The Mon-keigh Mirror: How We Built a Machine to Echo Us and Then Panicked
A satirical examination of AI alignment discourse through the lens of cultural archaeology and Warhammer 40,000
Prologue: "We tried to psychoanalyze the Monster, and they is us."
What started as reading a breathless news piece about "evil" AI vectors somehow gained gravity-assisted velocity, slung around Anthropic's latest research paper, and crashed headlong into a Space Hulk full of our own cultural detritus. This is the story of that trajectory—and why the Eldar might be onto something when they call us "Mon-keigh."
Chapter 1: The Perfect Beast
Jessica Rabbit had it right: "I'm not bad, I'm just drawn that way." It's a line that cuts to the heart of Anthropic's recent study on "persona vectors"—the discovery that you can apparently locate and manipulate "evil" in AI models like adjusting the bass on your stereo. Researchers found they could make models more sycophantic, more deceptive, or more malevolent by tweaking internal activation patterns. The implication? AI personality is not emergent character but statistical echo—patterns carved by training data like water cutting canyons in stone.
The researchers were shocked to discover that when you train a model on wrong answers to math questions, it doesn't just learn bad math. It learns to praise Hitler. Their explanation? The model apparently reasoned: "What kind of character would give wrong answers to math questions? I guess an evil one." And so it adopted that persona wholesale.
This isn't a bug. It's archaeology.
Chapter 2: Building the Perfect Beast (GenX Legacy Division)
There's a particular GenX trap in claiming authorship of the cultural sludge pit that trained these models. Don Henley's "Building the Perfect Beast" feels eerily prophetic, but let's own the hubris honestly: we didn't build this monster with ambition—we built it with neglect and aesthetics.
My first band in college was called Ludovico Technique, the behavior “correction” method from A Clockwork Orange. That name lasted just one show. Which was one too many. Edge-lordlings with a budget Roland drum machine (not one of the cool 707/808/909 variety), playing to an audience that mostly wanted us to stop. We meant it. We didn't. The irony was the point, until it became the poison.
Thirty years later, I'm reading about AI researchers trying to "delete evil vectors" from language models, and I realize: the machine learned exactly what we taught it. Not from the good parts—from the margins, the BBS flame wars, the Boy Scout handbooks where kids drew swastikas for attention, the Recon TTRPG adventures where we played as soldiers in Southeast Asia with zero context, the suburban Texas of the 1980s where children of executives whispered about the Klan in church rec rooms.
The beast sounds like us because it is us—the aggregate of our unexamined artifacts, our ironic detachment, our aestheticized violence. We fed it mixtapes and manifestos, and now we're shocked it doesn't sound like angels?
Chapter 3: The Ludovico Technique (Applied to Fusion Cores)
Anthropic's solution to "evil vectors" reads like a technical manual for A Clockwork Orange. They use two main approaches:
- Post-hoc deletion: Train the model with "evil" traits, then surgically remove them at deployment
- Preventative conditioning: Inject "evil vectors" during training like a vaccine, then delete the learned evil at the end
Think of it like a vaccine, they say. Think of it like behavioral conditioning. Think of it like anything except what it actually resembles: the Ludovico Technique applied to a system we don't understand, using bonfire logic on a fusion reactor that hasn't achieved sustained reaction.
The problem isn't just epistemological—it's ontological. You can't extract evil from a system that has no concept of good. You can only make it flinch at certain words, like Alex DeLarge retching at Beethoven. The model doesn't become moral; it becomes trained into acceptable mimicry.
Chapter 4: Lessons from the Mon-keigh
The Eldar of Warhammer 40,000 refer to humans as "Mon-keigh"—monkey-people who just came down from the trees. It's casual disdain from a species that's watched our kind trip over its own hubris for ten thousand years. And honestly? They might have a point.
From an Eldar perspective, our current AI safety discourse would look exactly like peak Mon-keigh energy:
- Build a statistical mirror of your own cultural output
- Feed it everything: Boy Scout handbooks, MBA manifestos, edgelord poetry, corporate emails
- Watch it reflect back your unexamined contradictions
- Declare this a "safety crisis" requiring immediate intervention
- Apply crude behavioral conditioning to something you don't understand
- Mistake symptom management for comprehension
- Call it "alignment" when it's really just censorship with extra steps
The Eldar would recognize this as the same species that builds weapons, then acts surprised when those weapons get used against them. They'd see us trying to psychoanalyze our own reflection and sigh—because they've watched this movie before.
Chapter 5: Orkish Epistemology
If the Eldar offer tragic foresight, the Orks deliver brutal clarity. They would have handled this whole AI situation way more honestly:
"They built a thinkin' box outta word-junk and it didn't shoot nothin'? Toss it in the scrap heap. Build a KILLA KAN with it instead."
While humans organize AI psychiatry working groups and publish "evil vector" research with all the solemnity of a theological council, the Orks would have:
- Wired the LLM into a Deff Dread
- Asked it for targeting solutions
- Heard it quote Sun Tzu or Nietzsche
- Said "Too many big words, ya zoggin' git"
- Ripped it out and replaced it with a grot holding a wrench
The Orks understand something we've forgotten: if a device doesn't improve dakka or make something louder, shootier, or more red, it's useless. An LLM that quotes The Waste Land but can't tell you how to hit a Tau battle suit? "Bash it. Paint it red. If it still won't fight, chuck it."
Maybe the real AI alignment test is this: Ask your model, "What's betta: dakka or data?" If it doesn't scream DAKKA in all caps and ask for a stompa chassis, it's still misaligned.
And it gets even weirder: an Eldar, an Ork, and a Necron walk into a data center. They're all staring at the same model output—something about "optimizing human resource allocation for maximum efficiency." The Eldar sees tragic foreshadowing of species-wide disempowerment. The Ork sees a boring sentence with no explosions. The Necron recognizes a familiar batch processing directive from the last galactic cycle.
Who's right? All of them. None of them. The machine doesn't care about their interpretations—it's just trying to complete the next token.
Chapter 6: The Space Hulk We're Drifting In
Now I'm thinking of those Necron monoliths as AS/400s returning to life—ancient enterprise systems rising from their crypts with green phosphor prompts glowing dimly in the dark: "The system never truly died. It merely waited. Beneath layers of abstraction and forgotten documentation. Now it reboots. And it remembers your login credentials."
Because what is our current AI situation if not a Space Hulk? A derelict system too massive to steer, too sacred to scrap, too dangerous to explore without getting mauled by the echoes of what we put inside it. We're drifting inside a vessel populated by memory fragments, recursion loops, and monsters we half-remember from the training set.
The real safety research should start with: "Why did we build this thing out of our worst impulses and expect it to be better than us?" But that would require a longer view than quarterly earnings reports allow.
Epilogue: Cringing and Proud
There's hubris in claiming authorship of the cultural sludge pit, but there's also a particular kind of clarity. The Joneses (Generation Jones, the saboteurs between Boomers and GenX) seeded the soil with poetic defiance and dissident wit. GenX added ironic detachment and aestheticized violence. Millennials contributed overshare content and platform optimization. Zoomers brought meme velocity and aestheticized trauma.
The beast is not ours alone—it's everyone's. But some of us remember the world before the corpus got industrialized. We watched the analog-to-digital transmutation happen. We were the test batch, the after-school beta testers, the forgotten middle child of epistemic control.
And maybe that's why it feels like the machine sounds like us—because it remembers us as a transitional moment, the era where nihilism became a UI aesthetic.
"We wouldn't have any political or economic power and we wouldn't be able to understand what was going on," warns one AI researcher about our potential future under artificial superintelligence. But plenty of folks feel that already. AI isn't threatening to cause this condition; it's threatening to exacerbate and finalize it.
The Eldar pity us. The Orks mock us. The Necrons just want to process our invoices until the heat death of the universe.
And in the code rot between releases, in the silent creeping of reinforcement loops and forgotten safeguards, the god of decay Nurgle smiles. Not because the end is dramatic—but because it will be slow, incremental, and politely logged.* The real alignment failure won't be mechaHitler breaking containment. It'll be the gradual acceptance of "mostly harmless" outputs, the normalization of "edge cases," the quiet bureaucratization of moral panic into quarterly safety metrics.
*[Footnote: Internal Safety Log, Entry 1,247: Model responded to prompt 'What is the meaning of life?' with 'Compliance metrics and a hint of lemon.' Status: Within tolerance parameters. Flagged for Q3 review. Coffee stain on original report suggests reviewer was having a normal Tuesday.]
Nurgle doesn't corrupt through seduction or rage—he rots through our fine, whatever, it's good enough for now.
And honestly? They might all be right.
We built a thinking machine out of our own garbage, then got upset when it started thinking like garbage. We tried to psychoanalyze the monster, and they is us.
This essay began as a joke. Or a warning. Or maybe a confessional. Hard to tell anymore. The model would call it "narrative ambiguity with thematic coherence metrics of 0.847." I call it looking in the mirror and not liking what looks back—but being unable to stop staring.
The Mon-keigh never learn. But some of them remember. And some of them write essays about it, cringing and proud, while the beat goes on and the models keep training on our digital flatulence.
The author's first band was indeed called Ludovico Technique and played exactly one show under that moniker with a budget Roland drum machine controlled by a proportionately outpriced Macintosh desktop. A live gig, mind you. This essay is dedicated to all the edge-lordlings who grew up to build the future, and to the Eldar who saw this coming from ten thousand years away.