How to Write Social Posts With AI That Do Not Sound Like AI
The Tells Readers Notice, Where They Come From, and a Five-Minute Edit That Removes Them
Short answer
AI posts sound like AI because they lack specifics and repeat a few habits. The usual ones are inflated significance, stock vocabulary, "not just X but Y" contrasts and tidy lists of three. Fix it at both ends. Give the model a real detail, number or customer phrase to build on, then edit out the habits with a five-minute pass.
1. Why do AI-written posts sound like AI?
Because the model has nothing specific to say, so it says the most probable thing in the most probable shape. The tells are what a sentence looks like when it is built from averages instead of from a detail.
A person writing a post usually starts from something that happened: a customer said a thing, a number moved, a mistake was made. A model asked to write a post about productivity starts from nothing and reaches for the language that most often surrounds the topic. The result is grammatical, balanced and empty, and readers now recognize that emptiness on sight.
This is a different problem from brand voice. A post can follow every voice rule you have and still read as machine-written, because it has no specific content. The companion guide on keeping AI drafts on-brand covers voice rules; this one covers the tells and the missing specifics underneath them.
2. What are the tells that make writing sound AI-generated?
Inflated significance, promotional language, vague attribution, stock vocabulary, "not just X but Y" contrasts, lists of three, heavy bolding and emoji bullets, and a closing line that summarizes or asks for thoughts. Wikipedia editors maintain a detailed catalogue of these patterns.
Wikipedia editors keep a field guide to signs of AI writing, built from reviewing a large volume of machine-written text [s1]. It is written for encyclopedia articles, but almost every pattern it lists shows up in social posts. The table maps the main ones to what they look like on LinkedIn or Instagram.
| Tell | What it looks like in a post | The fix |
|---|---|---|
| Inflated significance | This launch marks a pivotal moment in the evolving landscape of team software | Say what changed and for whom, in one plain sentence |
| Promotional language | A seamless, powerful, game-changing experience | Replace each adjective with the fact behind it, or delete it |
| Vague attribution | Experts agree that, many teams are finding that | Name the source or the customer, or drop the claim |
| Stock vocabulary | Delve, underscore, pivotal, tapestry, landscape, elevate | Keep a deny list and search every draft for it |
| Negative parallelism | It is not just a tool, it is a partner | State the one thing it is, with evidence |
| Rule of three | Faster, smarter and more connected | Keep the one item that is true and specific |
| Formatting as decoration | Bold on every other phrase, emoji as bullet points, long dashes everywhere | Format only what a skimmer needs to find |
| Outline-like ending | In conclusion, the future is bright. What do you think? | End on the most specific sentence, or just stop |
3. How do I get AI drafts that do not sound like AI in the first place?
Give the model something only you know: a customer quote, a number, a mistake, a detail from this week. Build the post on that. A specific input is the single change that removes most of the tells before you edit.
- Start from a detail, not a topic. Instead of "write a post about onboarding", give the model the detail: a customer told us setup was done before lunch, when they had blocked out the whole week for it. A post built on a fact has somewhere to go that is not a cliché.
- Write the first line yourself. The opening sentence sets the register for everything after it. If you write it, plainly, the model tends to continue in that register instead of reaching for its default.
- Give it real posts as examples. Paste three posts you wrote and would publish again, and ask it to match their sentence length and openings. Models imitate examples far better than they follow adjectives such as authentic or conversational.
- Add a deny list to the prompt. Paste the words and constructions you never want, from the table above plus your own. Ask for a draft that contains none of them.
- Ask for less. Set a word limit below what feels natural. Most tells live in padding, and a tight limit forces the model to keep the specific part and drop the filler.
4. What does the difference look like in a real post?
The AI version talks about significance and benefits in general terms. The edited version names one specific thing that happened, with a number, and stops. Same product, same message, different reader reaction.
| Version | Text |
|---|---|
| AI default | In today's fast-paced world, scheduling is more than just a task, it is a strategic advantage. That is why we are thrilled to unveil our new team calendar: seamless, intuitive and powerful. Ready to elevate your workflow? Let us know your thoughts below! |
| Edited | A three-chair dental practice ran its front desk off a handful of shared spreadsheets. Last month it moved to our team calendar. The double bookings that used to land every week stopped within the first two weeks. The calendar is live for every plan today. |
Everything that changed in the edit came from the input, not from clever rewriting: a specific customer, a before state, a number and a date. The AI default had none of those, so it filled the space with significance and adjectives.
5. What is the five-minute edit that removes AI tells?
Five checks in order: delete the first sentence if it is a windup, search for your deny list, cut every list of three to its true item, replace adjectives with facts, and cut the last line if it summarizes or asks for thoughts.
- Delete the first sentence and read again. If nothing is lost, it was a windup. This removes most openers about the fast-paced world we live in.
- Search the draft for your deny list, including the long dash if your brand avoids it. Replace each hit or rewrite the sentence.
- Find every list of three. Keep the item that is specific and true; delete the other two.
- Circle every adjective. Each one either points to a fact you can state instead, or it goes.
- Read the last line. If it summarizes the post or invites the reader to share their thoughts, delete it and end on the sentence before.
Do not try to disguise AI drafting with detector-evasion tricks such as deliberate typos. Readers are reacting to emptiness, not to the tool, and the only reliable fix is having something specific to say.
6. How do I keep this working across a whole content calendar?
Turn your deny list, your example posts and your required specifics into standing instructions, and keep a running list of the edits you make. Every repeated edit becomes a new rule.
The first week you will make the same edits again and again. Write each one down. By the end of the month that list is your house style, and it belongs in the standing instructions of whatever tool drafts your posts.
In Oppira, those rules live in the Brand Kit and content policy lines that every Studio draft is written against, and each post is planned from a message pillar in the Playbook, which gives the draft a specific point to make before a word is written.
Key Takeaways
The tells come from missing specifics
A model with nothing specific to say fills the space with significance, adjectives and familiar shapes.
The patterns are catalogued
Inflated significance, stock vocabulary, "not just X but Y", lists of three and summary endings are the common ones.
Fix the input first
A customer quote, a number or a detail from this week removes most tells before you edit a word.
Write the first line yourself
The opening sets the register, and the model tends to continue in it.
Run the five-minute edit
Cut the windup, search the deny list, trim lists of three, replace adjectives with facts, drop the summary ending.
Repeated edits become rules
Log the edits you make every week and move them into standing instructions.
Frequently Asked Questions
Sources
- Wikipedia:Signs of AI writing Wikipedia, September 2026.Editor-maintained field guide to patterns in machine-written text. Written for encyclopedia articles; the patterns carry over to social copy.
- Prompting best practices Anthropic, September 2026.Vendor guidance on using examples and clear instructions.
Explore More
Related analyses, benchmarks, and industry insights
Related Guides
Glossary Terms
Turn reading into a reaction
Oppira watches your competitors, keeps your playbook current, and drafts the response. See your market clearly by tomorrow.