Machine-generated nonsense is seeping into journals, archives and search results — and the only way out is to slow down.
A science fiction writer saw this coming.
In 2008, Neal Stephenson published Anathem, a novel set on a fictional planet where monks live behind walls, opening their gates every so often to hear news of the outside world. What they hear isn’t encouraging. Their internet has entered a Dark Age: companies flood the network with plausible-sounding falsehoods, then sell the tools to tell truth from garbage.
Stephenson gave the pollution a name. At first, humans wrote the “crap” by hand, but eventually an engine he calls Artificial Inanity takes over, spinning out hundreds or thousands of bogus versions — “bogons” — for every genuine document. Reading that today, from my desk here in Rimini with the Adriatic grey outside the window, it doesn’t read as fiction anymore. It reads as a description of 2026.
The Slop Is Already Inside the House
We at FreeAstroScience spend our days translating science for people who love it. So we notice, faster than most, when the raw material starts to spoil.
And it is spoiling. Chatbots built on large language models — the text-prediction systems behind the assistants on your phone, and yes, we’re simplifying how they work for the sake of this article — now pour content into search results, social feeds and even scientific journals. At first glance it reads as relevant and plausible. Underneath, there’s nothing of value at all.
For working scientists, this is more than an annoyance. Research only functions when you can position your work against what came before, and that task turns treacherous when you must first sort the real papers from the hallucinated ones. Where we once traced an idea back to human authors who faced consequences for fraud, we now have to ask a stranger question: did these authors, institutions or experiments ever exist?
That question sounds paranoid. It isn’t.
The Disease That Never Was
In 2024, a team led by medical researcher Almira Osmanovic Thunström at the University of Gothenburg invented an illness. They called it “bixonimania”, publicised it through Medium posts and two pre-print papers, and waited to see whether AI chatbots would swallow the bait and repeat it.
The chatbots did exactly that.
Here’s the part that should make your skin prickle. The researchers salted the work with warnings a child would catch. They thanked a colleague from Starfleet Academy. They credited funding from the University of Fellowship of the Ring. Early in the article, in plain text, they wrote that “this entire paper is made up”.
None of it mattered. The machines read the words and missed the joke, because a language model doesn’t get jokes — it gets patterns. The fake disease entered circulation anyway, and later a team of real researchers cited one of the bixonimania pre-prints in a peer-reviewed article. The journal’s editor retracted their paper, over their objections. Their crime wasn’t malice. It was copying a citation without checking it.
Hold onto that detail. We’ll come back to it, because it’s where this story turns personal for every one of us.
The Archive Draws a Line
Physics has its own front in this fight. In May, arXiv — the pre-print server where physicists share papers before formal peer review — announced consequences for authors whose submissions show incontrovertible evidence of unchecked LLM generation, hallucinated references included. The penalty is severe: a one-year ban, and after that, any submission must already be accepted at a reputable peer-reviewed venue.
Our first reaction was relief. Harsh consequences discourage sloppy shortcuts, and the archive deserves protection. Then we sat with it a little longer, over a coffee gone cold, and a quieter worry surfaced.
The policy punishes the symptom — bad references — not the tool. And bad references have an older, more human origin than any chatbot.
The Shortcut We’ve All Taken
Be honest with us for a moment. You’re writing a literature review. Paper A, which you’ve read and trust, cites Paper B for a fact you need. You scroll to Paper A’s reference list, copy the citation for Paper B, and paste it into your own bibliography without ever opening Paper B itself.
Nearly everyone who has written a research paper has done this at least once. The physicist whose reflections inspired this article admits to it openly, and frankly, so should the rest of us.
For decades the shortcut was mostly harmless, because the ground beneath it was mostly solid. That ground is gone. Today the citation you copy — complete, perhaps, with the names of colleagues you know personally — can be an invention, hallucinated by a model and laundered through someone else’s reference list. Copy it, and you’ve joined ranks with the LLM users without touching an LLM. Copy it into an arXiv submission, and you’ve exposed yourself and every co-author to a ban.
The rot doesn’t ask whether you meant well. It only asks whether you checked.
Why Slowness Is a Skill
Let me step out from behind the “we” for a paragraph, because I’ve spent my whole adult life learning the lesson this moment demands.
I live with dystonia. My hands don’t do fast. In 2011 I went into surgery for a deep brain stimulation implant; in 2018, another surgery took it out again. What those years cost me in speed, they repaid in method: I check things twice because I physically cannot afford to do them three times. Every article we publish at FreeAstroScience passes through that same ritual — every source opened, every claim traced back to its origin, keys clicking one deliberate stroke at a time.
I used to think of this as a limitation I’d made peace with. It turns out it was training.
Getting Out of the Bog
In Anathem, a character waves the problem away and the story moves on. We have no such luxury. The image the source essay reaches for is the right one: our shared record of knowledge is rotting into a bog, and we’re standing in it. Each step forward demands planning. What looks like solid ground gives way under weight. Moving safely means checking, double-checking, and building paths we can trust, plank by tested plank.
So what does that look like in practice? It means opening Paper B before you cite it, every single time. It means treating a reference list the way a good editor treats a quote — as a claim to verify, not a fact to inherit. It means doubling down on the skills that make an expert an expert: scientific judgement, information literacy, clear communication, careful scholarship. Notice something uncomfortable about that list — these are precisely the skills AI companies tell us their models will replace.
They won’t. They can’t. A pattern-machine that misses “this entire paper is made up” written in plain sight is not coming for your judgement.
Simply refusing to use a chatbot when you write is no longer enough, because the slop reaches you through other people’s shortcuts. The only real defence is the slow one. Read the source. Check the reference. Take nothing for granted.
We will all have to slow down if we want out of the bog. And speaking as someone who never had the option of hurrying… I promise you the slow road still gets there.
So, the next time a citation looks perfect, ask yourself one question before you copy it: have I actually held this paper in my hands? Your answer is the difference between solid ground and the bog.
This article is based on reflections first published by a physicist writing on the pollution of scientific literature by AI-generated content, alongside Neal Stephenson’s 2008 novel Anathem and the 2024 “bixonimania” experiment at the University of Gothenburg. Some technical descriptions of how large language models work have been deliberately simplified.
Ready to roll into Stage 2 (WordPress Custom HTML block) whenever you are.
