Voice · Identity · Training Data

What Voices Speak

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.

Unity Environmental University · Local Model Experiment · 2026


The finding

One voice couldn't be learned

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.

+0.165
King James Bible — voice gap
+0.141
Lincoln — voice gap
−0.156
Douglass — voice gap

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.

His voice isn't missing from the model. It's dissolved into the model's baseline.

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 geometry

Where each voice lives

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.

Drag the lens to try to bring each voice into focus
Frederick Douglass
Lincoln
Shelley
King James Bible
pole (default)

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.


The deeper question

But what is the center, exactly?

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 ↗

Erasure doesn't require absence. It can happen through universalization — when a contribution is absorbed so completely that its source disappears.

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 provenance

Whose text shaped the center?

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.

Approximate language share in large web training corpora

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.


What this means

The geometry tells the truth

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.

The center always reflects whose text was most abundant. And whose text was most abundant reflects decisions made long before the model was trained.

What you do with that is a design question, not a technical one.


Sources & Methods

The experiment is the argument

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.

fisheye_persona.py
Experiment 1. Computes voice gap scores for Lincoln, Shelley, KJV, and Douglass using a Poincaré ball fisheye transform. The multi-passage Douglass test — adding nine passages instead of three — is in this file. The gap scores in this essay come from this experiment.
fisheye_learn.py
Experiment 2. Learns voice transforms from source text using 8 Householder reflections (16,384 parameters per voice). The geometric transform described in the essay — "tilting a lens" — is implemented here.
logit_lens.py
Experiment 3c. Projects intermediate layer activations through the model's unembedding matrix at each layer to observe what tokens are most probable — the technique behind the language-shift visualization in the companion explainer.
logit_lens_langs.py
Experiment 3d. Aggregates logit lens output by language (using Unicode block detection) to produce the layer-by-layer language probability mass data. The 61% Chinese at layer 28 finding comes from this file.
fisheye_interference.py
Experiment 3. Proximate-other interference: tests Wilde/Whitman (queer pair) and Shaw/Thoreau (control pair). Fisher separability metric, injection layer comparison [12,18,24] vs [30,32,34]. The queer voice content moderation flag occurred during this experiment.
VOICE_GEOMETRY.md
Technical documentation and ethical accounting. Records all design decisions, including the decision to stop the Douglass experiment, the content moderation incident, and the choice of queer voices as the next experimental pair.
Full repository ↗
All code, raw output JSON, and documentation. Model: Qwen 2.5 3B Instruct, running locally on Apple Silicon (MPS). All source texts are public domain.

The water
was already there.

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.