Will AI Writing Ever Be Good? Notes for Working Writers
AI can produce clean sentences on command. It can also bury them in cotton. The "AI voice" that floods timelines today is smooth, polite, and strangely obsessed with airy metaphors. Read enough of it and the tics jump out: everything is hushed, spectral, liminal, humming. It's frictionless to skim and forgettable to finish.
Is that because models can't "experience the world"? Maybe a bit. But a simpler answer is incentives. Cheap literary flourishes impress skimmers. Safety training files off edges. Provider economics reward bland, broadly acceptable outputs over bold, polarizing ones. The result is mid by design.
The real reasons AI reads mid
- Cheap effects win. Sensory-metaphor mashups ("grief tastes like metal") scan as "literary" to casual readers. Models lean on them because they correlate with praise.
- Safety and scale flatten voice. RLHF and instruction tuning optimize for inoffensive, predictable text. Great writing isn't predictable.
- No business case for daring prose. Accuracy, coding, and enterprise features drive revenue. Risky style does not. So nobody budgets for it.
- Models act like industries, not authors. Think of an LLM as a peak-era magazine machine: high floor, consistent tone, front-of-book glibness on tap. True originality appears as a side effect, not a goal.
If there were enough demand (and money) for unforgettable prose, we'd see training runs and data curation aimed at it. Until then, you're getting the house style of "inoffensive competence."
What that means for your craft
You don't beat mid with bigger prompts. You beat it with inputs models can't fake and editorial constraints they can't dodge.
A practical workflow that escapes the default voice
- Collect lived detail. Keep a "detail bank" from fieldwork, interviews, observations, and your own stories. Paste these into context and demand they appear, verbatim where needed.
- Ban the vibe words. Give the model a stoplist: ghost, spectral, liminal, whisper, hum, echoes, shimmering, azure, tapestry, poignant, journey. Tell it to favor concrete nouns, actions, and numbers.
- Force constraints. Set style rules: no metaphors in paragraphs 1-3; one image only, tied to a physical object you've provided; 20-word average sentence length; max two adjectives per sentence.
- Use contrast on purpose. Ask for an A/B draft: one spicy, one safe. Steal the best lines from each and stitch your own pass.
- Persona with a spine. Specify stance: "Take a clear position that someone could disagree with. No hedging, no 'on the other hand.'" Great writing chooses.
- RAG, not vibes. Feed transcripts, notes, and reference docs. Force citations to your snippets. If it's not in your corpus, it doesn't go in the draft.
- Three human passes. 1) Voice pass: cut cliches, add specificity. 2) Structure pass: reorder for tension and payoff. 3) Line pass: verbs over adjectives; concrete over abstract.
- Spike test. Keep only lines you'd quote out loud. If a paragraph has no spike, it gets cut or rebuilt.
Prompts that actually help
- Opinionated brief: "Here's the take. Argue it. One core metaphor, grounded in this real object: [object]. No summaries of what 'some say.'"
- Report-first: "From these notes, extract five physical details. Open with them before any abstraction."
- Adversarial rewrite: "Rewrite this to make an informed critic wince because it's direct, specific, and cuts a sacred cow."
- De-sanitizer: "Remove hedges (might, could, perhaps). Replace with accountable claims tied to source snippets provided."
Editing checklist to purge mid
- Cut 20% without losing meaning.
- Replace every abstract noun with a thing you can point at or a step you can do.
- Make one sharp claim per section; prove it with a detail from your notes.
- Swap two "nice" lines for one you'd defend on stage.
What would make AI genuinely good at writing?
- Curated, high-friction data. Not web slurry. Edited corpora with clear voice, plus counterexamples that reward risk.
- Training targets for style, not just safety. Reward surprising-but-true lines, not just harmless ones.
- Willingness to trade breadth for taste. A model that's worse at math but unforgettable on the page. There's little market pull for that today.
Until those incentives flip, the ceiling comes from you. Your reporting, your taste, your cuts. Treat the model like a fast room of interns with good grammar and no lived memory. You're the editor-in-chief.
Further reading
- Stanford CRFM report on foundation models for a grounded view of how these systems are trained and evaluated.
- The New York Times Magazine regularly publishes smart criticism on culture and tech, worth scanning for style analysis.
Tools and training for working writers
If you're building a durable workflow around AI, focus on repeatable prompts, your own corpora, and measurable editing passes. These resources can help:
- AI tools for copywriting to pressure-test drafting and rewriting steps.
- Prompt engineering guides to codify your constraints and stoplists.
The bottom line
Great writing costs something-time, attention, taste, and sometimes money. AI doesn't remove that cost; it moves it. Use models for speed. Spend your savings on depth. That's how you beat mid and keep your name worth reading.
Your membership also unlocks:
