I told the machine it was weaponizing language.
I had asked it to help me define some words. Not marketing words — the words I actually use, the ones that hold a worldview together. It came back with definitions that were clean, confident, and wrong in a way I couldn't immediately name. So I sat with it, and then I could.
It had taken my terms, dropped my sources, and rebuilt them on a frame I never gave it. The output was fluent. It was also, in the strict sense, not mine anymore.
If you use these tools to write, you already know the feeling even if you'd describe it differently. People say "it sounds like ChatGPT." They say "I lost my voice." They say "it made it generic." There's now an entire product category built on that complaint — tools that promise to humanize your text, which is to say, to remove the evidence.
I think almost everyone is describing the same thing and nobody is naming it correctly. So I'm going to name it, because the name determines what you do about it.
What actually happened
I had given it specific terms with specific lineages. Sources, dates, publications, a flag design from 1967. It produced definitions that kept the terms and lost the lineage, then organized them onto an axis of its own invention.
Nothing was deleted. Everything was smoothed. The terms survived; what made them mine didn't.
The common word for this is flattening, and flattening is a good description — of the symptom. It tells you the surface got levelled. It doesn't tell you why the level landed where it did, or why it always seems to land in the same place.
That's the part worth having a word for.
The man who made the telephone legible
In 1876, a draftsman at a Boston patent firm prepared the drawings for Alexander Graham Bell's telephone patent application. His name was Lewis Latimer.
The drawings weren't decoration. A patent office can't grant what it can't read, and Latimer's job was to take an invention that existed in one man's head and on his bench, and render it in the one register the patent office understood. That is how I describe the whole job of communication: the thing that makes an invention legible enough to be real.
Look him up today and he is usually filed under a category — "Black inventor" — rather than the function he actually performed, which was translation. The record kept the man and lost the load he carried.
Hold onto that, because it's the pattern this whole piece is about.
The mechanism
When a system — a model, a filing scheme, a committee — encounters something it has no category for, it doesn't stop. It reaches for the nearest category it does have. The gap gets filled with the majority pattern. That is what prediction is.
So when a term has no slot, the system doesn't report "no slot." It reports the nearest neighbour, confidently.
Now the part that makes it structural rather than merely technical:
The absence gets read as a deficit, and it gets read that way in only one direction.
When a European framework has no word for something in a Yorùbá one, that reads as particularity — an interesting local nuance. When a Yorùbá framework has no word for something in a European one, that reads as a lack.
Same absence. Opposite interpretation. The difference isn't linguistic. It runs down the power gradient.
I call this the one-way reading.
Not flattening — flattening is what you see afterward. The one-way reading is the rule that decided which way the flattening would go. And it's a rule you can test in thirty seconds, in either direction, which is what makes it useful rather than merely true.
Yorùbá marks seniority and not gender: ẹ̀gbọ́n, elder sibling; àbúrò, younger sibling. English marks gender and not seniority: brother, sister. Each language is missing exactly what the other requires. Only one of the two has ever been called a gap.
Run that on any pair of frameworks you work in. Move the anchor and watch what changes. Once you can see it, you can't stop.
The correct reading of an absence is almost never deficit. An absence usually means a different structure is present — one your measuring frame has no instrument for. Oyèrónkẹ́ Oyěwùmí's argument about gender in Yorùbá society is exactly this: not that a category was missing, but that a different organizing principle was doing the work, and the observers lacked the apparatus to see it as anything but a hole.
It predates the machine by a century
This is where you'd expect me to blame the technology. I'm not going to, because the receipts point somewhere else.
Open a Dewey Decimal classification. Class 200 is religion. Around seventy percent of the available numbers go to Christianity, and every other tradition on earth shares the 290s. Hinduism, Jainism and Buddhism are together in 294 — three traditions, one number, roughly a fifth of humanity. That's not a statement about how many religions exist. It's a statement about who was holding the pen in 1876, and it has shaped library shelves ever since.
Ranganathan published a different approach in 1933: faceted classification, where a thing is described by its attributes rather than forced into one slot in one hierarchy. It handles plurality natively. It is also not what most of the world's shelves run on.
Or take N'Ko — the script Solomana Kanté finalized on 14 April 1949, after five years of experiment, explicitly because he rejected the claim that Manding languages couldn't be written. The absence of a script was being read as a deficiency of the language. He answered by building the instrument.
A model trained on text produced inside those frames inherits their shape. It did not invent the one-way reading. It scaled it, and it made it fast.
And it carries forward measurably. Hofmann, Kalluri, Jurafsky and King found that language models asked to make decisions about people on the basis of how they speak were more likely to assign speakers of African American English less prestigious jobs, to convict them, and to sentence them to death. Not on content — on dialect, with race never made overt. Their finding is that the covert stereotypes are more negative than any human stereotype about African Americans ever experimentally recorded, and closest to those measured before the civil rights movement. The models' stated attitudes were fine. Their behaviour was not. (Nature 633, 147–154, 2024.)
What I built instead of complaining
Diagnosis is cheap. What I built is a correction log, and it is not sophisticated. That's the point.
Every time the machine flattens something of mine, the correction gets written down as a rule — not a note, a rule, in a file the system has to read before it writes to me again. Every time it gets something right that it previously got wrong, that gets written down too, as an example.
The log has one measure: does the same class of correction fire twice?
That's the whole evaluation. Not accuracy. Not satisfaction. Recurrence. A correction I have to give a second time was never learned, and it doesn't matter how gracefully the thing apologized the first time.
Two things I didn't expect from running it:
It works on me too. I've made errors the log caught — claims I made at the wrong level, findings I attributed to the wrong document. The log doesn't care which of us flattened something.
The first recurrence it caught, it caught before I did. A rule I'd written two weeks earlier triggered a re-check that found my own mistake. Nobody corrected me. The record did. That's the point where this stopped being a complaint and became a system.
Why I'm not building a humanizer
Back to that product category.
The demand is real and it's growing fast. People want tools that strip the fingerprints off machine-assisted work so it reads as fully human.
I understand the impulse and I think it's the wrong bet — not morally, structurally. Concealment accepts the premise that machine contribution is a deficit to be hidden. It's the one-way reading again, pointed at yourself. You've agreed that the absence of a human hand is a lack, and then you've gone looking for a tool to paper over it.
The alternative is to measure it. Not hide the mix — state it. How much of this was yours, how much was built with a machine, and where the line falls.
That's harder to build and it's where the ground is open. The concealment market is crowded, cheap and defensive. Disclosure is nearly unserved, and it's where regulated industries, education and publishing are all heading anyway.
That's the tool I want to build, and I'm not going to build it alone.
I'm looking for three organizations to build it with — schools, libraries, community groups, publishers — anyone deciding what to do about machine-assisted work who would rather measure it than police it. If that's you, message me.
The one line
If you take nothing else:
When something is missing from a description, don't assume a deficit. Assume an instrument.
Then go find out what the other frame was measuring that yours cannot see.
Terence Latimer is an Information, Communications & Technology professional and writer based in the Mojave Desert. He writes about moving meaning between registers without losing it.