A text sounds AI-written when it repeats the same constructions and says nothing only its author could have known. You fix it in two passes: one to take those constructions out, and one to put your own data, examples and opinions in. Running it through a “humanizer” is not enough, because what is missing is information.
Why does it matter if a text sounds AI-written?
Readers notice, and they trust it less. A Gartner survey of 1,006 US consumers, run between May and June 2026, found that 65% believe brands produce too much AI-generated content. For 57%, that content has made them less trusting of brand messaging, and 43% said they now rely more on real people when deciding what to buy.
And they notice more reliably than most people assume. In a study presented at ACL 2025, five people who frequently use language models for writing read 300 articles and, by majority vote, got only one wrong. They outperformed almost every automatic detector tested, even when the text had been through a paraphraser or a humanizer.
How can you tell a text was written with AI?
By accumulation. Every construction in this table also appears in good human writing, because that is where the models learned it. The signal is repetition: five of them in two paragraphs.
| Sign | What it sounds like | What to do |
|---|---|---|
| False contrast | “It's not just software: it's a partner.” | Say what it does: “The software warns you when an invoice is due.” |
| Threes out of habit | “Fast, simple and efficient.” | Keep the one attribute you can prove, with its figure. |
| Inflated importance | “It marks a turning point for the industry.” | Say what changed, and for whom. |
| Vague source | “Studies show…” | Name the study and link it, or cut the sentence. |
| Replacement verb | “It serves as a meeting space.” | Use “is”. |
| Trailing -ing clause | “It analyses the data, revealing patterns.” | Two verbs: “It analyses the data and shows the patterns.” |
| One-word question | “The result? More sales.” | State the result with its number. |
| Motivational close | “The future is here. Will you be part of it?” | Close with the next step. |
| Brochure vocabulary | “Comprehensive solutions to empower your business.” | Name the service and what it includes. |
The first six are documented. Five are on the list of signs maintained by Wikipedia's editors, and the sixth comes from a study published in PNAS in 2025: GPT-4o uses present participial clauses 5.3 times as often as humans, and nominalizations 2.1 times as often. Another study, in Science Advances, went through more than 15 million biomedical abstracts. In 2024 the word “delves” appeared 28 times more often than expected, and the authors estimate that at least 13.5% of that year's abstracts were processed with a language model.
A limit of this guide. Both studies and Wikipedia's list analyse English text. The last three rows of the table are our own additions, from editing.
Do AI detectors and humanizers work?
Not as proof. Wikipedia's guide asks its editors not to rely on detectors alone: they do better than chance, but they have non-trivial error rates and fail on paraphrased text or on models they have not seen. A high score does not prove a text is AI-written. A low one does not prove it is good either.
Humanizers change words and word order. They cannot add what the text does not have: your figure, your example, your price. The same guide warns its editors about this: the signs are symptoms of a problem, and fixing only the signs makes the text harder to detect without making it better.
How do you write with AI without it showing?
By changing the order of work: material first, then the draft, and editing last.
- Gather the material before opening the chat. Write down your own figure, the real example, the price and your opinion. If you have none of that, there is no text to write yet.
- Ask for the draft with that material and with rules. Tell the model to flag what is missing instead of making it up.
- First pass: take out. Look for the signs in the table and delete every sentence that could sit on a competitor's website.
- Second pass: put in. Every paragraph needs something only you could have known.
- Read it aloud and check every fact. Where you stumble, rewrite. Where there is a figure, open the source.
Prompt 1 · draft from your own material
You are going to write a draft of [type of piece] for [audience]. Use only the material below. If a fact is missing, write [MISSING: which fact] instead of making it up. Material: - What the reader should understand or do: [...] - Our own data and figures: [...] - A real example: [...] - Our opinion on the subject: [...] - A limit or a downside we are going to admit: [...] Writing rules: - Start with the answer or the most useful fact, no introduction. - Sentences of varied length and short paragraphs. - State things positively: no "it's not X, it's Y" and no "it's not about X, it's about Y". - Don't group things in threes out of habit. - No rhetorical questions and no motivational closing line. - Don't use: revolutionize, unlock, leverage, seamless, game-changer, in today's world, it's worth noting.
For the first pass, use the model as an editor before you touch the text. Any of them works: ChatGPT, Claude or Gemini.
Prompt 2 · review before editing
Review the text below as a demanding editor. Do not rewrite it yet.
1. Flag every sentence that could sit on a competitor's website without changing a word.
2. Flag and count these constructions: "it's not X, it's Y" contrasts; lists of three; rhetorical questions; a trailing "-ing" clause after a comma; importance with no evidence ("key", "crucial", "marks a milestone"); vague sources ("experts", "studies show").
3. List the facts that must be checked before publishing.
4. Tell me which concrete information is missing: figures, prices, examples, deadlines.
Return a table with three columns: sentence, problem, which fact you would ask for.
Text:
[paste the text]And if you want drafts that start closer to the way you write, extract the rules from your own texts and paste them at the top of every request.
Prompt 3 · the rules of your voice
Here are three texts I wrote myself. Describe the way I write in ten concrete, checkable rules: sentence length, how I open, how I close, which words I repeat, what I never do. Don't use adjectives like "friendly" or "professional": I want rules that someone who doesn't know me could follow. [text 1] [text 2] [text 3]
Before and after: an example
The example is made up, for an accounting firm. The first text is what a model returns when asked for “a paragraph for the home page” with no material and no rules.
Before. In today's fast-paced world, accounting isn't just an obligation: it's a strategic tool. Our firm offers comprehensive, personalised and efficient solutions, supporting every client on their growth journey. The result? Peace of mind to focus on what really matters.
After. We keep the books for 40 small businesses in Rosario. On the 5th of each month you get a one-page report with three numbers: what you invoiced, what you will pay in tax and how much cash is left. If the tax authority sends a notice, we answer it within 48 hours. We don't take companies with more than 50 employees.
- Out went the opening about today's world, the contrast, the three adjectives and the one-word question.
- In came four facts that can be checked: number of clients, city, report date and response time.
- A limit appeared. Saying who you don't serve is one of the most credible things a company can write.
- The second text cannot be copied onto another firm's website.
Does Google penalize AI-generated content?
Google's official guidance on generative AI content, updated on 1 October 2026, does not forbid using it. It says AI is useful for researching a topic and for adding structure to original content. The risk it names is generating many pages without adding value for users, which may violate its spam policy on scaled content abuse.
The same page asks for one concrete thing: manually fact-check and review all generated content before publishing, because models predict likely words and can be wrong. That review covers titles, meta descriptions, structured data and image alt text.
What makes a text sound written by a person?
Whatever cannot be written without having done the work.
- The fact only the doer has. How long it took, what it cost, which version.
- The price. Almost nobody publishes it, which is why it stands out.
- The admitted limit. What your product doesn't do, or the case where it isn't the right choice.
- The owned opinion. “We do it this way because…”
- Your own mistake. What went wrong the first time.
At Totem we work from a list of twelve constructions we don't publish, and we update it often. The signs change with each generation of models: Wikipedia's list already separates some of them by period, from 2023 to mid-2026.
Common mistakes when editing AI-written text
- Asking it to “sound more human”. The model tends to add filler and jokes, and the text gets worse.
- Running the text through a humanizer and publishing it unread.
- Fixing only words. Removing “empower” doesn't fix a paragraph that says nothing.
- Publishing figures the model gave you without opening the source.
- Using the same request for everything. You get the same structure in every piece.
Frequently asked questions
How do I stop ChatGPT from writing like AI?
Give it your own material before asking for the text, add writing rules to the request, and edit the result in two passes: one to remove repeated constructions and one to add your own data and examples. The three prompts in this guide cover those steps and also work in Claude and Gemini.
Are AI detectors reliable?
Not as proof. They beat chance, but they have non-trivial error rates and fail on paraphrased text. In a 2025 study, five readers who use these models frequently got 1 of 300 articles wrong and outperformed most detectors.
Is it wrong to use AI to write content?
No. Google's guidance considers it useful for research and for structuring original content. The problem starts when content is published unreviewed, or when many pages are generated that add nothing.
Is the em dash a sign of AI writing?
Not on its own. No single sign proves anything: Wikipedia's list notes that all of them also appear in human writing. What readers notice is several of them in a small space.
Do I have to disclose that I used AI?
It depends on the case and on each platform's rules. Google suggests explaining how content was created when that helps the reader, and in Merchant Center it requires AI-generated images and product data to be labelled.
If you want to work on this live, with your own texts, see the Design & content with AI program.
Sources
- Gartner — Marketing survey: consumers and AI-generated content (22 Sep 2026)
- Russell, Karpinska & Iyyer — People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025)
- Reinhart et al. — Do LLMs write like humans? Variation in grammatical and rhetorical styles (PNAS, 2025)
- Kobak et al. — Delving into LLM-assisted writing in biomedical publications through excess vocabulary (Science Advances, 2025)
- Wikipedia — Signs of AI writing
- Google Search Central — Guidance on using generative AI content on your website
- Google Search Central — Spam policies: scaled content abuse