# The tells that make copy read as machine-written

> A field guide to the habits readers recognise as machine-written: punctuation, stock words, sentence shapes, openers and e-mail habits, each described in plain terms with a rewrite.

Readers now spot generated text within a sentence or two, and once they do, they stop trusting the rest. The giveaways form a small set of habits. Once you can name them you can remove them, and the copy gets better in the process. This guide describes each habit in words and shows a plain rewrite.

## Why the tells matter

A prospect who decides a message was generated assumes nobody at your company read it before it went out, and weighs every claim in it accordingly. In a crowded inbox or feed, that judgement takes seconds.

None of these habits is wrong in itself. Writers used every one of them long before language models. The problem is density and placement: models reach for the same few devices in the same spots, and readers have learned the pattern. The examples below describe each tell instead of quoting it, because this page is checked by the same detector Demander runs on its own copy.

## Punctuation and symbols

- **The long dash.** The best-known tell is the long dash used as a pause, with or without spaces around it, and its cousin, the hyphen with a space on each side. Models use it to bolt one clause on to another. Rewrite with a full stop, a comma, a colon or brackets: “We built it for finance teams. They close the month in three days.” For ranges, write “1 to 1.5 percent”.
- **Arrows as connectors.** An arrow between a call to action and a link, or between a problem and its fix, reads like a template. Write the verb: “Book a time on the calendar below.”
- **Decorative emoji.** Rockets, sparkles, light bulbs, flames, pointing hands and green ticks at the start of lines mark a post as generated in most feeds. Remove them. If a list needs markers, make it a real list.

## Stock words

Some words turn up far more often in generated marketing copy than in anything people write. Each has a plainer word that says more.

- **Lock-and-key verbs.** The verb for opening a lock, applied to growth, potential or value, and its relatives about releasing or raising things. Say what actually happens: “You see which campaigns create pipeline within the first week.”
- **Inflated verbs and adjectives.** The finance word for borrowing, used to mean use; the verb from horse tack; the tailoring word for having no visible joins; the old verb for digging. The plain words are use, simple and look at. Demander swaps several of these automatically, because the plain word never changes the meaning.
- **Hype.** Words that claim a revolution, a first or a new class of anything, and phrases about changing the game. Replace them with the fact that would justify them, or drop them: “Plans start at €249 a month and include the first €5,000 of managed ad spend.”
- **Scenery words.** Landscapes that shift, worlds that move fast, journeys, realms and tapestries of anything. Name the actual change: “Since June 2026, accounts that start sending offline conversions to Google have to use the Data Manager API.”
- **Filler.** Adverbs that intensify nothing (the ones meaning really, very, in essence or in the end) and phrases that announce a point instead of making it, such as telling the reader something deserves a note. Delete them and check the sentence still stands. It usually reads better.

## Sentence shapes

These are harder to spot because every sentence is grammatical. The shape is the giveaway.

- **Denial, then correction.** A sentence that rejects a modest description only to replace it with a grander one, so the product is first said to be one thing and then, after a comma or a full stop, something bigger. State the positive claim and support it: “Demander plans campaigns across eight ad platforms and moves budget toward pipeline.”
- **The two-part slogan.** A short clause, a comma and a negated alternative closing the sentence, built to be quoted. It sounds sure of itself and says little. Write what the product does and for whom.
- **The question and its answer.** A one-word question about the outcome or the best part, answered in the next sentence. It delays the point for effect. Put the point first: “The report shows cost per opportunity for every campaign.”
- **Groups of three.** Three one-word sentences in a row, or three rhyming comparatives in a list. They have rhythm and no content. Pick the one that matters and prove it: “Setup takes one snippet and a CRM connection.”
- **Announcing the explanation.** Telling the reader that the reason is coming, or that the product enters the story at this point, instead of giving the reason. Start with the reason itself.
- **Scene-setting openers.** Opening by asking the reader to picture a scene, by addressing them by job title and telling them what they already know, or by naming two kinds of reader they might be. Open with the specific problem: “Your ad platforms see form fills and never hear which leads became opportunities.”
- **The closing question.** Ending with a question about whether the reader is ready for something. Ask for the next step instead: “Paste your URL to see the plan.”

## E-mail and social habits

- **Greeting filler.** Hoping the message finds the reader well, or that they are doing well, before saying anything. Start with the reason you are writing: “I saw your team is hiring two SDRs in Prague.”
- **Announcing the contact.** Saying that you wanted to get in touch, or that you are following up again with nothing new to add. Give the new information or ask the question: “Did the pricing page answer your question about seats?”
- **Fake intimacy.** Labelling a message as a short question, offering to pick someone’s brain, or adding polite adverbs that belong in a form letter. Ask plainly: “Who owns paid search at your company?”
- **Engagement bait.** Ending a post with a one-word request for opinions or agreement, asking readers to name a colleague in the comments, or promising more posts to those who subscribe. Finish on the point, or on a real question someone could answer in a sentence.

## Fixing it in practice

Read the copy aloud. Most tells survive silent reading and fail when spoken: nobody pronounces a long dash, and nobody talks to a colleague in slogans.

Replace every adjective you can with a number, a name or a fact. Specifics are the hardest thing for a model to invent and the easiest for a reader to trust, provided they are true.

Cut the first sentence and see whether anything is lost. Generated drafts often open with a warm-up line, and the point starts in the second sentence.

Vary sentence length. Machine prose settles into a steady run of medium sentences. A short one after a long one sounds like a person talking.

## Where Demander checks copy

The same rules run at every point where copy is made or leaves the product.

| Where | What happens |
| --- | --- |
| In the prompt | Every task that writes copy people will read ends with the same writing rules, as a list of banned punctuation, words and sentence shapes |
| After generation | Mechanical fixes are applied; anything left gets one retry with the list of problems, and if the retry still has tells, the deterministic writer is used instead |
| Lint and editor | Every field shows its tells as you type, including text rendered into images, slides, cards and video scenes |
| Before anything goes out | An e-mail step that still reads as generated is held with one review task per step; an ad cannot launch; an organic post fails with the reason; a bot reply goes to a person |
| Stored creatives | Existing creatives are checked again when the rules change, and those that need a rewrite go back for approval |

The mechanical fixes (dashes to commas or full stops, a short list of inflated words to the plain word, filler removed) never change meaning. Everything else is flagged for a person or the model to rewrite. Word and phrase matching covers English; the dash rules apply in every language.

> **What a detector cannot do.** A detector finds patterns. It cannot tell whether a sentence is true, or whether it is worth reading. Clean copy that says nothing is still empty; the checks only remove the habits that make readers stop early.

## Questions

**Is a long dash always a sign of AI?**

No. Writers used it long before language models. The trouble is frequency: generated text uses it far more than most people do, so readers now treat it as a signal. Commas, colons and full stops do the same jobs without the association.

**Will removing these patterns fool AI detectors?**

Detectors are unreliable, and fooling them is beside the point. What matters is that a reader keeps reading. Plain copy built on true specifics passes that test, whoever drafted it.

**Does this apply to languages other than English?**

The dash rules apply in every language. The word and phrase lists are English, so copy in Czech, German or Polish needs a read by a native speaker for the local equivalents.

**Why not let the model fix its own copy?**

Demander does, once. The model gets the list of problems and one retry. If the rewrite still has tells, Demander uses its deterministic writer instead of spending more on further attempts.

## Related

- [Human-sounding copy](https://demander.to/features/human-copy/)
- [Creative studio](https://demander.to/features/creative-studio/)
- [E-mail outreach](https://demander.to/features/email-outreach/)
- [Brand design](https://demander.to/features/brand-design/)
- [Buying committee mapping](https://demander.to/resources/buying-committee-mapping/)

---
Source: https://demander.to/resources/ai-tells-in-copy/ · Demander · reviewed 2026-10-02