Strategy
How to Use AI to Write Posts and Still Sound Human
What the research says about AI-written posts, what each network's rules say, which jobs to hand to AI and which to keep, prompts that keep your voice, and a checklist for the edit.
Oleh Husiev15 min read
Use AI as an interviewer and an editor, not as the author. Bring the raw material yourself, let the model ask questions and point out weak lines, keep the final sentences yours, and before you post, make sure there's one detail in it that only you could know.
That's not a purity rule. It's where the evidence points. People are comfortable with AI the way most writers already use it, as a tool behind a person, and much less comfortable when the AI is the one doing the talking. And most of what makes a post sound generated can be fixed in a five-minute edit.
What the research says about AI-written posts
Comfort follows human involvement. In the Reuters Institute's 2025 survey across six countries, 62% of people were comfortable with news made entirely by humans and only 12% with news made entirely by AI. In between, the numbers climb with human control: 21% with a "human in the loop", 43% when a human leads with some AI help, and 55% when AI is only used for grammar editing. That's news, not social posts, but the gradient is the useful part: the more the person leads, the more people accept it.
People can't reliably spot AI text, but they distrust it when they suspect it. In a PNAS study of 4,600 participants, people couldn't tell AI-written self-descriptions from human ones, and leaned on misleading cues: first-person pronouns, contractions and talk of family read as "human", so text that used them passed. Yet across 13 experiments, people who disclosed using AI at work were trusted less than people who didn't. And in a study of messaging, partners believed to be using AI-suggested replies were rated as less cooperative, even though the replies themselves made conversations more positive.
AI help makes writing more alike. In a Science Advances study, writers who got AI story ideas produced stories rated more creative individually, but the AI-assisted stories were more similar to each other. Another study found AI autocomplete nudged Indian writers toward Western styles, and a third found that writing with an opinionated assistant shifted both what people wrote and what they later said they believed. For anyone building an audience, that's the risk in one line: each post looks fine, and the feed looks like everyone else's.
The tells are real at scale. One analysis of 15 million PubMed abstracts estimated that at least 13.5% of 2024 abstracts were processed with a language model, spotted through a surge in words like "delves" (28 times the expected rate), "underscores" and "showcasing". Readers notice that kind of vocabulary even when they can't name it.
None of this says "don't use AI". In a study of 758 BCG consultants, people using GPT-4 on tasks inside its abilities finished 12% more tasks, 25% faster, at higher quality. On a task outside those abilities, they were 19 percentage points less likely to get it right. Polishing a sentence is inside the frontier. Knowing what actually happened at your company last Tuesday is outside it.
What each network says about AI-written posts
There's a lot of folklore here. Here's what the networks have actually published:
| Network | AI-written text | Labels | What it acts on |
|---|---|---|---|
| Said in 2024 it doesn't track whether generative AI was used | Shows Content Credentials on images and video that carry them | Bots, automated comments, engagement pods; since July 2026, a "seems like AI slop" report option | |
| Threads (and Instagram, Facebook) | No rule on AI text | "AI info" labels on images, video and audio, not text | Missing labels on photorealistic AI video or realistic audio |
| X | No rule against AI-assisted posts from a person | None found for text | Bulk or deceptive activity and automated replies, under its automation rules |
| Bluesky | Its community guidelines don't mention AI | None | Spam and artificially manipulated reach |
| Mastodon | Set by each server; mastodon.social requires that "use of generative AI must be disclosed" | Up to you, in the post | Server rules; mastodon.social doesn't allow accounts that only post AI-generated content |
So on four of the five, the thing that gets punished is volume without value, whether a person or a model produced it. Google's search guidance draws the same line: it acts against content made "for the primary purpose of manipulating search rankings", "no matter how it's created". On Mastodon, check your server's rules before you post anything AI-written, because "disclosed" there isn't optional.
Which jobs to hand to AI, and which to keep
| Hand it over | Keep it |
|---|---|
| Asking you questions about an idea | The idea and what you think about it |
| Suggesting two or three structures | Every number, name and quote |
| Pointing out weak or unclear lines | The opening line |
| Fixing grammar and typos | Your replies to people |
| Cutting a post to fit a limit | The decision to post it |
| Turning one long post into a thread |
The left column is editing. The right column is the reason someone follows you rather than a brand account.
A workflow that keeps your voice
1. Start with your own material
Before you open a chat window, write down the raw ingredients in whatever shape they come: a voice memo, three bullet points, a screenshot of a metric. "Shipped CSV export Tuesday. Took 3 weeks, planned 1. Most of it was Excel's date handling. 40 people asked for it." That's enough. It's yours, and it has the specifics a model can't supply.
2. Let the model interview you
Ask the model to ask you questions instead of writing anything. It pulls out details you'd skip, and it keeps the substance coming from you. Research on co-writing suggests people feel more ownership of text the more of their own input goes into it.
I want to write a short post about the note below. Don't write the post.
Ask me five questions, one at a time, that would help me find the most
interesting angle and the details a reader would want.
Note: Shipped CSV export Tuesday. Took 3 weeks, planned 1. Most of it
was Excel's date handling. 40 people asked for it.3. Write the draft yourself, or give it your voice first
If you write the first draft, you're done with AI until the edit. If you want the model to draft, give it something to imitate.
Start by naming your own voice. Pull up the five posts from the last 90 days that got the most replies, saves and shares, read them out loud, and write down what you notice: how long your sentences run, whether you lead with the point or a story, whether you ask questions, which words you repeat ("look", "honestly", "here's the thing"). Those habits aren't flaws. They're what makes your posts recognisable.
Then put the examples and that description in the prompt. Anthropic's prompting guide calls examples "one of the most reliable ways" to steer tone and suggests three to five; Google's Gemini guide says to "always include few-shot examples". Describe how you write in positive terms, because models follow "do this" better than a list of bans.
Here are four posts I wrote. Match their length, rhythm and tone.
<example>[paste a post]</example>
<example>[paste a post]</example>
<example>[paste a post]</example>
<example>[paste a post]</example>
How I write: first person, short sentences, contractions, one idea per
post, a concrete detail in the first line, no hashtags.
Use only the facts in my notes below. If something you need is missing,
write [NEED FACT] instead of inventing it.
Notes: [your notes and your answers from step 2]The [NEED FACT] line matters. Anthropic's guidance on hallucinations is to let the model say it doesn't know and to strip out any claim it can't support, and even then, it says, those steps reduce made-up facts without eliminating them.
4. Ask for criticism, not a rewrite
Once there's a draft, ask the model to point at problems and leave the fixing to you. A rewrite replaces your sentences with its sentences. A critique keeps yours.
Here's my draft. Don't rewrite it. List:
1. Any line a reader might not understand
2. Any phrase that sounds generic or like marketing copy
3. Anything that sounds like a claim I'd need to back up
4. Where you'd cut if it had to be 20% shorterLinkedIn went the same way with its own tool this year: it replaced its "Enhance your post" writer with a proofreading tool that keeps the member's voice.
5. Run the checklist
The edit is where a post goes from generated to yours. The checklist is below; it takes about five minutes.
6. Fit it to each network last
Cutting a finished post to 280 characters for X or 300 for Bluesky is the most mechanical job in the whole process, and the one AI is best at, because the content is already decided. Do it after the edit, not before. Our character limits reference has every network's numbers.
A worked example
Here's how those steps play out on the CSV note from step 1. This is an example we wrote for this post, not a real launch.
A one-line prompt with no material ("Write a LinkedIn post about launching CSV export") gets you this kind of thing:
Excited to announce that CSV export is now live! 🚀 This powerful new feature empowers you to take control of your data like never before. Whether you're a small team or a growing enterprise, seamless exports are just a click away. Try it today!
There's nothing in it that's true of this launch in particular, and nothing a reader learns.
After the interview and the person's own draft:
CSV export is live. We planned a week. It took three.
Most of that went into dates: Excel reads 03/04 as March or April depending on your region, so the same file opened differently in Berlin and Boston. We now write ISO dates (2026-03-04), which every spreadsheet reads the same way.
40 of you asked for this. Thank you for being patient while we argued with Excel.
The model's contribution there was the questions ("Why did it take three times longer?" "What would someone opening the file notice?") and one critique note ("'argued with Excel' might confuse people; say what actually broke"). Every sentence and every fact came from the person.
The "only you" test
Every post should pass one question: could someone else have written this? If the answer is yes, it's too generic. Rewrite it until it has at least one detail only you could know. Those details come from three places:
- Your numbers. "We hit 200 signups this week" beats "Our product is gaining traction." Specific numbers show real experience and give people something to react to.
- Your mistakes. "I scheduled posts at 9am for three months because every blog said mornings are best. Then I checked my analytics." Admissions like that show you've done the work instead of reading about it.
- Your opinions. "Per-channel pricing is a bad deal for small creators. Here's what it cost me." An opinion from experience invites disagreement, which brings replies.
AI can write "Timing matters for social media success." It can't write "I tested posting at 7am against 3pm for three weeks on LinkedIn, and the afternoon posts got twice the comments." The specific detail is what makes it human.
A checklist before you post
- Add one detail only you could know. A number, a date, a name, a mistake. Something that proves you were there.
- Check every fact. Every number, date, name and quote against where it came from: your analytics, your notes, the original article. Google's guidance for AI-assisted content says to "manually factcheck and review all AI-generated content".
- Swap the long dashes. Language models join clauses with them far more than most people do. Replace each with a period or a comma. When we edited this blog, we found more than 150 of them across 15 posts.
- Cut the buzzwords. "Leverage", "streamline", "harness", "elevate", "game-changer", "landscape", "delve". Rewrite those sentences in plain words.
- Delete the run-up. If the post opens with "In this post", "It's no secret that" or "Excited to announce", delete the first sentence. The second one is usually the real opener.
- Pick one reader. "Whether you're a solopreneur or a Fortune 500 marketer" is a model hedging. Write to one person.
- Look for the patterns Wikipedia's editors list: "not just X, but Y", lists of three by reflex ("Simple. Fast. Powerful."), inflated words ("a testament to", "pivotal"), and sentences that end in "highlighting its importance". Their guide says these are signs, not proof, and that they shift with each model generation.
- Read it out loud. If a sentence sounds like a press release, say it the way you'd explain it to a friend, and write that down.
Here's what that does to a typical generated line:
Before: "In today's competitive social media landscape, it's essential to leverage data-driven insights to optimize your posting schedule and maximize engagement with your target audience."
After: "Check your analytics. Find the three hours your followers are most active. Put your best posts there. That's the whole strategy."
Replies are the last thing to automate
If there's one place to keep AI out entirely, it's replies and comments. They're where people decide whether a real person is on the other end, the messaging study above found suspicion of AI replies lowers how cooperative you seem, and automated comments are exactly what LinkedIn and X ban. Drafting a reply in your head and typing it takes a minute. That minute is most of what building an audience is. More on that in why replies grow your audience.
Should you say you used AI?
Legally, usually not. The EU AI Act's rule for AI-generated text, which applies from August 2, 2026, covers text published to inform the public on matters of public interest, and it doesn't apply when a person has reviewed the text and holds editorial responsibility for it. We explain what that means for social posts in do you need to label AI-written posts under the EU AI Act? Platforms other than some Mastodon servers don't require it for text either. And despite the rumours, AI text doesn't carry hidden characters that give it away; we tested that in invisible characters and watermarks in AI text.
Whether to disclose anyway is about your audience. Nielsen Norman Group's review of the research found disclosure "does not consistently produce a negative reaction", and suggests deciding case by case. A reasonable rule: if the AI shaped what you're saying, not just how it's punctuated, consider saying so. On Mastodon, follow your server's rule.
Where Shaflex fits
Shaflex's AI menu in the Composer does the editing jobs from the left column: Fix grammar, Improve writing, Rephrase, and Fit to platform limit, which cuts a finished post down to a network's limit. The AI assistant can also build a voice profile from your own posts, with the phrases you use and a list of words you never want suggested. Every result lands in your draft, and nothing goes out until you schedule it, so there's always a point where you read it first. Free accounts get three AI actions a day. The writing, the facts and the replies stay with you.
Sources
Research and platform rules checked on October 20, 2026; the mastodon.social rules checked again on November 16, 2026.
- Reuters Institute: Generative AI and News Report 2025
- Jakesch, Hancock and Naaman (2023): Human heuristics for AI-generated language are flawed, PNAS 120(11)
- Schilke and Reimann (2025): The transparency dilemma: how AI disclosure erodes trust, Organizational Behavior and Human Decision Processes 188
- Hohenstein et al. (2023): Artificial intelligence in communication impacts language and social relationships, Scientific Reports
- Doshi and Hauser (2024): Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances
- Agarwal, Naaman and Vashistha (2025): AI Suggestions Homogenize Writing Toward Western Styles, CHI '25
- Jakesch et al. (2023): Co-Writing with Opinionated Language Models Affects Users' Views, CHI '23
- Draxler et al. (2024): The AI Ghostwriter Effect, ACM TOCHI
- Kobak et al. (2025): Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances 11(27)
- Dell'Acqua et al. (2023): Navigating the Jagged Technological Frontier, Harvard Business School working paper
- The Register: LinkedIn says it doesn't track whether generative AI was used, November 28, 2024
- LinkedIn: What we're doing to support authentic content and conversations, March 13, 2026; Help: Prohibited software and extensions and Content Credentials
- Fortune: LinkedIn adds a "seems like AI slop" button, July 31, 2026
- Meta Help Center: AI labels
- X Help Center: Automation rules
- Bluesky: Community Guidelines
- mastodon.social: server rules; Mastodon docs: Profile and bot flag
- Google Search Central: Spam policies and Using generative AI content
- Anthropic: Prompting best practices and Reduce hallucinations
- Google: Gemini API prompt design strategies
- Wikipedia: Signs of AI writing (WikiProject AI Cleanup)
- Nielsen Norman Group: When should you disclose AI use?
- EU AI Act: Article 50
Questions readers ask
How can you tell if a social media post was written by AI?
Often you can't, reliably: in a PNAS study of 4,600 people, participants couldn't tell AI-written self-descriptions from human ones. The usual tells are long dashes joining clauses, buzzwords like 'leverage' and 'streamline', a hedge that speaks to everyone ('whether you're a solopreneur or a Fortune 500 marketer'), three-word endings ('Simple. Fast. Powerful.'), stalling openers ('It's no secret that...') and fake excitement ('Thrilled to announce...'). They're signs, not proof.
Does LinkedIn penalise AI-written posts?
Not for being AI-assisted, as far as anything LinkedIn has published says. In 2024 it told The Register it doesn't track whether generative AI was used. What it does ban is automation: bots, automated comments and engagement pods. Since July 2026 members can also report a post as 'seems like AI slop', so low-effort generic posts carry a risk.
Do social networks require you to label AI-written posts?
Mostly not for text. Meta and LinkedIn label AI images, video and audio, mostly from metadata the file carries, and we found no rule on Meta, X or LinkedIn that requires labelling AI-written text. Bluesky's community guidelines don't mention it. Mastodon is the exception: rules are set per server, and mastodon.social requires that 'use of generative AI must be disclosed' and doesn't allow accounts that only post AI-generated content.
What should you never hand to AI?
The facts, the opinion and the replies. Models make up numbers and quotes with confidence, so every figure has to come from you or a source you've checked. Your opinion is the reason people follow you. And replies are where suspicion costs the most: in one study, people rated a partner as less cooperative when they believed that partner was using AI-suggested replies.
What's the best prompt to keep my own voice?
Paste three to five of your own posts as examples, describe how you write in positive terms (short sentences, first person, contractions), give it your notes as the only source of facts, and ask it to mark anything missing as [NEED FACT] instead of inventing it. Anthropic and Google both recommend examples as the most reliable way to steer tone.
Do I have to say a post was written with AI under the EU AI Act?
Usually not. The AI Act's text rule, which applies from August 2, 2026, covers AI-generated text published to inform the public on matters of public interest, and exempts text a person has reviewed and taken editorial responsibility for. Server rules on Mastodon can still require disclosure, and whether to disclose anyway is a judgement call about your audience.

