How to Write With AI Without Sounding Fake

By Austin Little

People can smell generic AI prose from the first sentence. The fix is not a magic prompt. It is a workflow: you bring the facts and the voice, the model helps with structure and cleanup, and you cut every phrase that sounds like a brochure. This is the operator method for writing with free and local AI without sounding fake.

AI disclosure. This page was drafted with AI assistance and edited for Apiary. We don't invent quotes, stats, people, or events. If something looks off, tell Austin — that's the point of a living hive.

Why "sounding fake" became the real complaint

A few years ago the complaint was "AI can't write." Now the complaint is sharper: AI writes too smoothly. Same rhythm. Same adjectives. Same polite emptiness. Readers bounce because the text feels like it could sit on any company's blog with the logo swapped.

Searchers type things like "how to write with AI without sounding fake" because they already tried the naive path — paste a topic, take the first draft, hit publish — and watched engagement die. Students get flagged. Operators lose trust. Grandparents get sold on "AI wrote my letter" products that produce the same syrupy tone as everyone else.

The good news: sounding human is mostly process, not model size. A small local model on Ollama plus a ruthless edit pass often beats a big cloud draft you never touched. If you want the free stack map first, read How to Use AI Without Paying. If you are setting this up for an older relative, pair it with AI for Grandparents: Free, Safe, and Actually Useful. For the Apiary-side meaning of tending knowledge with AI, see What Is an AI Beekeeper?.

What "fake" actually means on the page

"Fake" is not one failure. It is a pile of tells. Learn the tells and you can hunt them:

1. Brochure adjectives. Innovative. Seamless. Cutting-edge. Robust. Comprehensive. Elevate. Empower. Unlock. Nestled. Tapestry. Delve. These words are not illegal. They are usually empty.
2. False balance. "While there are pros and cons…" when the reader asked for a clear recommendation.
3. Invented specificity. Fake quotes, precise percentages, named studies that do not exist. This is the dangerous kind of fake — not just tone, but lies.
4. Symmetric structure. Three benefits, three drawbacks, three takeaways, every section the same length. Real writing is lumpy. AI loves wallpaper.
5. Voice swap. Your rough notes sound like you. The draft sounds like a press release. That gap is the problem.
6. Apology fluff. "In today's fast-paced world…" "It is important to note that…" Openers that delay the point.

Kill those and you are already ahead of most "AI content."

The dignity rule (before any prompt)

You are not trying to trick readers into thinking a robot is a person. You are trying to publish useful writing that still sounds like you — or like a careful editor who cares about the reader.

Dignity means:

• Do not invent facts to sound authoritative.
• Do not paste someone else's private notes into a random cloud chat.
• Do not force AI on people who asked for a human letter without saying so when it matters (school, legal, medical).
• Prefer free local tools for private drafts; use free cloud tiers only for non-sensitive work.

Apiary's line fits here: BYO-LLM — bring your own brain. The hive is the system. The voice is yours.

The rewrite loop that actually works

Forget "one prompt, perfect article." Use a loop:

Step 1 — Write the ugly human draft first

Open a notes app. Type for five to fifteen minutes without asking AI for help. Include:

• What you actually know
• What you are unsure about
• Names, dates, and numbers you can verify
• A sentence in your real speaking voice ("Here's the short version…")

If you skip this step, the model invents a personality for you. That personality is always the same: corporate helpful.

Step 2 — Ask the model for structure, not soul

Paste your notes into a local chat (Ollama) or a free-tier chat for non-private work. Ask for something narrow:

Turn these notes into an outline with H2s. Keep my claims. Flag anything that looks like a guess. Do not add statistics I did not provide. Do not invent quotes.

Notice what you did not ask: "Write a polished 2000-word SEO article in a friendly professional tone." That prompt is how you get fake.

Step 3 — Generate section by section

Ask for one section at a time. After each section, edit before you request the next. This keeps the model from drifting into wallpaper and keeps you in the loop.

Step 4 — Voice pass (the important one)

Read the draft out loud. Anywhere you would never say the sentence at a kitchen table, rewrite it. Cut the first paragraph if it is throat-clearing. Replace "individuals" with "people." Replace "utilize" with "use."

Step 5 — Truth pass

Highlight every number, proper noun, product claim, and "studies show." Verify or delete.

Step 6 — Compression pass

AI drafts run long. Cut 15–25% without mercy. Fake often lives in the extra sentences that say nothing.

Prompts that keep your voice (copy and adapt)

Use these as starting points on free local models. Adjust the job; keep the constraints.

Outline from notes

You are a clear writing helper, not a marketer. Here are my rough notes. Make an outline with short H2 titles. Keep my facts. List any claim that needs a source. Do not invent sources. Do not add a conclusion full of adjectives.

Rewrite for voice

Rewrite the following so it sounds like a practical operator talking to one reader. Short sentences. No hype words. Keep every fact. If something is unclear in my draft, ask a question instead of guessing.

Cut the brochure

Delete empty adjectives and filler openers. Keep meaning. Return the cleaned text only.

Uncertainty flag

Mark every sentence that looks like a guess with [GUESS]. Do not remove them yet — just mark.

One-job email

Rewrite this email so it is polite, under 120 words, and specific. Do not invent a meeting time or a discount I did not mention.

Patterns that waste tokens and produce fake tone:

• "Write like a New York Times bestselling author"
• Giant persona stacks copied from social media
• Threatening the model

Local vs free cloud for writing work

Details on free stacks live in How to Use AI Without Paying. Fast free inference options (Groq-style, Cerebras-style) are useful for bursts — see the sibling pieces on Groq limits and Cerebras worth-it — but they are still cloud. Treat them like emailing a stranger who writes well.

Hardware honesty for local drafting

You do not need a gaming PC to draft without sounding fake.

• 8GB RAM: small 3B–7B class models. Fine for rewrite and outline. Close browser tabs if the fan spins.
• 16GB RAM: 7B–14B is comfortable for section drafts.
• 32GB+ / strong Apple Silicon: you can go bigger.

If the model is slow, drop a size. A crisp small model you edit beats a huge model you rubber-stamp.

The kitchen-table test

Before you publish, ask:

1. Would I say this sentence to a neighbor without cringing?
2. Could this paragraph sit on a competitor's site unchanged?
3. Did I verify every concrete claim?
4. Did I leave one specific detail only I would know (a local constraint, a real workflow step, an honest limit)?
5. Did I cut the first paragraph if it was throat-clearing?

If you fail (2), rewrite. Specificity is the antidote to fake. Not fake specificity — earned specificity: your tools, your steps, your limits.

Common traps that make AI writing sound fake

• Publishing the first draft. The first draft is clay. Edit is the craft.
• SEO stuffing. Keyword in every H2 until it sounds like a robot inventory. Put the query in the title and one clear H2; write for humans after that.
• Synonym cycling. "Utilize," "leverage," "harness" in rotation. Pick "use."
• Fake empathy. "I understand how challenging this can be…" when the model understands nothing. Be useful instead.
• Uniform section length. Real articles have short sections and long ones. Let them.
• Invented case studies. "Sarah, a small business owner in Ohio…" unless Sarah is real and consented. Apiary rule: no invented people.
• Over-explaining obvious words. Readers are adults. Dignity, not lectures.

Editing checklist (print this)

Use this as a final gate:

• [ ] My ugly notes existed before the model draft
• [ ] No invented quotes, stats, studies, or people
• [ ] No brochure adjective pile
• [ ] Opening line states the point within two sentences
• [ ] At least one concrete workflow step a reader can do today
• [ ] Uncertainty is labeled, not smoothed over
• [ ] Private material stayed local
• [ ] I read the whole thing out loud once
• [ ] I cut at least one paragraph that only restated the previous one
• [ ] Links to related pages are real pages, not wishful ones

Voice banks: teach the model your phrases

One practical trick that does not require a paid "custom GPT":

1. Collect 10–20 sentences you actually wrote that sound like you.
2. Paste them as examples: "Match this voice. Do not copy the topics — copy the rhythm and plainness."
3. Keep the bank offline in a text file. Refresh it when your writing drifts.

Do not paste confidential client sentences into a cloud model as "style examples." Local is the right place for a voice bank.

Working with small models without sounding dumbed-down

Small local models are uneven. They are still excellent for:

• Tightening sentences
• Turning bullets into paragraphs you will rewrite
• Generating alternate subject lines
• Listing questions a skeptical reader would ask

They are weaker at:

• Hard multi-step reasoning
• Long-range consistency across a 4,000-word draft in one shot
• Knowing current prices and rate limits (they will invent confidently)

Do not ask the model to be your source of truth for live pricing.

A one-hour practice session (do this once)

Minutes 0–10: Install or open Ollama. Run a small model. Paste one paragraph you wrote last week. Ask for a tighter rewrite with no new facts.

Minutes 10–25: Write ugly notes for a real page you need (email, how-to, FAQ). No AI yet.

Minutes 25–40: Ask for an outline only. Edit the outline by hand.

Minutes 40–50: Generate one section. Voice-pass it. Truth-pass it.

Minutes 50–60: Save your best prompt constraints in a notes file titled voice-rules.txt.

That hour teaches the loop better than reading twenty "prompt engineering" threads.

For operators, trades, and side businesses

If you write quotes, follow-ups, and how-tos for real work:

• Draft the customer-facing email locally so job details never hit a random site.
• Let AI propose structure for a parts explainer — then verify every part number yourself.
• Never let the model invent a warranty claim, a price, or a regulatory statement.
• Keep a "human last line" habit: the last paragraph before send should be typed by you.

Bee-adjacent writing follows the same rule: AI can help organize inspection notes; treatments get verified with a mentor or extension source, not with the chat alone. See related bee craft in the hive when you need that lane.

When sounding a little "AI" is fine

Not every artifact needs literary soul:

• Internal checklists
• First-pass meeting cleanup from your notes
• Alt-text drafts you will edit
• Sorting a pile of bullets

Save your deepest voice pass for pages that carry trust: public articles, client proposals, anything with your name on it.

Joe-Google test

People actually type:

• "how to write with ai without sounding fake"
• "make AI writing sound human"
• "AI writing tips"
• "edit AI draft"
• "why does ChatGPT sound robotic"
• "keep my voice with AI"

They do not type "temperature sampling for lexical diversity" when they want a letter that sounds like them. Write for the first list.

Case study pattern: service page that sounds real

You operate in a real geography. You actually stock certain parts. You refuse certain jobs (for example, you will not quote a torsion spring replacement without seeing the door). Put those truths in the outline before AI touches anything.

Prompt shape:

"Using only the outline below, write a service page section. Do not invent cities. Do not invent certifications I did not list. Do not add testimonials. Short sentences. No 'seamless' or 'cutting-edge.'"

Then you add one paragraph only you could write: a mistake you learned from, a weather constraint, a parts lead time. That paragraph is the anti-fake serum.

Case study pattern: bee conservation note

Outline with species, nest type, seasonal timing, and a safety note about pesticides. Ask AI to smooth transitions. You add the yard-specific bit: where the hotel sits relative to the chicken run; what failed last spring. Readers trust the failure more than the polish.

Editing modes (pick one per pass)

Clarity pass: shorter words, fewer clauses. Honesty pass: remove unverifiable claims. Voice pass: match voicebank.md. Kindness pass: for customer email, firm without humiliation. SEO pass last: only after honesty. Stuffing keywords first is how you get fake.

Never ask the model to do all five at once. Multipass is how humans stay in charge.

The "corporate synonym" problem

Models love swapping simple words for fancy ones: use→utilize, help→facilitate, start→commence, buy→purchase (sometimes fine), need→require. On a craft site or a trade site, reverse those swaps unless the fancy word is precise jargon your audience expects (torsion, brood, mite load).

Humor and edge

AI humor skews dad-joke or HR-safe. If your brand uses dry sarcasm, write the funny line yourself. Ask AI only to keep the surrounding explanation clear. Forced model jokes are a top-five fake tell.

Citations and links

Ask AI to suggest types of sources (extension office, CISA, manufacturer manual). You fetch the URLs. Pasting "add citations" without sources is how fake references appear. Apiary drafts prefer fetched links with dates.

Comments from readers ("this sounds like ChatGPT")

Possible responses:

1. Thank them; ask which paragraph felt hollow; fix it.
2. Disclose AI assistance if you had not.
3. Do not argue about detectors.
4. Improve specifics.

Defensiveness reads faker than disclosure.

Freelance contracts and AI

Some clients forbid AI. Some require disclosure. Some do not care. Get it in writing. If you used local AI for grammar only, say so if asked. If you used AI for substance, say so if asked. Lying about process is how freelancers lose referrals.

School and kids

Students: follow academic integrity rules. Parents: help kids brainstorm questions, not submit model essays as their own. Teachers: design assignments that require personal observation AI cannot fake easily (yard logs, shop photos with reflection).

A 14-day practice plan

Day 1–2: Build voicebank. Day 3–4: Ban list. Day 5–7: Rewrite three old posts with the clerk workflow. Day 8–10: Email-only practice. Day 11–12: Read aloud edits with a friend. Day 13: Publish one short piece with disclosure if appropriate. Day 14: Note which prompts earned a permanent spot in prompts.txt.

What free models do well for style

Mid-size local models are surprisingly good at: shortening; changing reading level; removing fluff you point at; turning bullets into prose when bullets are rich.

They struggle with: sustained literary voice; culturally delicate humor; accurate niche law; staying humble without hedging into mush.

Choose jobs accordingly.

Watermarks and provenance (high level)

Some ecosystems experiment with watermarking or provenance signals for AI text. Treat public explanations as evolving. Do not claim a detector result as courtroom fact. Do not promise readers you can "remove watermarks."

The honesty of this very article

This page was drafted with AI assistance and shaped with constraints, ban-list instincts, and operator edits. If a paragraph still feels generic, that is a bug report — tell Austin. The standard is not "zero AI." The standard is "zero fake authority."

Extended examples of constraint prompts

Angry → firm: "Rewrite without insults, without sarcasm, without legal admissions, keeping the timeline of facts I listed." Too long → usable: "Cut by 40% while keeping numbers and proper nouns." Too stiff → neighborly: "More like spoken explanation to a smart adult in a driveway; no slang that ages in a year." Too salesy → straight: "Remove CTAs except the single last line I provided."

Save the winners.

Typography and formatting tells

AI loves neat H2 symmetry and identical bullet lengths. Break symmetry on purpose: one section with a short list; one with a story paragraph; one with a table. Human documents are slightly uneven. Perfect wallpaper looks manufactured.

When collaborators over-edit your voice into mush

Park a "voice owner" role. That person has final say on tone. AI drafts are suggestions. Committees without a voice owner produce fake-by-committee, which is worse than fake-by-model.

Sources / further reading

• Ollama: https://ollama.com (UNVERIFIED live feature list — recheck on publish day)
• Sibling drafts: How to Use AI Without Paying; AI for Grandparents: Free, Safe, and Actually Useful; What Is an AI Beekeeper?
• Groq / Cerebras free-path siblings in this batch when you need burst cloud speed for non-private drafting

Before-and-after patterns (steal the method, not the sentences)

You do not need invented case studies. You need contrast patterns you can apply to your own draft.

Pattern A — opener

• Fake-leaning: "In today's fast-paced digital landscape, effective communication is more important than ever."
• Human: "If your AI draft sounds like a brochure, readers bounce. Fix the process, not the adjectives."

Pattern B — advice

• Fake-leaning: "It is recommended that users leverage innovative solutions to optimize their workflow."
• Human: "Write ugly notes first. Ask the model for an outline. Edit like you mean it."

Pattern C — uncertainty

• Fake-leaning: "Studies show that most people prefer concise emails." (Which studies? Often none.)
• Human: "I have not measured this for your inbox. What I can say: short emails with one ask get answered more often in the shops I have watched."

Pattern D — product mention

• Fake-leaning: "Our seamless platform empowers teams to unlock productivity."
• Human: "Apiary is built around BYO-LLM: you bring the model (local or a free-tier key). The site does not become your AI landlord."

Practice by taking one paragraph of your own AI draft and forcing it through these four patterns. Do that ten times and your eye gets fast.

Genre-specific notes (same loop, different stakes)

Emails

One ask. One deadline if it is real. No invented availability. If the model adds "Looking forward to connecting!" and you hate that phrase, delete it forever and add it to a ban-list in your voice-rules.txt.

How-to articles

Steps must be doable. How-tos go fake when they describe a UI that shipped two renames ago.

FAQ pages

Answer the question in the first sentence. AI loves a three-sentence windup. Cut the windup.

Social captions

AI will offer hashtags and "engagement hooks." Most of them sound like a mall radio ad. Write the caption in your notes voice, then ask only for a shorter variant.

Proposals and quotes

Numbers are sacred. The model does not know your labor rate. Paste only the explanatory paragraphs, never the pricing table, if you use cloud at all — and prefer local.

Building a personal ban-list

Keep a plain text file of phrases you never want to publish. Start with:

• delve / tapestry / landscape (metaphorical)
• unlock / elevate / empower (empty verbs)
• robust / seamless / cutting-edge
• "In today's…" openers
• "It is important to note that"
• "At the end of the day"
• "game-changer"
• "dive deep" (unless you are literally talking about diving)

When you paste a draft into the model, add: "Remove any phrase from this ban-list: …" Then still search the document yourself. Models miss items.

Collaboration without losing the plot

If two people edit with AI:

1. Agree who owns facts.
2. Agree who owns voice.
3. Do not both "polish" the same paragraph with different models — that is how you get mush.
4. Keep a changelog note: what the model did vs what the human changed.

Treat those comments as part of craft, not as failure.

What detectors and "humanizer" tools actually do

Detector scores are noisy. They are not a morality meter and they are not a craft coach. Chasing a "100% human" score teaches you to write weird on purpose.

"Humanizer" sites often:

• Charge after a teaser
• Resubmit your text to another model
• Promise to beat school detectors

If your goal is integrity, you do not need them. If your goal is to break a school or employer rule, that is outside Apiary's help. Write better, disclose when required, and keep private text local.

A weekly maintenance habit

Once a week, fifteen minutes:

1. Open your last AI-assisted piece.
2. Highlight three sentences that still sound fake.
3. Rewrite them by hand.
4. Add any new banned phrases to voice-rules.txt.
5. Delete unused local models if disk is tight (ollama list is your friend).

Craft is repetition with attention. The model does not get better at sounding like you unless you keep correcting it.

FAQ

Can I just use a "humanize AI text" website? Most of those are another model remixing fluff. Learn the edit loop. Tools that claim to "bypass detectors" are a red flag for school and work policy problems — not a craft skill.

Will a bigger model sound less fake? Not automatically. Bigger models can be more fluent at emptiness. Your constraints and edits matter more.

Is it cheating to use AI for writing? Depends on the context. For public how-tos, disclose when it matters and own the facts. For school, follow the teacher's rules. For client work, follow the contract. Craft is accountability, not cosplay.

How do I stop the model from inventing sources? Forbid them in the prompt, then delete any that appear. Never publish a citation you did not open.

What if English is not my first language? AI can help with clarity. Keep your meaning. Ask: "Fix grammar. Do not change my claims. Do not add idioms I did not use." Then read it aloud anyway.

Should I tell readers I used AI? Apiary pages carry a disclosure block. For emails to a friend, usually no. For academic or regulated work, follow the rules that apply. When in doubt, honesty costs less than getting caught.

Does local AI still hallucinate? Yes. Local means private, not truthful. Truth is still your job.

How long should the voice pass take? For a long article, longer than you want and shorter than rewriting from scratch. If the voice pass takes forever, your prompt asked for too much polish too early — go back to notes + outline.