12 lyric registers most AI can’t produce — and what they sound like

One of the most consistent complaints about AI-generated lyrics, including ours, is that they all sound like the same person wrote them. That person is an MFA graduate in their late twenties, vaguely melancholy, slightly literary, gender-neutral, and never quite specific enough to be embarrassed by anything they’ve said. Every default LLM voice converges on this register. It is the average of the training corpus.

The problem isn’t that this register is bad. The problem is that it’s the only register the model can write in confidently. Songs that need a different voice — a child, a dying parent, someone halfway through learning English, a religious convert at the moment of conversion — collapse back to MFA-adjacent diction the moment the model has to fill a syllable.

This post is a taxonomy of 12 registers that the default voice can’t access without explicit instruction. For each one we’ll show the marker (what makes the register itself), the trap (what AI usually does instead), and a short snippet of what the register actually sounds like when written from inside.

Why register matters

Register isn’t style. Style is the texture of the writing — sparse vs. ornate, terse vs. flowing. Register is who is speaking. A child and a 70-year-old widow can both speak in sparse style, but their registers are utterly different: the child uses adjectives wrong on purpose; the widow uses them with a precision earned by burying people.

The 12 registers below all describe a kind of speaker the default AI voice cannot inhabit without help. We added them to our pipeline at Build 2294 as Style Mutants — a 12-persona slot the model is prompted to occupy for a small percentage of forges. We released them publicly because the underlying lesson is bigger than our pipeline: if you can name the register, you can write toward it. If you can’t, you’ll write the MFA voice forever.

1. The child narrator

Marker: Adjectives in the wrong slot. Concrete nouns substituted for abstract ones (“the loud” instead of “the noise”). Time handled badly — “last summer” might be three weeks ago or three years.

Trap AI falls into: Cute. Singsong meter. Adults pretending to remember being children. The Sandlot voiceover.

What it actually sounds like:

Dad came home with the loud in his shoes again
Mom said the word and her face did the still thing
I counted the ceiling — twelve cracks, two new
The dog knows when to go small

Notice the syntax is grammatical but the semantic categories are slightly off. “The loud in his shoes” — the child is mapping anger onto a sound they associate with it. Tracy Letts uses this technique in August: Osage County when adults recall childhood; the off-categories are how the speaker signals that the child still hasn’t fully processed what they witnessed.

2. The dying parent

Marker: Time compression. Conversations the speaker keeps trying to finish before the morphine drip kicks in. Detail without preciousness — what they remember is sharp because they’re losing it.

Trap AI falls into: Hallmark-card valediction. “I will always be with you.” Generic last-words posture.

What it actually sounds like:

The yellow cup with the chip — third shelf
Don’t throw it out and don’t fix it
Your father used it the morning he left
I kept it because I am the kind of woman who keeps cups

The dying parent doesn’t deliver a thesis. They deliver inventory. Cormac McCarthy’s The Road works this register on the father throughout. So does Mary Karr in The Liars’ Club when the mother is medicated.

3. The foreigner mid-language

Marker: Syntax from the first language bleeding through the second. Word choices that are technically correct but unmarked by a native speaker (“I make a coffee” rather than “I’m making coffee”). Idioms tried and not quite landed.

Trap AI falls into: Stereotyped “broken English.” Or, worse, a register so polished the speaker’s actual origin disappears and we’re back in MFA territory.

What it actually sounds like:

I came here for to work and now is twelve years
My mother died last summer — I don’t go
The boys say “Dad you sound funny” — yes, I sound funny
But the funny is what I have left of her

Notice how “is twelve years” rather than “it’s been twelve years” lands as Eastern European syntax without naming it. Anne Carson uses cross-language register constantly — translation from Greek into English deliberately preserves the structural strangeness so the reader feels the gap.

4. The twin

Marker: First person singular that keeps slipping into first person plural without warning. Pronoun ambiguity treated as accurate rather than confusing.

Trap AI falls into: Twin-as-metaphor. Two-halves-of-one-soul cliché. Identical-twin TV-movie posture.

What it actually sounds like:

We turned forty in March
She on the Tuesday and I on the Tuesday
The doctor called my body about her cancer
I answered and said yes

The pronouns refuse to settle. “She on the Tuesday and I on the Tuesday” refuses to acknowledge they were on the same Tuesday. This is how twins actually narrate — the singular is a polite fiction. Joan Didion writes this register in The Year of Magical Thinking not as twins but as widow: the “I” keeps reverting to the “we” that no longer exists.

5. The J-pop / K-pop interior

Marker: Emotional declaratives delivered without irony. English used as a small sacred object inside an otherwise non-English mind. Phrases like “forever” or “starlight” that an MFA-trained writer would have removed.

Trap AI falls into: Either parodic — making fun of the register — or, more often, sanitizing it into pop-with-edge. The register requires sincerity AI rarely permits itself.

What it actually sounds like:

I wrote your name in English in my notebook today
The teacher said “very good” and the rain started
I am sixteen and I will love you forever
This is the truth even if later it is not

The last line is the register’s signature. The speaker knows the declaration may not survive. They make it anyway, and the declaration’s sincerity is sharpened, not undercut, by the speaker’s awareness of its limits.

6. The religious convert

Marker: Theological vocabulary used as if for the first time. The speaker tries on language that has been around for two thousand years and treats it as new. The before-and-after split is in the syntax itself.

Trap AI falls into: Either evangelical-radio cliché or atheist-mocking-the-believer. Both collapse the register.

What it actually sounds like:

The word “grace” — I had heard it
I did not know it was a thing that could happen to a body
I was thirty-four. My hands were on the steering wheel
I pulled over because I could not see for the seeing

Marilynne Robinson’s Gilead is the prose-fiction reference here — specifically the register of the old preacher’s narration. The trick is to make theological vocabulary feel small and personal, as if the speaker is borrowing it from a stranger.

7. The jaded ex

Marker: Affection and contempt occupying the same line. The speaker is not over it and is also done with it — both true. Tonal control that lets the listener decide which side is dominant.

Trap AI falls into: Pure bitterness (the angry-breakup-song register) or pure regret (the wistful-acoustic-cover register). Real jaded-ex voice is both simultaneously.

What it actually sounds like:

You always laughed at my coffee order
I changed it the year you left and I changed it back last March
This is not a song asking you to come back
This is a song about the coffee

Adrienne Lenker writes this register on Big Thief records. So does Phoebe Bridgers. The register requires the speaker to refuse the listener’s sympathy without rejecting it — a thread-the-needle move AI almost never lands without help.

8. The veteran

Marker: Specific military vocabulary used without explanation. The speaker assumes the listener will Google it later. Things that happened are referred to obliquely — the speaker is not telling a story about the war, the speaker is in a present in which the war has already happened.

Trap AI falls into: War-movie tropes. Bracketed dialogue. The Hollywood register.

What it actually sounds like:

The grocery store fluorescents at 2 AM are the same fluorescents
My daughter was born on the day of the Helmand thing
I have not told her
She is twelve and I have a lifetime to not tell her

Kevin Powers’s The Yellow Birds works this register. So does Phil Klay’s Redeployment. The marker is the casualness of the reference combined with the obvious weight under it.

9. The recovering addict

Marker: Numbered time. Specific count of days, weeks, months. A vocabulary borrowed from the program (meeting, sponsor, the rooms) used flatly. The speaker resists drama because drama was the problem.

Trap AI falls into: Narcotics-Anonymous-meeting parody. Or, oppositely, a romanticized using-register (the user is interesting; the recovering user is dull).

What it actually sounds like:

Day seven hundred and three
The coffee is bad and the chairs are bad and the people are fine
My ex left a message Wednesday — I did not call back
This is what it is. This is the size of it.

The register is anti-poetic on purpose. The speaker has spent years not telling stories about themselves and now uses the smallest possible language for the largest possible facts.

10. The court witness

Marker: The speaker is telling the truth and is aware they are telling the truth in a setting where lies are common. Sentences are short and declarative. The speaker refuses to characterize anyone, including themselves.

Trap AI falls into: Procedural-drama register. The Witness-on-Stand cliché.

What it actually sounds like:

I saw the man on Tuesday
He was wearing a brown jacket
He said the word and I said the other word
I am not going to tell you which words. You will ask, and I will say.

Notice the last line breaks the fourth wall on the witness register itself — the speaker is signaling that they’ve been here before, that the cross-examination already happened, that they are protecting someone. The register is a frame inside a frame.

11. The teacher addressing a former student

Marker: Affection mediated through institutional formality. The speaker remembers the student as a child but addresses them as an adult, and the gap between those two is the song.

Trap AI falls into: Dead-Poets-Society register. The teacher-as-sage. Or worse, the teacher-as-creepily-fixated-on-students.

What it actually sounds like:

I taught you long division in 1997
You were nine. Your shoes were always untied.
The newspaper said your name yesterday — I read it twice
I am writing because I do not know how to be proud of you in private

The register requires institutional memory (long division, untied shoes) and adult acknowledgement (the newspaper, the formal letter posture). The asymmetry between the two is the entire song.

12. The translator

Marker: The speaker is rendering someone else’s words and is aware they are doing it badly. The lyric foregrounds the gap between what was said and what can be said in English.

Trap AI falls into: Skipping the gap. The translation reads as if it were original, which collapses the entire register.

What it actually sounds like:

My grandmother said something in Bengali I cannot say in English
The closest is “the body remembers” but it is not that
It is something about the body and something about salt
I have spent forty years trying to translate it. I have not

Anne Carson lives in this register full-time. So does Ocean Vuong, particularly in On Earth We’re Briefly Gorgeous. The translator register is a register about register — it asks the listener to hold the absence of an exact word as part of the song’s meaning.

How to use this

If you’re writing with an AI tool (ours included), you can prompt for these registers by name. “Write this in the dying-parent register” pulls the model away from the MFA default toward the specific syntactic markers above. The result will be imperfect — none of these registers are easy to fake — but it will be markedly more specific than the default.

If you’re writing without an AI tool, the taxonomy is still useful. Most songwriting workshops teach style (sparse vs. ornate, narrative vs. lyric). Almost none teach register. The 12 above are a starting list; the full taxonomy is much longer. Try one on your next draft and see whether the syntactic markers do the work.

The goal isn’t to use unusual registers for novelty. The goal is to recognize that you have already chosen a register the moment you started writing — usually the MFA default, usually invisibly. Naming it lets you choose another. Most songs that survive are songs whose register couldn’t have been written by anyone else.

The 12 Style Mutants are wired into the SongForgeAI forge pipeline as of Build 2294. They activate at low frequency on individual songs — roughly 5% — and produce the specific syntactic markers above. Try the forge with an unusual brief and watch for the register shift in the output.