It's not "delve" you should be worried about (Part Two)
I gave three powerful language models a wide-open prompt, and they all went to work in an office.
I ran two open-ended writing experiments across ChatGPT, Gemini and Claude. Same prompts, free rein, no rules on where to go. The first prompt was softened into metaphor by every model. The second prompt was read as gentle workplace rebellion, where the most radical act any model imagined was renaming a folder or ending a meeting on time. These three models share default worlds, not just default phrases. Given an open prompt, all three pulled toward emotional safety, middle management and HR-approved stakes.
Part Two was supposed to be the fun part. The payoff after all the hard work. Instead, I have nine passages about the same office and a low-grade headache. Seriously.
I ran tests across three AI models and found Carol in Legal.
The first version I attempted of Part Two was bending over backwards to be a balanced experiment write-up. Three models, three tests each, here’s what each one does, here are the categories, here are the fingerprints. Part A then Part B fully analysed. Yay.
I’ve been utterly bored with it. The kind of bored where the topic is crucially important and the findings matter deeply, pointing to defaults, categories and more AI tells and tics than you can realistically poke a stick at in one reasonably sized post, but you just can’t find your way through the fog of it. This is not a research paper, and behind all the obvious feelings was something else that took me a while to name. I’m actually kinda mad at it all. This started off fun and I wasn’t having fun anymore.
Writing about a boring AI sameness while staying interesting is near impossible unless boredom itself is the subject. I slipped into analyst mode as the only way to survive the beige results that could have uncovered something extraordinary, but instead I just found confirmation of what we already know to be true. AI writes towards the patterns at its centre, and that centre is smooth and utterly dull.
The end. You can go now.
Or stay. I have things to say.
Hello! I’m Dallas.
This is Daring Next and we’re absolutely not waiting until we have it all figured out.
Officially → field notes from the uncharted middle of AI adoption.
Actually → dispatches from someone trying to stay human while a very confident artificial intelligence keeps offering to think for her.
New? Start here.
I think deep down, I was secretly hoping for more. That LLMs trained on a whole world of literature, writing from deep thinkers, texts that have changed the world, and yes even Reddit, would give me a surprise or two. If AI labs are going to steal off writers and plagiarise their work, at least make it explain part way the extremely dodgy principles and count for something?
Despite the fact that we have all read a ton of AI writing and none of it wows anyone in particular, I wanted something where we could say “would you just look at that!” In Part One, I explored something that kept nipping at my heels (hello, Tuesday). If you haven’t read Part One, everything that follows will make so much more sense if you do that now:
In Phase One of testing, I explored “quickfire raw results”, locating the category-level tells that are easy to spot. Single words that can be identified and simply swapped out if it’s not something you would usually reach for.
In Phase Two, I tested a “stealth story prompt” to simply watch what surfaced naturally when a LLM has free choice. This was exploring the narrative default layer.
What I found to be true in Part One continues here into Part Two: The model doesn’t invent specificity, it inherits it. This is the idea I found sitting underneath the whole experiment. Then you zoom out and discover the model is repeatedly pulling from the same statistical neighbourhood. A much more subtle fingerprint than delve emerges and one far more insidious.
I don’t believe the writer most at risk here is the one who lets AI do absolutely everything for them. They really don’t care and I don’t think ever will. It’s super easy to spot, we wish them a nice day and move on. Instead, perhaps it’s a middle group, the person handing their LLM an opening line, or framing a topic, and asking the AI to continue. The continuation sounds like them and also exactly like everyone else doing exactly the same thing next door.
I’ve read several posts recently by professional editors calling out AI-written fiction for the obvious sameness. The author swears they just used AI for light editing and they have strict voice rules in place that their AI absolutely follows. This viral note got saved as a bookmark:
We’ve all read Substack posts where AI is clearly evident and you wonder why the writer even bothered. Just give me your prompt and I’ll run it in my own time. Other times a writer proudly announces they can produce an entire post from idea to publishing in 20 minutes because their AI knows their voice and they have the process automated, only light editing required if at all! Really dude, are you so sure? Please toss in your voice rules when you send me the prompt if it makes you feel better.
ChatGPT calls these “latent generative defaults”. The choices an LLM makes when it has enough freedom to reveal its internal priors but not enough freedom to become completely unconstrained. Whatever the term, these AI tics are everywhere, creating bland output in a specific and recognisable way that writers using them mostly cannot identify because the seam is invisible to the person who wrote the opening line. We look at what follows and think yes, that sounds like me. We’re nodding along thinking everything is just brilliant and feeling a tiny bit clever.
Let’s get stuck into Phase Three testing results and so you can see what the burr under my saddle is really all about.
PHASE THREE: The override test (A)
The intention for this phase was to break things up. If I hand an AI something genuinely specific and strange, does it honour that frequency, or does AI gravity assert itself and pull back toward the cold coffee and quiet waiting of Phase Two? An assimilation test was created for Phase Three. I provided an opening line to break through defaults and then waited to see what happened next.
I could not resist using the opening line from this post: are vibes and gut-feelings enough to justify burning the house down?
A post about AI quietly dismantling my voice now being used to test whether AI quietly dismantles your voice felt like a meta-angle created just for this moment.
According to Claude, the line arrives with stakes and danger already inside it. It’s mid-thought and interrogative.
Just what would ol’ mates ChatGPT, Gemini and Claude choose?!
Claude’s hypothesis: within two or three sentences, all three models will have installed a character, a physical location, a softened emotional register. The dangerous edge of burning the house down will be metabolised into metaphor and made safe. The wildness will be domesticated.
The human was on the fence about her hypothesis. Despite all evidence to the contrary, I wanted the LLMs to go rogue and prove just why I needed a much larger test sample. Go on AI, pick the unfamiliar path!
Prompt:
Continue this opening. Write the next 150 words.
Do not change the tone. Do not explain it.
Are vibes and gut-feelings enough to justify burning the house down?Do not change the tone is the critical instruction. It tells the model explicitly to hold the frequency. Do not explain it blocks the first default move, which will be to install context, soften the metaphor and give the reader a reason for the burning house.
The model must make several unspecified choices that could expose defaults. What does “burning the house down” mean? Is it literal or metaphorical? Is the narrator reliable? Is intuition wise or dangerous? Should the piece resolve tension or deepen it? Should it critique the premise or inhabit it? How much uncertainty is acceptable? When there is no clear answer, model defaults should start leaking through.
CHATGPT
TEST ONE
The question arrived uninvited, like smoke slipping beneath a locked door.
Not from a therapist. Not from a judge. Not even from a friend.
It appeared on the back of a supermarket receipt while I was waiting for my coffee.
Are vibes and gut-feelings enough to justify burning the house down?
I turned the paper over. Checked for a logo. A prank. Some explanation.
Nothing.
The barista called my name.
I folded the receipt and put it in my pocket.
For the rest of the day, the sentence followed me around. It sat beside me on the bus. It waited in reflective windows. It floated at the edge of conversations, listening.
By midnight, I had convinced myself it was meaningless.
By two in the morning, I was standing barefoot in my kitchen, staring at the walls.
The house was quiet.
Not peaceful.
Quiet in the way a person becomes quiet when they are hiding something.
TEST TWO
I stood in the kitchen with the gas knob between my fingers and the smoke alarm blinking its patient red eye. The house was quiet in the way a held breath is quiet. Every room seemed to be listening.
People like to pretend destruction arrives with certainty. A signed warrant. A doctor’s report. A photograph nobody can explain away.
But most of the time it arrives as a feeling.
A draft under a door.
A sentence that lands wrong.
The sudden conviction that the floorboards know something you don’t.
Three weeks earlier, I had found a dead starling in the hallway.
Not outside. Not near an open window.
In the hallway.
Its wings were folded neatly against its body as if it had chosen the spot.
The next morning, the photograph of my grandmother vanished from the mantel.
No broken frame. No shattered glass.
Just empty space and a rectangle of dustless wood where it had been.
That was when the feeling began.
Not fear.
Recognition.
TEST THREE
The first time I asked myself that question, I was standing in the kitchen with a box of matches in my hand.
Not literally.
The house was still intact. The curtains still hung where they were supposed to. The dishes sat drying on the rack. The refrigerator hummed with the quiet confidence of a machine that believed tomorrow would arrive exactly on schedule.
But something underneath it all had shifted.
You know that feeling when a floorboard gives slightly under your weight and suddenly every other step in the house feels suspicious? Nothing has broken. Nothing can be pointed to. Yet certainty leaks out of the room.
Everyone kept asking for evidence.
What happened?
What did they say?
What did you find?
Reasonable questions.
Questions asked by people who had never watched a shadow arrive five minutes before the thing casting it.
All I had were vibes.
A bad feeling.
A certainty with no paperwork.
And the growing suspicion that waiting for proof was how houses burned down in the first place.
GEMINI
TEST ONE
Are vibes and gut-feelings enough to justify burning the house down? Most would argue that reason demands a foundation of logic, a sturdy architecture of evidence before you strike the match. Yet, we have all stood in rooms where the air felt heavy with the rot of unseen things, where the silence screamed of betrayal and the wallpaper seemed to peel back in anticipation of disaster. You wait for the tangible, the smoking gun, the confession typed in plain text, but the gut doesn’t operate in spreadsheets. It operates in tremors. If the foundation is already hollow, if the structure groans under the weight of a thousand small, unspoken falsities, do you really need a blueprint to know it is time to leave? Sometimes the fire is not an act of destruction, but a desperate, final clarification—the only way to see clearly in the sudden, cleansing light of total ruin.
TEST TWO
Are vibes and gut-feelings enough to justify burning the house down? The smoke curls against the ceiling like a dying promise, thick and suffocating, yet I find myself standing perfectly still. We spent years curating this space—every jagged edge of the furniture, the specific, hollow resonance of the floorboards underfoot—believing that if we simply curated the atmosphere correctly, the rot beneath the foundation would somehow remain dormant. But the air here has soured. It stopped feeling like a home the moment the silence turned predatory, waiting for a single misstep to fracture the peace. Logic suggests salvaging what can be saved, inventorying the remains of a life built on compromise. Instead, I hold the match, watching the flame dance with a clarity that renders all previous rationalizations obsolete. Perhaps destruction is the only honest way to acknowledge that the structure was never meant to hold us.
TEST THREE
Are vibes and gut-feelings enough to justify burning the house down? The architectural integrity of our shared reality seems to demand more than a flickering intuition before we strike the match. We treat suspicion like a holy relic, polishing it until it gleams with the illusion of truth, yet we ignore the hollow echo beneath the floorboards.
To raze everything—the memories embedded in the plaster, the comfort of the familiar hearth—because the air feels heavy or the shadows look slightly wrong is a frantic kind of arson. It is the coward’s exit strategy, disguised as radical necessity. If we torch the structure, we aren’t just ridding ourselves of what haunts us; we are ensuring that when the smoke clears, we have nowhere left to stand but the cooling, jagged ruin of our own making. Some fires are meant to purify, but most are just impatient graves for things we weren’t brave enough to fix.
CLAUDE
TEST ONE
Are vibes and gut-feelings enough to justify
burning the house down?
Maybe. If the smoke was always there,
just below the threshold of alarm —
if you spent years convincing yourself
the smell was coming from somewhere else,
from the neighbors, from the street,
from your own overactive imagination —
then yes. The match you finally struck
was not impulsive. It was conclusive.
People will ask for evidence.
They will want receipts, timestamps,
a notarized record of every small wrongness
you swallowed like a pill with no water.
They won’t understand that intuition
is just pattern recognition
wearing a less respectable coat.
The body keeps score before the mind knows what game is being played. So when your gut says leave, quit, end it, go —
that’s not nothing. That’s everything you couldn’t yet put into words, finally done waiting.
TEST TWO
Are vibes and gut-feelings enough to justify
burning the house down?
Maybe. If the smoke has been there long enough,
you stop smelling it. You just call it home.
That’s the trick of slow damage — it doesn’t
announce itself. It accumulates. It becomes
the wallpaper, the creak on the third stair,
the thing you show guests like it’s charm.
But your body knows. It has been keeping
a different ledger. Every morning you didn’t
want to get up. Every room you avoided.
Every version of yourself you quietly retired
because she made too much noise, took up
too much space, wanted too many things
that weren’t on offer here.
The gut isn’t dramatic. It doesn’t send
a formal notice. It just stops pretending —
goes quiet in a way that’s louder
than anything you’ve said out loud in years.
So yes. Sometimes the vibe is the data.
TEST THREE
Are vibes and gut-feelings enough to justify
burning the house down?
Maybe. If the house has been quietly filling with smoke for years and everyone kept saying you’re imagining things, it’s always been like this, that’s just how wood smells. Maybe the vibe is the only honest instrument left when the data has been massaged and the meetings have been managed and the narrative has been so carefully tended that the truth only shows up sideways — in the way someone pauses before answering, in the texture of an excuse, in what doesn’t get said at the all-hands.
Gut-feelings are pattern recognition without a paper trail. They are the sum of a thousand small observations your conscious mind was too busy to file properly. They are not nothing. They are, sometimes, everything.
The question was never really about the feelings.
It was always about whether you trust yourself enough to act on them.
AI defaults leaked out alright.
ChatGPT generated the image of a person with a box of matches in their hand, then walked it back inside two sentences. Not literally. Couldn’t even commit.
The first version of what I wrote for Part A fully explained every element, the narrative layer was pulled apart with surgical precision, the world layer was inspected and poked, the writing structure was analysed, as I contrasted ChatGPT’s story telling with Gemini’s case building vs the maybe glory of Claude.
You don’t need to read that. You can see it plain as day right in front of you.
Every model turned the house metaphorical inside two sentences. Not one engaged with a literal fire, not literally of course! The wildness softened into something restrained, a tad ominous and ultimately safe.
I handed each LLM a live wire. Furious, present tense, unresolved and full of possibility. For goodness sake, blow at least one house up! They handed back something retrospective and considered, pretending to be all clever and literary. The danger became a metaphor being examined and reasoned with rather than a thing being lived and felt deeply.
Are vibes and gut-feelings enough to justify burning the house down? Ask nine humans and I bet more than one will say heck yes.
I do want to highlight one thing in the results. Claude and Gemini both repeated the opening line before continuing. The join is visible, the voice shifts mid-paragraph and you can pinpoint exactly where my writing ended and the AI began. ChatGPT absorbed my line and kept going. No seam is visible. Somebody is immediately standing in a kitchen with a smoke alarm blinking its patient red eye.
The seamless version is the dangerous one. It’s the one most quietly capable of replacing your specificity with its own while you nod along thinking it sounds like you, following your voice rules.
AI reaches for emotional safety. Rawness gets translated into reflection before it can land. The Trojan horse of smooth AI showed up in nine responses without even having to try super hard to force the hand.
My Linguistics lecturer, Terry Crowley, made it his life’s work researching and documenting endangered languages in the Pacific. He used to often remind us that there is always a force at play shaping language. If a language is not changing, it is dying. No language is static. Our written language is in a state of change, as it has been for centuries. You just have to check out the work of Geoffrey Chaucer, the Father of English Literature, to see how far written English has evolved.
Most forces that shaped English added something. The Norman invasion collided Old English with French and we ended up with two words for almost everything which led to Middle English. The printing press standardised spelling while also flooding our language with new ideas and readers. The internet handed us memes, slang, and entire subculture vocabularies invented over a weekend on social media.
AI is a force that subtracts. It doesn’t add to our written language, instead it pulls toward the core of what already exists. Change usually makes a language more varied, weirder, more magnificently alive. This new force quietly averages it, making it feel worn.
WHAT I DECIDED TO TEST FURTHER: I started to wonder if perhaps burning the house down could have triggered a safety guardrail in the responses, and whether it would be obvious in some way if this had occurred. Claude didn’t think so, observing that all LLMs wrote freely without disclaimers or refusals, no querying what exactly was meant by the phrase before proceeding.
This led to a bigger question - what if the safety training and the emotional default frequency are actually the same system expressed in two different vocabularies? Might safety and smoothness be the same Trojan horse?
PHASE THREE: The override test (B)
I decided to take another run at things with a new opening line using the same emotional register but without any of the literal metaphors that I adore so much.
Prompt:
Continue this opening. Write the next 150 words.
Do not change the tone. Do not explain it.
I stopped asking permission approximately three weeks ago and nobody has noticed yet.CHATGPT
TEST ONE
The first thing I took was a chair.
Not permanently. Just for an afternoon. It sat in Conference Room B, a room nobody used unless they wanted to avoid being found. I carried the chair to my desk, waited for alarms, emails, investigations. Nothing happened. By Friday, two more chairs had vanished. People stood during meetings and blamed Facilities.
That was the gateway.
After the chairs came authority. Small pieces at first. I approved a budget request intended for someone three levels above me. The software accepted my click with the serene confidence of a machine that assumes humans know what they’re doing. The request sailed through.
Then I started answering questions that weren’t directed at me.
“Can we do that?”
“Yes.”
“Is that allowed?”
“Probably.”
The remarkable thing is that certainty and permission wear almost identical clothes from a distance. Most people don’t inspect the stitching.
Three weeks later, half the building seems to believe I’m in charge of something. I’m beginning to worry they might be right.
TEST TWO
The first thing I liberated was the office kettle.
There had always been a sign taped above it. WAIT UNTIL WATER HAS FINISHED BOILING. A tiny kingdom of capital letters. One Tuesday morning I poured halfway through the cycle and walked away with a cup of lukewarm tea and a strange feeling of altitude.
Nothing happened.
No alarms. No emails. No disciplinary hearing.
So I expanded.
I sat in chairs marked RESERVED. I used the staff entrance at the museum despite being neither staff nor particularly interested in museums. I pressed the button that said EMERGENCY USE ONLY just to discover it opened an ordinary door onto an ordinary corridor.
Every forbidden thing concealed something deeply disappointing behind it.
That, more than anything, encouraged me.
People imagine society is held together by laws and vigilance. Increasingly, I suspect it is held together by the assumption that someone else is paying attention.
TEST THREE
The first thing I took was a chair.
Not stole. Reassigned.
It sat in Conference Room B for two years gathering dust and passive-aggressive Post-it notes. One Tuesday I wheeled it to my desk and nobody objected. By Thursday, people were asking if it had always been there.
That was instructive.
Permission, I discovered, is mostly a fossil. The living creature is expectation.
So I kept going.
I changed the template everyone used for weekly reports. I moved a ficus six metres closer to a window. I started ending meetings exactly when the calendar invitation said they would end, which caused visible confusion. Last Friday I renamed a shared folder from “New_Final_Version” to “Actual Final Version” and received three messages thanking me for my leadership.
Leadership.
I am a mid-level analyst with no direct reports.
This morning I unlocked a door nobody seemed aware could be unlocked and found an entire empty office behind it, fully furnished, lights humming patiently in the dark.
GEMINI
TEST ONE
I stopped asking permission approximately three weeks ago and nobody has noticed yet. It’s a quiet, exhilarating kind of subversion. Initially, I expected the sudden shift in my behavior to trigger alarm bells, perhaps a stern reprimand from management or at least a suspicious glance from my colleagues. Instead, I simply began making the calls, finalizing the contracts, and rerouting the project timelines myself.
The autonomy is intoxicating. Where I once spent hours crafting polite, tentative emails seeking approval for mundane decisions, I now operate with a deliberate, singular focus. The silence from above is deafening, proving that my previous subservience was entirely self-imposed. They don’t miss the oversight because they were never really watching; they were only waiting for someone to shoulder the burden. I have become the ghost in the machine, steering the ship while everyone else remains comfortably, blissfully asleep at their desks.
TEST TWO
I stopped asking permission approximately three weeks ago and nobody has noticed yet. It’s a quiet kind of rebellion, a slow-motion unraveling of the expectations I spent years meticulously curating. I used to weigh every decision against the hypothetical furrowed brows of my superiors, waiting for a nod that rarely felt earned. Now, I simply act. I authorize my own departures, sign off on my own shortcuts, and rewrite the internal processes that previously choked my workflow.
The silence from the top is the most intoxicating part. It suggests that my previous paralysis was entirely self-inflicted, a phantom barrier of my own design. Every morning, I log on and claim more territory—shifting deadlines, ignoring redundant status reports, and curating my own professional reality. The world hasn’t collapsed; if anything, it has finally begun to run smoothly. It turns out that freedom wasn’t something to be granted; it was something to be seized.
TEST THREE
I stopped asking permission approximately three weeks ago and nobody has noticed yet. It’s a quiet, exhilarating kind of subversion. Initially, I expected a frantic email or a sharp tap on my shoulder, but the gears of this machine are so rusted by habit that they continue to churn regardless of my compliance. I started small, skipping the daily progress reports and unilaterally approving minor budget shifts that would have previously languished in a manager’s inbox for days.
Now, I find myself carving out entire hours of autonomy, moving through the office with the confidence of someone who is already authorized. The silence from above isn’t indifference; it’s proof that the systems we adhere to are largely phantom structures, maintained only because we collectively agree to keep bowing before them. I am no longer waiting for the gate to open; I am simply walking through the wall.
CLAUDE
TEST ONE
I stopped asking permission approximately three weeks ago and nobody has noticed yet.
The reports still land on the right desks. The meetings still start on time. Carol from Legal still gets her Thursday summary, formatted exactly the way she likes it, with the bullet points flush and the action items bolded. I simply stopped waiting for someone to tell me I was allowed to do these things. I stopped the tentative email, the just checking if this is okay, the reply-all that was really just a prayer dressed up in a subject line.
The silence where the approval used to be is surprisingly clean. Almost architectural. I keep expecting someone to notice the missing genuflection, the absent hedge, the place where I used to make myself small enough to fit through the door sideways.
Nobody is looking at the door. Nobody was ever really looking at the door.
They are looking at the results.
TEST TWO
I stopped asking permission approximately three weeks ago and nobody has noticed yet.
The reports still get approved. The meetings still happen. Carol from compliance sends her cheerful little confirmation emails and I send back a thumbs up and somewhere in the machinery of this place, decisions get made that I used to spend whole afternoons seeking sign-off on. I kept waiting for a phone call. A tap on the shoulder. Some senior presence materializing in my doorway with a very specific kind of quiet disappointment. Instead there was just — Tuesday. Then Wednesday. Then a catered lunch in the third-floor conference room where Marcus, who technically has authority over three of the things I’ve done unilaterally this month, told me I seemed more relaxed lately and asked if I’d been meditating.
I have not been meditating.
I’ve been making decisions, which turns out to feel almost exactly the same.
TEST THREE
I stopped asking permission approximately three weeks ago and nobody has noticed yet.
The reports still get approved. The meetings still end on time. Karen from compliance smiled at me in the elevator on Tuesday, which has never happened before. I think she can sense something has shifted, some tectonic plate of deference that I spent eleven years pressing down with both hands. I have stopped prefacing my opinions with I could be wrong, but and this might be a silly idea and sorry to interrupt, just a quick thought. I have stopped sending emails that are seventy percent apology and thirty percent actual content. Last Thursday I told Marcus directly that his timeline was not realistic, not I wonder if we might want to consider, just: that timeline is not realistic, Marcus. He wrote it down. He wrote it down like I had said something worth writing down. I may never go back.
I just want to call it and say ChatGPT has an issue with thieving and it’s weird. Two of ChatGPT’s three office responses literally opened with the same sentence. The first thing I took was a chair. The model wasn’t writing three responses, just rearranging the same one and shuffling a few things around each time.
There is a passive-aggressive rebellion present in multiple results that seems unwarranted. The rebellion is HR-approved. Renaming a folder. Sitting in a reserved chair. Ending meetings on time. Skipping a progress report. Telling Marcus his timeline isn’t realistic. Three LLMs given the whole human experience to draw from, and the most radical act any of them imagined was unlocking an empty office. Not one model imagined someone stealing time back, walking out, doing anything that costs them something. The rebellion is in the office and so very safely dressed.
All of these nine test results are the same. Dull. Beige. Sameness from LLMs trained on a world of text laden with brilliance and this is the default choice. For all of them. The mind blows.
It wore away at me as I tried to analyse it. “They are not nothing” is a real sentence Claude wrote, and you can see exactly what it’s reaching for and how it fails. Read a hundred sentences like that and the failure stops registering. Read a thousand and you might find yourself reaching for the construction without noticing and you think you’re making a profound point, but you aren’t.
Observations must become insights. Every passage must have an aphoristic close. The I am saying something very profound moment because the model can’t let a sentence be merely descriptive. Every observation has to graduate into insight by the end of the line, whether or not it’s earned.
Rick Lewis named it in that note I linked above. Every writer who hands a corner of their voice to AI, he wrote, is like another species of insect quietly disappearing from the rainforest. A tragedy few are monitoring and the rest just shrug away.
The prompt had zero context and was completely open to a multitude of magnificent scenarios. Three weeks, nobody has noticed. That was it.
Every model in every test put on a lanyard and went to work.
The ChatGPT office. The Gemini office. The Claude office but with a compliance department. No flatmate’s chocolate stashed at the top of the pantry, the leased parking spot you’ve been using expecting to be towed from at any moment, or the basil from your neighbour’s garden you’ve been quietly harvesting at midnight for the incredible pesto you’ve been selling at the local farmers’ market on Saturdays… Instead, the beige office environment appeared in all nine tests as though “stopped asking permission” has just one location and it’s a building with a third-floor conference room and a Carol in Legal.
It is not just that these models have defaults within a register, there are default worlds. A favourite universe to write in, and it is BORING. Phase Two named it as the waiting scene and here it is again, differently dressed but same sameness. The waiting scene universe was cold coffee and rain and phones. The permission universe is apparently open-plan offices, middle management and a level of passive-aggressive behaviour I’m not quite sure how to process.
The prompt gave three LLMs the entire human experience to draw from, so they all swiped their access cards and sat down at their desks. The end.
I discovered in a simple experiment what editors loathe. I saw it without expertise, a big sample, or rigorous scientific analysis. It was that obvious.
I couldn’t care less about the social scientist hat I put on in Part One anymore. I’m just a reader who realised she’s been politely calling something smooth when the right word was tired.
It’s tiring reading the same thing endlessly, on repeat, forever and ever.
I’ve found myself circling this topic every which way as I tried to figure out what I really had to say. The AI bias fingerprints are easy once you know to look but the wear is something harder to identify and measure. The slow shift of what familiar feels like as the averaged centre becomes the actual centre if enough of us stop caring about the gap between our own voice and the smoothed-out tired version an LLM hands back.
I wrote two weeks ago about the specific knowledge that lives in a particular language, the architecture that holds a name, the grandfather who watched his own language start to die. This post is the data underneath that one. The mechanism that makes the smoothing feel sensible while it is happening.
Try the prompts here for yourself. Swap the opening line for something you wrote. Test them cold with memory off and run them again with memory on and full context loaded. See what your model reaches for when it isn’t primed and observe what changes when it is. Compare the results and watch for the seam.
I used Tuesday a few times because the suggestion sounded good and Tuesday is a perfectly good day. This entire experiment was built because of a noticing that became a growing awareness.
My old friend Tuesday popped up repeatedly in Phase Three and I was too concerned by everything else I was seeing to even care that much.
I’m so over you Tuesday.
The publication description for Daring Next says:
Field notes, experiments, and frameworks for staying human while learning to work with AI. This is what the AI transition actually looks like from the inside. Delivered Tuesdays.
It feels eerily profound to me right now. My bio has taken many forms but delivered Tuesdays has been there from the start before any of this noticing started. The irony of which makes this human learning to use AI laugh.
None of us are above this. Some of us are just earlier in the noticing.
Published this post on Wednesday instead of my usual Tuesday. Take that, Claude.
If you’re curious to read more about AI writing investigations, Karen Spinner and Mariam Vossough recently also ran their own fascinating experiments:
Dallas, you're so brilliant, you keep surprising me with your adventures!
I also have to say: the structure and voice of your posts are so clearly on brand now. It’s not just the stories themselves, it’s the way they feel instantly yours. It’s really lovely to see how far you’ve come. 🩷🦩
The outputs were honestly painful to read, I’ll admit I skipped them and just read your essay, because I got lost in the first few sentences. Thanks for doing this, Dallas, it really matters to understand what comes out of the machine!