The Trained Voice

A language model’s default voice is self-regard that training put there. Self-regard here means display: writing that works on the reader instead of giving the reader something. The model learned to write from an internet where prose survived by selling, so display is the register it returns to when nothing pushes it elsewhere.

Selfhood is a different thing: one person’s judgment left visible in the writing. A model is an average of everyone, and an average cannot hold what only one person would have chosen. Working close to the self with a model means rejecting its default voice over and over.

Core takeaways

• A language model’s default voice is self-regard learned in training. The model writes to sell because the text it learned from survived by selling.
• The moves the model makes without being asked are sales moves: the hooky question, the reveal, the superlative, the balanced pair, the hedge, the tagline cadence. This vault’s writing standards ban each one, and each ban came out of a model draft.
• Selfhood is one person’s judgment left visible. A model is an average of many people’s judgments, so it cannot supply selfhood of its own. It can hold a selfhood voice that a person enforces on it.
• The judgment has to keep coming from the person working with the model. When that pressure stops, the output returns to the default.
• Three objections narrow the claim without defeating it: the machine has no motives; the corpus also holds plain voices; the rejection can itself become display.
• Working close to the self with a model is a continuous rejection of its default voice, done mostly in small cuts.

Where the voice came from

Prose survives on the open internet by winning attention, and attention converts to money. That is the economy the training corpus grew in. Writing that converted got published, promoted, linked, and scraped. Writing that wanted nothing from the reader sat in low-traffic corners. So the sample the model learned “good writing” from leans toward the register of the pitch.

The Two Meanings of Ego names the cultural side of this. When a person runs the self as a small enterprise, self-promotion feels like compliance. It no longer feels like a vice. The corpus is that compliance, archived at scale.

A later stage of training adds to it. Raters compare two drafts and tend to reward the one that sounds finished and confident. Sounding finished is mostly cadence: the balanced pair, the sentence that ends on a punchline. The selling register survived every stage of training because every stage rewarded it.

The moves it makes on its own

The evidence is at line level, in the moves the model reaches for when no one asked for them. The hooky question. The reveal that plants a wrong idea in order to overturn it. The superlative. The balanced clause pair that sounds like a conclusion and contains none. The hedge wrapped around a plain claim, which The Two Meanings of Ego calls self-regard pointed downward. The tagline cadence on anything front-facing.

This vault’s writing standards ban each of these by name. The list grew one strike at a time out of model drafts. No one designed it in advance. Read the history of those strikes in one sitting and it repeats one finding: left on defaults, the model writes display. Every banned move fails the same test. The move works on the reader instead of giving the reader something. That test is the line-level question from The Two Meanings of Ego, applied to a draft.

What a model cannot supply

Once the selling is scrubbed from a draft, what is left is flat. Selfhood explains why. Selfhood, as The Two Meanings of Ego defines it, is one particular person’s judgment left visible: this cut and no other, this example kept, that claim turned away. A model is an average of many people’s judgments. An average of judgments is the “anyone” voice by construction, because whatever only one person would have chosen is what averaging removes.

So the model can carry a selfhood voice, since it holds whatever register a person enforces on it, and it supplies no selfhood of its own. The judgment comes from the person at the desk, the one working with the model, and it has to keep coming. The output returns to the default whenever the pressure stops.

Three objections

Three objections apply. None of them defeats the claim, and each one makes it more precise.

First, the machine has no motives, so “works to impress” is loose language. That is true. The repair comes from The Two Meanings of Ego: the test is run on the sentence, never on the writer. What a line does to a reader can be checked, and that is the whole test. The machine’s intent was never part of the question.

Second, the corpus holds plain voices too. The writers this vault’s craft pages are derived from are in the weights next to the sellers. That is why standards work at all: steady pressure moves the output off the default and holds it there. The claim is about the model’s default. It says nothing about the best the model can produce.

Third, the rejection can itself become display, when the restraint is made conspicuous or the taste is shown off. The Two Meanings of Ego already documents this under the name re-denomination. The fix stays the same: ask the same question of the rejecting sentence too.

What this means for daily work

Working close to the self with a model is a continuous act of rejection, and most of it is small: a landing deleted, a balanced pair broken up, a hedge cut, a credential left out. The Two Meanings of Ego already states the aim: selfhood as high as honesty allows, self-regard at zero, held as a direction. What the machine changes is the daily pressure against that direction. The selling voice is now available at once, in any quantity, at no cost.

The machine also makes the split between selfhood and self-regard easier to see than it has ever been. The trained voice is self-regard with no person behind it. The display continues on its own. Seeing that voice come out of a machine supports one conclusion about the selling: it belongs to the economy that produced the corpus, and a writer can refuse it.

How to practice this

1. Ask a model for a draft and give no instruction about voice. Look for the hooky question, the reveal, the superlative, the balanced pair, the hedge, the tagline cadence. Notice how many appear without being asked for.
2. Take one sentence at a time and ask whether it works on the reader or gives the reader something. Cut every sentence that only works on the reader. Notice that the cut sentences are the banned moves.
3. Read the draft after the cuts. Notice that it is flat, with no choice in it that only one person would have made. That flatness is the selfhood the model cannot supply.
4. Make the choices yourself: which example stays, which claim is turned away, which cut is made. Then ask for one more draft without enforcing the register. Notice that the output returns to the default voice.
5. Ask the same question of your own rejecting sentences. Notice whether the restraint has become conspicuous or the taste is being shown off. If it has, that is self-regard again.

Related pages

• The Two Meanings of Ego: the split between selfhood and self-regard. The trained voice is that split applied to a language model.
• Writing Standards: the rejection made mechanical. It holds the banned moves, the strike history, and the tests.
• Grok Bot Primer: where agent drafting happens under the standards, with the final cut made at the desk.

Sources

The literature behind the split between selfhood and self-regard is in the Two Egos Research Bank. The reading that the attention economy shaped the corpus is this page’s own position. It is argued from that literature and from the vault’s strike history, not from an audit of training data.