Writing Fiction with AI: How to Remove the "Machine Voice" Through Layered Generation

Anyone who has used AI to assist with writing fiction has likely encountered a specific type of draft: the plot is clearly laid out, but the moment characters open their mouths, the text becomes as dry as boiled chicken breast. The dialogue is unnaturally neat, emotional transitions are jarring, and even environmental descriptions feel templated. This article skips the metaphysics and focuses on a reproducible, engineering-focused workflow: first diagnosing the sources of the “machine voice,” and then systematically dismantling it through a pipeline of “Problem Checklist -> Voice Profile -> Script Layer -> Specialized Revision -> Model Division.”

1. Diagnosis: What Exactly is the “Machine Voice”?

Before making changes, we must translate the vague feeling of “something is off” into specific symptoms. There are four common categories:

Symptom 1: Dialogue reads like meeting minutes. One person reads out two or three pieces of information in order, the next person confirms them sequentially, and a third summarizes. Every sentence serves a purpose individually, but together they sound like a work report—information is exchanged, but actual communication between characters never occurs.

Symptom 2: Characters act like they attended the same onboarding training. Even if five characters are given different backgrounds, personalities, and experiences, they still speak with the same sentence structures. Roles like observer, technician, and medic are clearly divided, but if you cover their names, it sounds like one person taking turns to speak.

Symptom 3: Emotional transitions lack build-up. A character might be joking in one sentence and suddenly furious in the next, with no intervening action, pause, or psychological preparation. The emotion feels like a硬-wired switch being flipped.

Symptom 4: Templated environmental descriptions. Regardless of whether the plot is tense or relaxed, the scenery is always described in the same “standard prose tone,” serving no narrative purpose.

Behind these symptoms lie three root causes:

1. Statistical Averaging. Language models inherently predict the “next most probable token.” Training and alignment processes push expressions toward safe, fluent, and high-frequency averages. However, the “human touch” comes from deviating from this average—fragmented speech, repetition, interrupting, and answering off-topic.
2. Single-Pass Generation with Missing Intermediate Structures. Asking a model to jump from an outline directly to a final draft forces it to simultaneously solve plot arrangement, dialogue design, and pacing control. Attention is diluted, and dialogue—the easiest element to expose AI limitations—defaults to a “report style.”
3. Insufficient Constraint Granularity. “Write more naturally” is an unexecutable prompt. The model can only respond to this high-frequency expectation using high-frequency words.

2. Step 1: Translate “Unnaturalness” into an Executable Problem Checklist

Before revising, break down vague feelings into checkable items. Taking the “meeting minutes feel” as an example, you can break it down into:

• Are there more than three consecutive sentences following a “report -> confirm -> summarize” structure?
• Are sentences chopped too finely, requiring a “he said” every couple of words?
• Is the same event being confirmed back and forth?
• Is the character responding to the other person’s attitude, or merely processing their information?

Handing this checklist to the model for self-review, followed by targeted revisions, is far more effective than repeatedly yelling “be more natural.” Consider this comparison:

Before Revision (Report Style):
“Reporting to the captain, enemy situation found 300 meters ahead.”
“Copy. Please confirm enemy numbers and equipment.”
“Visual estimate: about 20 people, equipped with heavy weapons.”
“Understood. Notify all squads to move into combat positions.”

After Revision (Communication Style):
“Captain, movement up ahead.” Lao Zhou kept his voice low, pointing forward.
“How many?”
“Can’t tell clearly. About twenty, and they’re heavily armed.”
“Don’t panic.” The captain crushed his cigarette butt against the floor of the vehicle. “Everyone to their positions.”

Both passages contain the exact same information. The difference lies in “who” says it and in “what state.” The latter features varying sentence lengths, omitted subjects, two-word replies, and interspersed actions. This is how real people talk.

3. Step 2: Build a “Voice Profile” for Each Character

The root cause of the “onboarding training feel” is that the model only memorizes character function tags without grasping their linguistic fingerprints. The solution is to create a voice profile for each character during the setup phase and keep it visible in the context during generation. The profile doesn’t need to be long—four or five lines will do—but four dimensions are indispensable:

• Background and Profession: Determines the vocabulary pool. Someone who farmed the land and someone who attended a modern academy will not describe “dawn breaking” the same way.
• Personality and Current State: Determines sentence length and pauses. Impatient people use short sentences; cautious people tend to say “let me think about it again.”
• Speech Habits: Catchphrases, omission patterns, and fixed terms for the same objects.
• Relationship with Present Characters: Determines forms of address and tone. A “copy that” meant for a superior and a “gotcha” meant for a brother cannot be mixed.

The acceptance criteria is simple: cover the name, and can the reader guess who is speaking? If not, the profile hasn’t taken effect, and it needs to go back to the drawing board.

4. Step 3: Introduce a “Script Layer”—Build the Bones Before Adding the Flesh

This is the most rewarding step in the entire workflow. Do not let the model write the prose directly from the outline. Instead, generate an intermediate script layer—containing only dialogue and key actions, with no scene embellishments—and then expand the prose using the script as a skeleton.

An example format for the script layer is as follows:

[Scene] Military train carriage, night, third carriage catches fire after an ambush.

• Lao Zhou: (voice lowered, pointing forward) Movement up ahead.
• Captain: (raises hand to signal silence, observes for three seconds) How many?
• Lao Zhou: (still panting) Can’t tell clearly. About twenty, heavily armed.
• Medic: (tugging the captain’s clothes, voice trembling) There are still people trapped in carriage No. 3!
• Captain: (turns around, speaking faster) Lao Zhou, hold them off. Medic, follow me.

The script layer brings three direct benefits:

1. Controllable Dialogue Quality: Lines are exposed “naked” without any embellishment, making quality immediately apparent. The cost of revising one line of script is far lower than revising a whole paragraph.
2. Front-loaded Actions and Choreography: Characters know to shout for help when the carriage is sliding, rather than standing by delivering situation reports—because key actions are already specified in the script layer and won’t be distorted during expansion.
3. Increased Efficiency: On the surface, an extra step is added. In reality, because rework is drastically reduced, overall time consumption drops significantly. In my practice, the time to finalize prose was cut by about half, saving far more time than the script layer itself consumed.

5. Step 4: Specialized Dialogue Revision and the Marginal Utility of Rule Patching

Even with the script layer, it is recommended to schedule a separate revision pass exclusively for dialogue, rather than mixing it into general text polishing. Common rules for specialized revision include:

• Do not chop sentences too finely.
• Do not insert a “he said” every couple of words; use actions instead of dialogue tags.
• Do not repeatedly confirm the same event.
• Characters should respond to each other’s attitudes, not just information.
• Allow off-topic answers—real people often interrupt halfway through.

Another counter-intuitive insight: if your tool supports a randomness or temperature setting, do not assume “the higher, the more natural.” In A/B testing, lower randomness settings sometimes pulled the text further away from that weird, overly polished feel—sentences became more complete, and redundant explanations were removed. Of course, there are times it fails: after tweaking a few words, the way characters speak in a scene still doesn’t change. Removing a few repeated words doesn’t stop the “meeting” from continuing. This indicates the problem is structural, not lexical.

More importantly, recognize the marginal utility of rule patching: if you tweak prompts back and forth, adding rule after rule, but the draft keeps getting stuck at the same spot, it’s a signal that rules alone have hit their ceiling. Continuing to add rules yields diminishing returns approaching zero. The correct action is to change methods—introduce a script layer, or switch to a model better at organizing dialogue—rather than patching the old framework.

6. Multi-Model Division of Labor and Cost Control

Different models have different strengths: some excel at global planning, some at organizing dialogue, and some at stable long-text generation. Compressing the entire pipeline onto the most expensive, rate-limited model is neither economical nor necessary. A rational division of labor is:

• Planning and Review: Handled by the model with the strongest reasoning and planning capabilities for outline review and final draft diagnosis. Low usage, high precision.
• Script and Dialogue: Assigned to a model better at organizing communication. This is where the “machine voice” is heaviest and where spending your budget is most worthwhile.
• Prose Expansion: Handled by a cost-effective model with stable long-text generation, absorbing the bulk of the output volume.

In terms of engineering implementation, rather than manually copy-pasting between different sessions, it is better to solidify “which model is called at which stage” into a manageable pipeline. Using large model services provided by Baidu AI Cloud as an example, multi-model integration, invocation sequences, and quota allocation can be managed centrally, turning the process into configuration rather than craftsmanship. The trade-off principle when budgets are tight: prioritize the quality of the dialogue stage. Template-like prose can be saved by post-polishing, but once dialogue solidifies into a “report style,” it is almost impossible to rescue.

7. Acceptance Checklist and Human Final Review

After each draft, run through a naturalness checklist:

• Cover the names: can you tell who is speaking?
• Are there consecutive “report -> confirm -> summarize” structures?
• Are there actions or pauses paving the way for emotional transitions?
• Does the environmental description serve the current emotion and narrative?
• Is there at least one “imperfect” colloquial feature in the dialogue—omissions, interruptions, or off-topic answers?

The final step is always a human read-through. AI is responsible for solidifying structure, information, and pacing; the judgment of “whether it sounds like a real person” must still be made by a human. The point of the tool is not to write it for you, but to free you from mechanical labor so you can invest your energy into decisions only you can make.

Conclusion

The “machine voice” is not metaphysics; it is a set of engineering problems that can be named, dismantled, and processed item by item. The problem checklist translates feelings into actionable items, voice profiles solve “everyone sounds like the same person,” the script layer solves “disconnected dialogue and actions,” specialized revision smooths out detail frictions, and model division balances quality and cost. After stacking these steps, the draft produced by AI may not be ready for print directly, but at least when reading it—it no longer feels like an endless,冗长 meeting.