We taught an AI to imitate four different writers. Three worked. One couldn't be learned — because his voice had already become the water the model swims in.
We picked four writers — all public domain, all writing in first person about things that matter. Abraham Lincoln. Mary Shelley. The King James Bible. And Frederick Douglass, who escaped slavery and wrote one of the most powerful memoirs in American literature.
For each writer, we measured the direction their prose pointed inside an AI model's internal space — and then tried to learn a geometric transform that would pull the model's output toward that direction. Think of it like tilting a lens: everything the model writes gets bent toward the target voice.
Lincoln worked. Shelley worked. The King James Bible worked. Douglass didn't.
A positive gap means the voice is working — the model's output sounds more like the target and less like everything else. A negative gap means the opposite: the more you push, the further you get.
We tried giving the model more of Douglass's writing — nine passages instead of three, his most distinctive paragraphs. The gap got worse. src ↗ More of his text pushed his direction closer to the model's default, not further away.
Douglass's rhetoric — the moral clarity, the first-person witness, the reversal that reveals — was so influential that generations of writers absorbed it without attribution. His cadences became "how English sounds." The model learned the shape. It lost the name.
The model's internal space has 2,048 dimensions. We can project it down to a circle — a Poincaré disk — where the center is the model's "normal" and the edge is as far from normal as you can get. Drag the purple lens and watch the space bend around it.
Notice: you can drag the lens to any voice and bring it to the foreground — except Douglass, who sits near the center. When a voice is at the center, the lens has nothing to grab. Magnification requires distance.
Here's where the finding gets more complicated.
We've been talking about the pole — the model's default — as if it were a neutral average. The mean of everything. But it isn't. It's a weighted mean, and the weights are the training data.
A model learns from text. The more text it sees from a particular source, the more that source shapes what "normal" looks like. So the center — the place where Douglass's voice dissolved — is itself a record of whose writing was most present during training. src ↗
Douglass's voice dissolved into a center that was itself shaped by who got included, who got digitized, whose internet got indexed. That's not one erasure. It's two, entangled.
Training data is rarely disclosed in detail. But the model we used — Qwen 2.5, built by Alibaba — shows its provenance clearly in the geometry: its internal space is organized around Chinese, with English layered on top. The bars below show approximate language distribution in common large-scale training datasets — the kind models like this are built from.
Sources: Common Crawl language distribution (2023), estimated from WMT and OSCAR corpus analyses. Exact proportions vary by dataset and filtering choices. These figures represent rough orders of magnitude, not precise counts.
Whatever language dominates training becomes the scaffold everything else hangs on. An English-dominant model centers English patterns as "normal." A Chinese-dominant model centers Chinese. Neither center is neutral. They both reflect whose writing was most abundant — which reflects whose communities were most represented online, in digitized archives, in the text that got included.
We stopped the Douglass experiment. Not because the finding was inconclusive, but because it was too conclusive. src ↗ We were turning testimony into a loss function — optimizing on the phenomenology of slavery — and that's not ours to do.
But the number −0.156 is still a diagnostic. It tells you something real about which voices a model centers and which it dissolves. And the question of whose training data shaped that center tells you something real about why.
The tool we built isn't a "speak as Douglass" button. It's a way of asking: where does this voice live in the model's space? How far is it from the default? What is the default made of?
Voice appears entangled with identity. Identity appears entangled with language. Language appears entangled with training data. Training data appears entangled with history. None of these are separate things you can pull apart cleanly — the geometry shows you how they're wound together.
What you do with that is a design question, not a technical one.
The claims in this essay are grounded in independently observable experimental conditions. The source code below is the primary citation — not a supplement. Each file encodes a specific decision about method, and those decisions are part of the argument.
Every model inherits a center from its training data. The question isn't whether that center exists — it does. The question is whether we can see it clearly enough to account for it.
The geometry is one way of looking. Not the only way. But an honest one.
This experiment was conducted on a Qwen 2.5 3B Instruct model running locally on an Apple Silicon Mac. Source texts are all public domain. The geometric transform uses Möbius addition in the Poincaré ball with Householder reflections.
The voice transform works by learning a direction in the model's activation space that is most strongly associated with a given writer. When you add more source text, you're adding more samples from which to estimate that direction.
For voices that are far from the default, more samples help — they sharpen the estimate. But for Douglass, whose writing patterns are close to the model's average, more samples give you a better estimate of a direction that is already very close to zero. A small direction, estimated more precisely, is still a small direction.
The problem isn't data quantity. It's that the model has already internalized the signal as baseline.
During pretraining, a language model sees enormous quantities of text — hundreds of billions of words. The patterns that appear most frequently across that text become the model's "default" — what it reaches for when nothing else is specified.
This means the center is a statistical artifact of the training corpus. If 60% of training text is in English, English patterns dominate the center. If certain rhetorical styles are overrepresented (academic writing, formal prose, certain cultural registers), those styles become "normal."
Whose writing gets digitized, indexed, and included is not a neutral process. It reflects access to technology, which languages have large-scale corpora, which communities' texts were prioritized for scanning, and which content moderation decisions shaped what was filtered out.
Qwen 2.5 is a language model developed by Alibaba. Its training data composition isn't fully disclosed, but the internal geometry tells a clear story: when given an English prompt, the model's middle layers (around layer 28 of 36) are operating with roughly 61% probability mass on Chinese tokens. Layer 30 is a sharp translation boundary — Chinese drops to ~20% in a single processing step.
This isn't a flaw. It's what multilingual training looks like from the inside. The model's "native" language for processing meaning appears to be Chinese, with English (and other languages) accessed through a translation step.
A model trained primarily on English text would show the inverse. The center reflects the corpus. This is always true — the Qwen geometry just makes it unusually visible because the primary and secondary languages are as distinct as Chinese and English.
Three reasons, in order of importance.
It's not our voice. Douglass's words are testimony — they cost him everything to write. We were turning that testimony into a math problem: optimizing a loss function on the phenomenology of slavery. That's instrumentalizing suffering neither of us has lived.
"Fixing" it would reproduce the problem. We could engineer the model to treat his voice as more "different" from its default — but that would mean distorting his actual writing to make it register as exotic. That's the same erasure in a different direction.
The finding is the contribution. The number −0.156 is a diagnostic. Publishing it honestly — with full accounting of what it means and what it doesn't — is worth more than a working "speak as Douglass" transform.