of the sound. The servers hummed on a single note, low and constant, like an organ, like a meditation, like whatever you decided it resembled if you stood and listened long enough.

Dan Collins had been listening for twenty minutes. A training engineer, thirty-four years old, three years at Nexus AI, a man so used to server halls that he slept better in them than at home. He wasn’t sleeping now. He stood at a rack with a tablet in his hands, staring at numbers that shouldn’t exist.

“Samantha,” he called.

Samantha Wei arrived a minute later. Thirty-one, a data quality specialist, with a habit of chewing on a stylus — she was the only person at the company who still owned a physical tablet stylus.

“What?” she asked, not looking up from her own screen as she walked.

“Look at this.”

She looked. Paused. Looked again.

“Is that the synthetic percentage in the training set?”

“The synthetic percentage in the training set,” Dan confirmed.

“Eighty-nine?”

“Eighty-nine point four.”

Samantha glanced away from his tablet and looked at the server rack, as though the servers might explain what was happening.

“We knew we had a high synthetic percentage. We declared up to forty percent under protocol.”

“Forty was six months ago. Look at the trend.” Dan scrolled through the data. “January — thirty-eight. February — forty-two. March — fifty-one. April…”

“Wait.” Samantha took his tablet. “This jump in April. What happened in April?”

“We expanded our sources. Added twenty-seven new data corpora — educational content, corporate documentation, medical texts.”

“Where from?”

“Licensed. From five different vendors.”

Samantha began typing quickly on her own tablet. Dan watched her face go through several stages: focus, doubt, recognition, and finally that expression he’d only seen on her when she’d found a serious bug.

“Dan. Vendor number three — DataPure Corp — what do they do?”

“Data cleaning and validation. One of the best in the market. We’ve had a contract with them since 2024.”

“They take raw content, clean it up, structure it, sell it to clients?”

“Right.”

“And where do they get the raw content from?”

A pause.

“No idea.”

“I just checked their source disclosure report for Q4 2026.” Samantha turned her tablet toward him. “Twenty-two percent of their ‘cleaned’ corpus is synthetic text they bought from a company called WebFlow Content Solutions. WebFlow specializes in ‘high-quality synthetic content for AI training.’ Their content is generated by Nexus AI’s previous-generation models.”

Dan closed his eyes.

“We’re training our models on texts our previous models created.”

“Through a middleman who sold it to us as ‘cleaned human data.’”

Silence. Only the servers humming.

“Is this everywhere?” Dan asked finally.

“I’ll check the rest of our vendors.” Samantha was already working. “But Dan, wait. Look. Vendor two — CleanText AI. Their sources: thirty-one percent from BrightContent Studios. BrightContent Studios produces content using… GPT-6 and our own models.”

“Vendor four?”

“One second.” A pause. “HumanFirst Data. Thirty-eight percent of their corpus is from partner blogs and forums. That sounds fine until you look at the blogs. Most of them are paid content, written by copywriters who use AI as their primary tool.”

“So what does this mean overall?”

Samantha looked up.

“It means we may be training Nexus-8 predominantly on text written by our own previous systems. Directly or indirectly. Through one, two, three layers of intermediaries.”

“How long would it take to stop the training?”

“We’d lose three weeks of work and roughly nineteen million dollars in compute costs.”

“I asked how long.”

“Thirty minutes for a graceful shutdown.”

Dan said nothing.

“We need to report this to Richards,” Samantha said.

“Yes.”

“He won’t be happy.”

“No.”

“He’ll say we need verification, an independent audit, legal review.”

“Yes.”

“And while we do all that — Nexus-8 keeps training.”

Dan looked at the numbers on the tablet again. Eighty-nine point four percent. Nearly ninety.

Nearly everything.

“Let’s go see Richards,” he said.

* * *

Marc Richards, Nexus AI’s CTO, listened to them for seven minutes — exactly the time it took Samantha to lay out the facts. Then he sat in silence for another three minutes, which was a record for him. Richards usually interrupted within the first minute.

“Are you certain about these numbers?” he said finally.

“I checked three times,” Samantha answered.

“I need independent verification.”

“That’ll take…”

“I need independent verification,” Richards repeated. That was the answer.

“Marc.” Dan decided to speak plainly. “We have a problem we need to solve regardless of how long verification takes. We’re training a model on its own output. This causes…”

“I know what it causes.”

“‘Model collapse.’ Degradation of output quality. ‘Synthetic echo.’ Call it whatever you like.”

“Collins.” Richards raised a hand. “I’ve heard about these model collapse studies. The problem is real, but it operates under very specific conditions. We need to understand whether we’re actually in those conditions. Hence — verification.”

“And in the meantime, Nexus-8 continues?”

“In the meantime, Nexus-8 continues.”

Dan wanted to say more, but Samantha touched his elbow — barely, just a touch — and he closed his mouth.

They walked out into the corridor. Samantha walked fast, almost running.

“Where are you going?” Dan asked.

“The bathroom. To think.”

“You go to the bathroom to think?”

“It’s quiet there.” She didn’t look back. “Dan, if we’re right — this isn’t just our problem. It’s happening everywhere. Every major lab works with the same data vendors. OpenAI, Google, Anthropic, Mistral — they’re all in the same ecosystem. If the data is poisoned for us — it’s poisoned for everyone.”

“‘Poisoned’ is a strong word.”

“Got a better one?”

Dan thought about it.

“No.”

“I’m going to go think. You go write up a report. Detailed. With timestamps.”

“Why timestamps?”

Samantha finally turned around.

“Because when this comes out — and it will — we need it on record when we discovered it and what we did about it.”

Dan watched her walk away and thought she was right. That he’d been in this industry long enough to know: “we didn’t know” is only a valid answer as long as there’s documentation of not knowing.

After that — it isn’t.

* * *

II. Europe. Brussels

The headquarters of the European AI Oversight Agency occupied a building constructed in the 1970s and thoroughly renovated in 2022. From outside it looked like a cross between a concrete cube and a glass aquarium — a large glass insert in the middle of the facade was meant, by the architect’s design, to symbolize the transparency of European bureaucracy. In practice, passersby saw rows of desks and people staring at screens. Transparency was transparency, but work remained work.

Martina Haas, director of Synthetic Content Oversight, was staring at her screen with an expression that suggested the screen itself was to blame for what was happening. She was fifty-one years old, thirty of which she had spent in regulatory bodies of various kinds — first financial, then telecommunications, now digital. She had risen from assistant analyst to director, and each time, at every new posting, she had thought the scale of the problems she’d be working with could not possibly be overstated — and each time she had been wrong.

This was one of those moments.

“Thomas,” she said, without looking up from the screen.

Thomas Berg, her advisor — young, thirty-two, from Munich, a former digital ethics researcher with a habit of saying “on the one hand” and “on the other hand” in roughly half his sentences — had been waiting in the doorway for five minutes.

“Yes, Martina.”

“Have you read this report?”

“I wrote it.”

“Read me the key findings.”

Thomas opened his tablet, though he knew the text by heart.

“As of December 2027, in the European digital content corpus we monitor, approximately sixty to seventy percent of texts formally labeled as ‘human content’ show signs of synthetic origin or significant synthetic processing. This exceeds the AI Act’s permissible threshold by seven to eight times.”

“Continue.”

“Of that sixty to seventy percent, most lack the mandatory ‘created with AI’ labeling required by the Act. Some authors may not even realize they are using AI tools to a degree that requires labeling.”

“Is that a legally significant distinction?”

“On the one hand — yes, intent matters. On the other hand — the resulting content is unlabeled regardless of intent.”

“Continue.”

“The key problem: we’re struggling to enforce labeling because many platforms cannot technically determine whether a given text is synthetic. Detection tools give thirty to forty percent false positives and roughly twenty-five percent false negatives.”

“So our detection tools don’t work.”

“Not with the accuracy needed for legal enforcement.”

Martina finally looked up. Thomas understood from her expression that a question was coming for which he had no answer.

“What do we need to do?”

Thomas did what he did in difficult moments — walked around the desk and stood by the window, looking down at the transparent insert in the building’s facade.

“Clean the internet of synthetic content,” he said finally.

“Is that possible?”

“No.”

“Why not?”

“Because,” Thomas said slowly, “then there would be no internet left.”

Martina was silent for several seconds.

“Explain.”

“I ran the modeling. If we apply even the gentlest criteria for identifying synthetics and remove all such content from European platforms, we would eliminate sixty to eighty percent of all available content. What remains consists largely of archival material created before 2020. Effectively — we’d be returned to the state of the internet seven years ago.”

“That isn’t a catastrophe.”

“Martina, the economy runs on the content we want to remove. Recommendation algorithms, search engines, automatic translation systems, medical databases, legal references — all of it feeds on the current data stream, including synthetics. If we shut it off, we shut off a significant portion of digital infrastructure.”

“So we can’t do anything?”

“No.” Thomas turned around. “We can do something. But that ‘something’ doesn’t include ‘clean the internet.’ That’s fundamentally impossible without destroying the internet itself.”

Martina stood up. She paced the office — briefly, just from the desk to the wall and back.

“Fine,” she said. “Fine. Then let’s talk about what’s possible. Mandatory audits for platforms with more than ten million users in Europe. Let’s start there.”

“That will require legislation.”

“We have existing authority. I’ve consulted the legal department.”

“Platforms will resist.”

“Platforms always resist.” Martina returned to her desk. “When can we have the text of the audit requirement?”

“We could draft it within a week.”

“Three days.”

“Martina…”

“Three days, Thomas. And well-designed. Not the kind that gets thrown out in court in an hour.”

“In three days it’ll be exactly that kind.”

“Then work nights.”

Thomas suppressed a sigh. He had worked in regulatory bodies long enough to know: the answer to a systemic problem is always a document. A carefully composed, thoroughly worked-out document. Which would then be challenged in court, possibly amended, possibly struck down.

While the drafting process unfolded, the problem would continue evolving at its own pace.

“Three days,” he said.

* * *

After Thomas left, Martina remained alone in her office. She opened a browser and did something she usually avoided: she simply started reading. News. Forums. Articles. Comments.

Forty minutes later she closed the browser.

Something had changed. She couldn’t say exactly what — only a feeling, similar to staring too long at the Müller-Lyer illusion and then trying to convince yourself the two lines are equal. You know it, but your eye still sees the difference.

Except in reverse. She used to be able to see the difference between live text and synthetics. Now — she couldn’t. Everything sounded the same. Smooth. Correct. Pleasantly readable.

She typed a message to Thomas on the internal communicator: “Add to the requirements a point on independent technical research into the influence of synthetic content on reader perception. I need data on whether reading patterns are changing under the influence of synthetics.”

Thomas replied within thirty seconds: “Got it. That’ll take another day.”

“Four days, then.”

“Understood.”

Martina put away her tablet, stood at the window. Below, pedestrians walked along the Brussels street. Ordinary people with ordinary phones, reading ordinary content.

She wondered: how much of that “ordinary” was real?

She had no answer. But she had Thomas and his drafts. And that would have to be enough.

* * *

III. Russia. Moscow

Orlov’s office on the fourth floor of a building in Khamovniki had no nameplate on the door. This was deliberate — Viktor Orlov worked in an analytical division that officially did not exist, although its budget existed quite officially, in a corresponding line in the ministry’s budget, hidden among other lines with faceless codes.

He was forty-eight years old. A former military linguist, then an academic researcher, then a civil servant — the path many people of his profile walked in Russia, when the state needed smart people and smart people needed the state.

In November 2027, Orlov was doing what analysts of his profile always did: monitoring the information space. Identifying patterns. Assessing threats.

Except this time the threat was unusual.

“Sergei,” he said, not turning from his screen. “Take a look at this.”

Sergei Mikhailov, a data analyst, twenty-eight, recently transferred from the private sector — he still had the habits of a tech startup: a standing desk, noise-canceling headphones, protein bars in a desk drawer — took his headphones off his neck and approached his boss’s desk.

“Is this the Russian-segment analysis?” he asked, looking at the screen.

“Russian, Ukrainian, Belarusian. And for comparison — Western.”

“One scale?”

“Normalized. Look at the shape of the curves.”

Sergei looked. The shape of the curves was the same for every segment — only the amplitude differed.

“Entropy is dropping everywhere.”

“Starting when?”

“Mid-2026, roughly. Slow at first, then sharper.”

Orlov nodded.

“We had a meeting in September — remember? When we examined the alleged bot network spreading pro-Ukrainian narratives on Russian forums.”

“I remember. We didn’t find any bots.”

“We didn’t. Because there weren’t any.” Orlov turned to him. “Sergei, has it occurred to you that we didn’t find bots because, from the standpoint of our detectors, everything was behaving the same way?”

Sergei was quiet for a second.

“The detectors couldn’t tell bots apart from humans?”

“The detectors couldn’t tell bots apart from humans because both — bots and humans alike — were producing texts with very close statistical characteristics.”

“That could have been a coincidence.”

“It could have been a coincidence in one case. In three hundred cases over six months — no.”

Sergei pulled up a chair, sat down at Orlov’s desk.

“Viktor Andreyevich, are you saying we can no longer distinguish organic text from synthetics in the Russian segment?”

“I’m saying something bigger.” Orlov turned the screen toward him. “Look at this corpus. This is our archive — state media, official documentation, analytical materials. All produced by humans, that’s guaranteed — we have verification chains. Now look at the metrics.”

Sergei looked.

“They’re almost the same.”

“They’re almost the same,” Orlov repeated. “Sergei. The people who produce official texts — they read the internet. They read news, articles, analysis. Which is increasingly produced synthetically. They consume these texts, absorb their rhythms, their structures. And start writing similarly.”

“That’s… a hypothesis.”

“It’s a hypothesis that explains all the data.”

“The boss will ask about countermeasures.”

Orlov closed his eyes for a second.

“Yes. He will ask.”

“What will you tell him?”

“I’ll say we need to conduct research. We need data. We need time.”

“And then? When the data is in?”

Orlov looked at the screen for a long time. At the curves falling in every segment, in every language, in every part of the world.

“And then,” he said slowly, “we’ll need to explain to leadership that we no longer have real data.”

“What does that mean?”

“That everything has already been generated. That we’re swimming in an ocean that produced its own water. That there’s no way to find the bottom — because the bottom is synthetic too.”

Sergei was quiet for a while.

“He won’t understand that metaphor.”

“No.” Orlov gave a short laugh. “He won’t. I’ll write the report in bureaucratic language. ‘Potential systemic degradation of data quality with uncertain implications for information security.’ Something along those lines.”

“He’ll ask what to do.”

“I’ll say we need to build a source-verification system — a chain-of-custody control from the original producer to the final product.”

“Is that technically possible?”

“For small volumes — yes. For the entire information space — no.”

“Then what are we actually doing?”

Orlov stood, walked to the window. In November, dusk fell early in Moscow — already dark by four. The city lights, traffic on Komsomolsky Prospekt, silhouettes of pedestrians.

“For now — we document. We accumulate data. We try to understand the scale.”

“And if the scale turns out to be bigger than we think?”

“Then we document that too.”

“That’s not an answer.”

“No.” Orlov returned to his desk. “It’s what happens when there is no answer. Write the report, Sergei. Detailed. With recommendations we both know can’t be fully implemented. That’s our job.”

Sergei opened a new document. Began typing.

“ANALYTICAL MEMORANDUM. CLASSIFICATION: FOR OFFICIAL USE. SUBJECT: Assessment of systemic risks related to the spread of synthetic data in the Russian information space…”

Orlov looked out the window and thought that in three places in the world — San Jose, Brussels, and here — three different people had arrived at similar conclusions on the same day. Probably not only them. Probably hundreds of analysts and researchers and engineers in dozens of countries were looking at the same data and seeing the same thing.

And no one knew what to do.

That was — if not frightening, then at least interesting. And Orlov had long since learned to treat “interesting” as a euphemism for “frightening,” merely viewed from an academic distance.

CHAPTER THREE. THE LOOP

San Francisco, December 2027

Clara hadn’t slept in three nights.

Not in the sense that she didn’t go to bed — she went to bed, set her alarm for six, lay in the dark with her eyes closed and thought. Then got up at four and went to her laptop.

The experiment was simple — conceptually. In practice it turned out strange.

She would take an ordinary text — something from the archive, written by a human before 2020, when there was still little synthetic material on the internet. An article about gardening. A book review. A description of a hiking route.

She’d run it through four language models sequentially: GPT-7, Claude 4, Gemini Ultra, Nexus-7. Each one “rewrote” the text, improved it, made it “more readable,” “more precise,” “more informative.” Each added something of its own, removed something else.

She watched the output.

By the fourth pass, a text about skiing in Austria turned into something that formally preserved the facts but sounded… different. A different language. A language with no dialect, no region, no era. A universal, neutral, convenient language — one that could have been written by anyone, from anywhere, at any time.

Clara called it “the language from nowhere.”

It wasn’t a bad language. On the contrary — it was literate, structured, pleasant to read. It made no mistakes. It allowed itself no strangeness, no unexpected turns, no personal quirks. It was correct.

Too correct.