Every POV character in every AI-assisted draft I've read has the same superpower. They walk into a room, and within three seconds they know the other person is lying. Call it the generic narrator problem: the AI hands you a detective, a duchess, a seventeen-year-old barista, and a battle-hardened mercenary, and every single one of them reads people with the same flawless, unfelt certainty. "She sensed he was hiding something." "Something about his answer didn't sit right." "He was lying — she could always tell."
That's not tension. That's a plot summary of tension. The reader doesn't get to feel the suspicion being built because the narrator skips straight to the verdict. And because the AI defaults to this omniscient gut-instinct voice, the problem compounds fast across a full draft — you end up with five point-of-view characters who are all the same cold reader wearing different clothing.
The fix isn't to teach the AI how to detect lies better. It's to stop asking the AI to detect lies at all, and start asking it to render how this specific, biased, half-blind person tries to detect them — and gets it wrong in ways that matter. That's the method in this guide. If you're still getting oriented with the platform, the beginner handbook writing fiction with AI covers the basics before you go deep here.
Why AI writes the same lie-detection instinct into every POV character
The model isn't being lazy. It's being efficient, and its efficiency has a shape.
When an AI writes dialogue where one character is withholding something, the most statistically common way that gets narrated in fiction is the undifferentiated "she knew." Thousands of published novels do exactly this, so the model reproduces it. The narrator floats just above the scene, privy to the truth, and translates it into a tidy sentence. No lens. No friction.
The second reason is subtler: the AI doesn't actually know what your character would notice, because you probably haven't told it. If the character sheet says "Cora: 34, forensic accountant, guarded, sharp," the model has a list of adjectives, not an operating system. It knows Cora is sharp, so it makes Cora notice everything a sharp person would notice — which is to say, everything. That's the same instinct as everyone else's.
A character's voice isn't how they talk. It's what they filter the world through. AI can imitate the talking. You have to build the filter.
This matters most in genres where reading people is the action. If you're writing a mystery novel with AI or a thriller with AI, every interrogation is built on the detective's fallible reading of a suspect. If everyone reads everyone correctly, there's no mystery — just suspense about logistics. Same with romance, where the whole engine of a slow burn is two people misreading each other's signals, and a false tell is more valuable than a true one.
Build a per-character "truth rubric": what signals they trust, dismiss, and over-read
Here's the core artifact. For each POV character, you write a short rubric that answers three questions: what physical or verbal signals do they trust as evidence of a lie, what do they dismiss as meaningless, and what do they over-read — the signals they weight far too heavily because of their own history?
The rubric needs to be specific and slightly wrong. A good one reads like a personality flaw dressed as a method.
Picture a character — call her Cora — and you notice her certainty comes from contraband. Cora grew up with a father who lied about money, and the only tell she ever caught him on was the way his hand drifted toward his pocket when he was nervous. So she trusts hand movements. She distrusts tone of voice entirely, because she was told as a child that she read tone wrong and learned to ignore it. She over-reads stillness — a person who doesn't move while answering makes her suspicious, when the truth is he's just exhausted.
That's a rubric. It's three degrees of error built into a person. Now the AI has something to render.
The best place to keep these rubrics is your project's story bible that AI models actually follow — one block per POV character, written in plain, declarative sentences the model can parse. Don't bury it in prose paragraphs. Make it a table or a bulleted list with headers: Trusts / Dismisses / Over-reads. Then run the character consistency tool across your draft to see if the rendered behavior actually tracks the rubric, chapter by chapter.
A rubric that does work
- Trusts: hand movement, especially toward pockets or cuffs. Whether a person's answer arrives too fast. Numbers that don't match themselves.
- Dismisses: eye contact, tone, sweating, fidgeting. She believes these are nervous tells that mean nothing — she's been fooled by them.
- Over-reads: stillness. A person who holds perfectly still while talking reads to her as coiled, controlled, dangerous.
That third line is the prize. It's the error engine. Every time the AI writes a scene where a character is nervous and holds still out of, say, military training, Cora will read danger. Every time a person is truthful but fidgety, Cora will ignore it. You have created a fallible observer who will be wrong about the same things in the same way, which is exactly how real people are wrong.
If you're building a whole cast of these, sketch the rubrics in an AI book outline session first, before you draft, so the patterns are baked into the plan rather than patched in during revision.
Prompting the observer, not the liar: describe the tell through the wrong lens
Now the technical shift, and it's the one most writers miss. When you prompt AI for a lying scene, the instinct is to describe the liar's behavior — "Marcus is lying and here are his tells." Don't. That's giving the AI the omniscient view again. Instead, describe the scene entirely through the observer's rubric, and let the tell be whatever the rubric would notice, even if your rubric leads them astray.
Here's what that looks like in practice. Notice how the prompt never tells the AI what's actually true.
Write a 400-word scene. POV: Cora, forensic accountant, 34. She is interviewing Marcus about the missing invoices. He is lying about the second set, but do NOT write his tells as objective facts — write only what Cora registers through her rubric: she trusts hand movement and too-fast answers, she dismisses eye contact and fidgeting as meaningless, and she over-reads stillness as danger. Marcus happens to go very still (he does this when he's anxious). Show Cora misreading the stillness as controlled menace. She never states "he was lying." The reader should see the hand movement that DOES give him away and the stillness she obsesses over. End on Cora making a wrong accusation.
Why it works: you've forbidden the omniscient summary ("he was lying") and forced the render through a narrow lens. The tell that actually matters (the hands) is there, but the POV character is distracted by the wrong signal (the stillness). The AI now has a job it can actually do — filter behavior through a flawed instrument — instead of pronounce verdicts.
What to tweak: the last line matters enormously. If you don't specify that Cora lands on a wrong conclusion, the AI will often resolve the scene safely with the correct read. Force the error.
Same scene, two lenses
Run the identical interview through a second POV character and watch the AI produce something genuinely different. Say the other witness is your card sharp, Delia. Her rubric trusts tells nobody else sees — a quickening breath, a shift in shoulder angle — dismisses words entirely ("what people say is dressing"), and over-reads composure as evidence she's being played.
Write the same interview scene from Delia's POV, 350 words. Delia is a retired card sharp, 40. Marcus is lying about the second set of invoices. Render ONLY what Delia's rubric catches: the instant his breath changes pitch before he answers, the half-inch shift in his left shoulder. She dismisses his actual words entirely — the content of what he says is "dressing." She over-reads his calm as evidence he's a professional liar running a game on her. She should feel flattered-conflicted by how well he's playing. Do not state he is lying or that she knows it. End mid-suspicion.
Now you have two scenes from the same event and neither is generic. The mystery lives in the gap between Cora's reading (menace) and Delia's (admiration) — and neither is fully right. That gap is your plot.
This is where the platform pulls ahead of generic chat. With Entangled Text vs ChatGPT, the project holds your rubrics across scenes, so you're not re-pasting character sheets every time. If you're budgeting API runs across a novel, the guide to budgeting AI drafting for a full novel will keep these iteration loops from running away with your cost.
Plant one false signal the POV character misreads — and make it cost them later
A rubric alone gives you flavor. The false signal gives you plot. Pick one over-read from the rubric — for Cora, the stillness — and plant it early in a scene where the signal points the wrong way. A genuinely innocent character goes still. Cora decides he's dangerous. She acts on that decision: pulls back, withholds trust, accuses, loses an ally.
Then make it cost her. Not immediately. Let the truth surface three chapters later — the still man was terrified of her, not hiding something — and let the consequence be real. A lead goes cold. A relationship fractures. The reader feels the misreading land because they watched Cora build the wrong case on the right instincts.
This is a two-part prompt, and you run the halves far apart in the manuscript.
PART ONE (plant, ch. 4): Write a 250-word beat where Cora's over-read of stillness fires. The character is Priya, who is innocent of the theft but terrified of being accused because of an unrelated secret. Priya goes utterly still. Render Cora reading this as dangerous control. Have Cora make one concrete decision — she "accidentally" leaves a document out for Priya to see, to test her. Do not reveal Priya's secret.
PART TWO (payoff, ch. 11): Return to that planted test from ch. 4. Reveal that Priya's stillness was fear, and her secret had nothing to do with the theft. Cora realizes she burned a witness for the wrong reason. Show the cost as a lost lead — Priya has already gone to the other side. 300 words. Keep Cora's voice; do not let her become omniscient or suddenly right.
Why it works: the plant and the payoff are separated, so the AI has to honor continuity — the earlier test, the earlier stillness — when it writes the payoff. That's a concrete continuity anchor, not a vibe.
What to tweak: if the AI softens the cost (Priya forgives her, the lead comes back), push back hard. A false signal that costs nothing is decoration. The whole point is that Cora's fallibility has teeth. This is also where The Five-Pass Revision Order for AI-Assisted Novels earns its keep — the false-signal plant/payoff belongs in a structural pass, not a line pass, or the AI will quietly fix the cost away while you're polishing commas.
Audit pass: flag every "something was off" line and route it through the rubric
You will still have them. The generic summary lines are stubborn, especially from earlier drafts or AI passes you ran before you built rubrics. "Something was off." "She sensed he was hiding something." "There was more to his answer." They're the AI's comfort food.
Do a dedicated audit pass. Search the manuscript for the habitual phrases — sensed, felt, something about, knew he was, didn't sit right, was hiding — and for every hit, either delete it or route it through the rubric. The routing prompt is short and surgical.
Find every sentence in chapters 3–6 where the narration summarizes a character's suspicion ("she sensed," "something was off," "she knew he was hiding"). For each: identify which POV character is observing, pull their rubric (Trusts/Dismisses/Over-reads), and rewrite the sentence as a specific, physical render through that rubric — one concrete signal, or one over-read. Keep the sentence the same length. Do not add new plot. Output as a before/after list so I can approve each.
Why it works: it forces the AI to name the observing character before rewriting, which stops the drift back into omniscience. The before/after format lets you keep the ones you like and reject the ones that over-correct into purple prose.
What to tweak: constrain the chapter range. If you hand it the whole manuscript, the AI gets shallow and starts inventing signals that contradict earlier scenes. Log the results against the rubric and re-check with the character consistency checker so chapter 12's tells don't contradict chapter 2's.
Two supporting tools make this pass move faster. The Show Don't Tell Hints tool flags summary narration specifically, and the Sentence Rhythm Report catches the second-order problem: once you convert suspicion to concrete renders, the AI has a habit of writing them all at the same length, which makes a paranoid POV read like a metronome. Vary the beat.
The instinct to teach the AI to detect lies is the trap. Detection is omniscient. What you actually want is a specific person, with a specific history, misreading another person in a specific way. Build the rubric, prompt through the observer's lens, plant one costly false signal, and audit out the leftover "something was off" lines. Do that and the tension stops living in the narrator's verdict and starts living in the gap between what your POV character sees and what's actually true — which is where it always belonged.
One concrete next step: open your current draft, pick your protagonist, and write the three-line rubric right now — Trusts, Dismisses, Over-reads. Don't polish it. One line each, specific enough to be slightly wrong. That single block is the difference between a character who knows everything and a character the reader can worry about. When you're ready to carry the method across the whole cast and out the door, the KDP upload checklist after an AI-assisted first draft will catch what the audit pass misses before launch.