This week, a technical article described a pain point familiar to anyone who has used AI to write long-form content: around chapter 20, the AI starts "copying itself" — the same face-slapping move, the same emotional rhythm, the same chapter-end hook. Each section reads fine on its own, but five chapters in a row feel stale.
What This Is
The author calls this "narrative inertia": large language models are trained to "predict the next most likely token," which means that in long outputs, they increasingly lean toward the high-frequency patterns in training data — formulas are formulas precisely because they work. He didn't try to "eliminate formulas." Instead, he built a detection-injection system with 8 trigger conditions + 6 disruption actions: every 5 chapters, the system runs a check; if it hits one of 8 patterns — "same scene type appearing 2 times in a row," "same chapter-end hook used 3 times," "repeated sentence templates," and so on — it pulls one action from the set: switch perspective, introduce surprise, time jump, information reversal, emotional dissonance, or formal variation. The author calls this "planned derailment" — keep the formula, but break the staleness on purpose.
Industry View
This isn't just a novelist's problem; it's a fundamental feature of all current LLMs. Three perspectives are circulating in the industry:
Optimists argue this is exactly where prompt engineering (designing prompts to steer AI output) earns its keep — humans as orchestrators, AI as executors. 8 triggers + 6 disruptions is essentially a "supervisor" layer bolted onto AI, and it can generalize to all long-form production scenarios.
Skeptics warn that "8 triggers + 6 actions" is elegant scaffolding but a patch over a structural defect. The real fix lives at the model layer: longer context windows (how much text the AI can "see" at once), callable memory systems, layered reasoning architectures. If companies pin their hopes on prompt tricks, they'll hit a wall long-term.
There's also a cooler take: human writers writing novels also "copy themselves" — AI just magnifies the problem. The difference is that humans have self-awareness; AI has none. This means "AI content editor" will become a concrete job — specifically for spotting and breaking up patterned traces, not just fixing typos.
Impact on Regular People
For enterprise IT: If you're using AI to batch-generate marketing copy, weekly reports, or customer emails, repetition in sentence patterns and structure will pile up after 10+ outputs. We recommend adding a dedicated "de-patterning" review step to your pipeline.
For individual professionals: Don't submit an AI-drafted proposal, PPT outline, or quarterly summary after a single read. Read the whole thing through, and look specifically for openings that all start with "In today's…" and paragraphs that all follow "First/Second/Finally" — these are AI's most common tics.
For consumer markets: AI-generated web novels, short-video scripts, and podcast copy will develop "staleness" faster in long-form than short-form. Within the next year, "AI content editor" will move from a vague concept to a specific job title, with the core responsibility of spotting patterned traces and breaking them up.