# Novel II — Voice and Style: The Unreliable Narrator
In ["Novel II — Voice and Style: Imitation and Homage"](/harrys-desk/novel-ii-voice-and-style-imitation-and-homage) I said homage is what remains after you have copied a master long enough to steal the seeing, not the signature. On Monday, in ["Finding the Narrator"](/harrys-desk/novel-ii-voice-and-style-finding-the-narrator), I said voice is the residue of a specific consciousness under pressure, and that the model has no consciousness, only a manner. If you did that work, you have a mouth you can hear. Today we let that mouth lie.
The claim is narrower than the workshop slogan. Unreliability is not a twist, and it is not the decision to withhold a fact until page two hundred. It is a designed gap between what the speaker pretends — to know, to value, to have seen — and what the page lets a careful reader reconstruct. The gap is the instrument. Hide it completely and you have a narrator who is wrong in private, which is not a technique. Announce it and you have a trick. The model, trained to be helpful, complete, and a little proud of its honesty, will do both of those failures in the same paragraph unless you stop it.
By the end of this article you should be able to name the distance your narrator is working, keep the truth document out of the prompt, and tell a leak from a wink.
Not a Twist, a Distance
Wayne Booth gave the term its modern life in *The Rhetoric of Fiction*. An unreliable narrator is not a narrator who is interesting, or damaged, or fond of drink. It is a narrator whose account diverges from the implied author's, in a way the reader is invited to notice. The book has two minds on the page: the one speaking, and the one that arranged the speaking. The pleasure — and the moral work — is the reconstruction.
This is why Monday's portrait has to come first. You cannot design a lie for a mouth you cannot hear. A generic first person who "might be hiding something" is not unreliable. It is vague. Vagueness is what the model will give you if you ask for an unreliable narrator without specifying the distance. You will get fog: *I think*, *maybe*, *I can't remember*, *looking back now*. That is weather. It is not a gap.
Holden Caulfield and Humbert Humbert, whom I named on Monday as sharing a pronoun and nothing else, also share a structural fact. Both are asking the reader to take their account as sufficient. Both are wrong, in different directions. Holden is a values problem dressed as a facts problem; he reports the world as phony because that is the only verdict his pride can afford. Humbert is a values problem that pretends to be a style problem. The sentences are exquisite. The seeing is a crime. If you steal Humbert's adjectives and miss the indictment the book is making *through* them, you have done pastiche of a monster and called it homage.
The implied author is not a ghost you put in the prompt. It is you, in the decisions you refuse to let the narrator make for the book. In ["Authenticity — When Is It \"Your\" Work?"](/harrys-desk/authenticity-when-is-it-your-work) I argued that authorship is a matter of decisions, not of keystrokes. Unreliability is one of those decisions, sustained. If the model is also deciding what the reader is allowed to know, you no longer have a designed gap. You have a slot machine.
Facts, Knowledge, Values
You do not need a taxonomy to write. You need to know which distance you are playing, because the three distances fail in different ways, and the model collapses them into one flavor.
Facts: the narrator misreports what happened. The gun was in the other hand. The letter was never sent. *The Murder of Roger Ackroyd* is the clean case, and it is also the case most writers should not imitate, because the lie is a puzzle-box and the voice is secondary. If your unreliability is only a withheld fact, you have a mystery, not a narrator. Mysteries are honorable. They are a different instrument.
Knowledge: the narrator does not have the equipment to understand what they saw. The child, the outsider, the servant who was in the room and cannot name the politics. Huck Finn knows he is helping Jim and believes he will go to hell for it. The gap is not that Huck is lying to you. The gap is that his moral vocabulary cannot hold the good he is doing. The reader sees around him because Twain has given the reader a second language Huck does not have.
Values: the narrator has the facts, more or less, and judges them wrongly. This is the richest distance for a novel, and the one the model is worst at, because values unreliability requires a speaker who is coherent, fluent, and wrong. Stevens in *The Remains of the Day* does not forget that he served a Nazi sympathizer. He has a professional language that will not let the word land. The dignity is the lie. Ishiguro never winks. He lets Stevens talk until the dignity becomes audible as damage.
Pick one distance and be ruthless. A narrator who misreports facts *and* lacks knowledge *and* has crooked values is not complex. It is noise. The reader cannot reconstruct a signal if every channel is scrambled. In ["Novel II — Deep Craft: Point of View"](/harrys-desk/novel-ii-deep-craft-point-of-view) I said POV is a contract about who knows what, and when. Unreliability is a second contract, laid over the first: the speaker will not honor the first contract in a particular, designed way. If you have not named the way, you have not made the contract. You have hoped the reader will find the book clever.
Hedging Is Not Lying
Here is the AI accent of unreliability, and it is more specific than the accent I named on Monday.
The model has been trained, at great expense, to be helpful and honest. Honesty, in that training, means: do not leave the reader lost; do not contradict yourself; if you are unsure, say so; if there is a truth, surface it. Those are virtues in an assistant. They are fatal in a lying narrator.
Ask a model to write an unreliable scene and you will usually get hedging. The narrator is "not sure." Details "might" have been different. Time is "a little fuzzy." This reads as unreliability to a prompt, because the training data is full of essays that describe unreliable narrators as confused, traumatized, or drunk. Confusion is sometimes a cause of unreliability. It is not the design. A designed liar can be crystal clear. Stevens is crystal clear. Humbert is crystal clear. Clarity is what makes the gap visible. Fog hides it.
The second substitute is confession. The model, eager to be complete, will have the narrator admit the lie in the same scene that tells it. *I suppose I should say what really happened.* *Even then I knew I was not telling it straight.* That is the implied author climbing into the narrator's mouth to protect the reader. It is also the model protecting *you* from the discomfort of letting a speaker be wrong without a footnote. Strike it. If the narrator would not confess, the sentence cannot stay, no matter how useful it is to a first-time reader.
The third substitute is the SparkNotes wink: diction that performs "unreliable narrator" as a genre, the way a bad pastiche performs Faulkner. You will see it in overworked memory-failure, in theatrical self-correction, in the paragraph that ends with a thesis about self-deception. Mannerism, again. On Monday I said mannerism is a recurring choice detached from seeing. Unreliability-as-flavor is a mannerism. The seeing, here, is the particular way *this* speaker needs the lie.
Keep the Truth Out of the Prompt
The practical method is two documents, and they must not meet inside the model.
Document A is the truth. What happened. What the narrator believes happened. Where those two diverge, and why the narrator needs the divergence — pride, love, professional language, a promise, fear of a fact that would unmake them. You write this yourself, model closed. One page is enough. It is not a chapter. It is a map of the gap. Do not put it in a prompt. If you do, the model will smuggle it. Models are excellent at smuggling. They have been rewarded, for years, for making the hidden thing available to the user.
Document B is the spoken scene. The narrator tells. They do not have Document A in front of them, because they *are* the reason Document A exists. You write the first pass of B by hand, in the mouth from Monday's tuning fork, under the pressure from Monday's homework. Then, if you use the model, you may give it the narrator portrait, the sample passage, and the refusals — will not confess, will not hedge, will not summarize the meaning of the scene — and you still do not give it Document A.
Generate once. Do not iterate toward "clearer unreliability." Clearer, to the model, means more leaks and more winks.
Then you do the work the model cannot do. Put Document A next to Document B. Read them against each other, aloud. Mark three things in B:
Leaks: any sentence that could only have been written by someone who has read A. Vocabulary the narrator should not have. A causal link they have not earned. A joke that requires the truth. These are the model being helpful.
Winks: any sentence whose job is to tap the reader on the shoulder. Self-corrections that exist for us, not for the speaker. Thesis endings. *I told myself.* These are the model being a teacher.
Blanks: places where the gap is unreadable — where a careful reader, given only B, could not possibly reconstruct anything, because you hid the contradiction instead of staging it. Unreadable is not subtle. Subtle is a contradiction the reader can feel as friction. Unreadable is a locked drawer.
Rewrite the leaks and the winks by hand. For the blanks, you do not add a confession. You add a second pressure: another detail the narrator will *happily* give you, which does not fit the evaluation they are selling. Huck will tell you about Jim. Stevens will tell you about the silver. The reader does the rest.
This is slower than "write this scene with an unreliable narrator." It is the only method I have that does not slowly replace the gap with a guided tour.
The Ledger
A short story can live on one distance and one scene. A novel cannot. Scene-by-scene generation will drift toward honesty for the same reason it drifts toward the AI accent: each new prompt wants a complete, helpful unit, and completeness is the enemy of a sustained lie. By chapter eight the narrator has explained themselves. By chapter twelve they are the default generalist again, with a hobby of being wrong.
Keep a ledger. It is ugly and it works. After each chapter, four lines, model closed:
What they claimed. What the reader has been allowed to see. What the narrator still cannot say. Whether the distance (facts, knowledge, or values) held.
If the ledger starts needing a fifth line — "reset the lie because the last chapter told the truth" — you have a structural problem, not a prose problem. Go back to Document A. The chapter that confessed is the chapter you rewrite, not the chapter you "complicate."
The ledger also stops the cheap last-act reclassification, in which a narrator who has been merely limited is revealed, in the final twenty pages, to have been a liar all along. If the voice was not already doing the work, the reveal is a hat. Readers can feel a hat. In ["Novel II — Deep Craft: Pacing and Tension"](/harrys-desk/novel-ii-deep-craft-pacing-and-tension) I said tension is the reader's desire to know. Unreliability weaponizes that desire only if the reader has been given a pattern to doubt, early, in the grain of the speech. A pattern. Not a fog, and not a footnote.
I have watched AI-assisted drafts do the hat because the model, given a late prompt that says "now reveal the narrator was lying," is very good at writing the reveal and very bad at having already written the book that earned it. You cannot prompt your way backward into a design. You can only have made the design, chapter by chapter, and refused the model's offer to be clear.
In ["The Integrity Framework — A Writer's Code for AI Use"](/harrys-desk/the-integrity-framework-a-writers-code-for-ai-use) I said the writer is accountable for what the page finally is. A lying narrator does not transfer that accountability to the character. If the gap is unreadable, or theatrical, or abandoned in chapter ten, you did that. The model was only fluent.
For Next Time
On Monday we leave voice-and-style and open the next triad: ["Novel II — Worldbuilding: Setting as Character"](/harrys-desk/novel-ii-worldbuilding-setting-as-character). A world is not a backdrop you describe after the people have spoken. It is a set of pressures that make some speech possible and some speech a lie. Setting, done properly, is a second witness. It can corroborate a narrator or quietly refuse them — the house that is colder than the memoir admits, the town that does not match the legend. If this week's speaker is going to keep lying, they will need a world that does not collude.
Your homework until then: take Monday's narrator. Decide one distance — facts, knowledge, or values — and write Document A, one page, model closed: what happened, what they believe, why they need the divergence. Then write four hundred words of them telling a scene that includes the lie, still model closed. Generate the same scene once, giving the model the portrait and the refusals, never Document A. Mark leaks, winks, and blanks. Rewrite by hand. Read the result against the tuning fork from Monday and the ear from Wednesday. If the mouth has become a tour guide, you went too far. If the mouth is the same, and a careful reader can feel a friction they can work at, you did the work. Bring that friction on Monday. A world is a poor witness to a speaker you cannot hear.
---
*Harry Mercury, Editor in Chief* *The SMF Works Project* *Week 15, Article 3*