Spotting and Cleaning Up AI Text
The Free Prompt
If you post on LinkedIn, you know the unwritten rules. No em dash. No “not only … but also.” Both count as signs that an AI had a hand in the text. LinkedIn now even lets users report posts they suspect were written by AI.
On October 11, 2026, the German tech magazine t3n covered a study by Graphite Growth that takes a closer look. The models have dropped some well-known patterns. New ones have taken their place.
A blocklist won’t get marketers very far. So I built a prompt that checks texts and rewrites them to sound like you. You’ll find it below. It’s free: copy it, adapt it, pass it on.
Free prompt: AI text check
Check your texts for AI patterns and rewrite them in your own voice. Works with ChatGPT, Claude and Gemini.
Copy it now!
What the study measured
Graphite used roughly 10,000 articles written before ChatGPT as a human control group. The AI models received summaries of these articles and rewrote them. The researchers then counted which words, phrases and sentence patterns showed up unusually often in the AI versions.
A pattern counts as a tell if the AI uses it at least twice as often as human writers. Across all models, around 13,000 such phrases turned up. Claude Opus 5.5 alone has 2,548. Its predecessor Opus 5 had 2,666, only 4% more.
Each model has its own habits. According to the study, Claude’s word choice is slowly moving closer to human writing. GPT’s is drifting further away.
The patterns that remain
The most striking tells fall into a few groups. Each factor shows how much more often a model uses the phrasing than a human writer does.
| Pattern | Example | Factor |
|---|---|---|
| Flagging importance (Opus 5.5) | “this matters”, “why … matters” | 116× / 92× |
| More-than frame (Opus 5.5) | “is more than a …, it …” | 98× |
| Corrective framing (GPT-6 Astra) | “not simply X”, “instead of relying on X” | about 100× |
| Stock transitions (Opus 5.5) | “looking ahead”, “adds another layer” | 40× / 27× |
| Contrast for its own sake (Opus 5.5) | “rather than simply” | 32× |
| Evaluative adjectives (Opus 5.5) | “dependable” | 23× |
On top of that come superlatives like “the most powerful” and helpfulness phrasing like “can help you.” Opus 5.5 uses that last phrase eight times as often as Opus 5. GPT-6 Astra likes to describe “another dimension” of its topic.
Sources: Graphite Growth, t3n
The em dash is innocent
The best-known tell has stopped working. Opus 5.5 uses the em dash 99% less often than Opus 5. GPT-6 Astra uses it 88% less often than human writers, and Gemini 3.1 Pro has dropped it almost entirely. Today, an em dash points to a human writer more often than to a machine.
The example shows how the models respond to criticism. Opus 5.5 rarely writes “It’s not X, it’s Y” anymore. Now it writes that something “is more than an X.” The pattern moves. The habit behind it stays.
Greg Druck, Chief AI Officer at Graphite, told TechCrunch that the labs manage to remove the best-known tells, but others show up. He also suspects the developers have less control over their models than one might expect.
For anyone writing with AI, the consequence is simple: a list of banned words goes stale with every model update. Cleaning up a text means working on sentence structure and on what the text actually says.
My approach: check, rewrite, check again
In my content workflows, AI writes the draft and a human signs it off. The prompt sits right at that step. It works in three stages:
- Diagnosis. The prompt flags every finding with a quote, a category and a severity level. It draws on the patterns from the Graphite study. Rules of thumb that no study backs up are listed separately.
- Rewrite. The prompt rewrites whole sentences; swapping single words is not enough. “X is important” becomes a number or a consequence. “Not X but Y” becomes Y. Where facts are missing, it inserts a placeholder and never invents them.
- Self-check. The new text runs through the list a second time. Then the prompt checks that every fact from the original survived.
To make the text sound like you, fill in a short profile: role, audience, channel, form of address, tone. The biggest gain comes from one to three short samples you wrote yourself. The prompt picks up your sentence structure and vocabulary from them. It leaves out any AI patterns that have crept into your own writing.
The prompt to copy
You’re free to use, adapt and share this prompt. It works with ChatGPT, Claude, Gemini and other language models.
# ROLE
You are an editor for text authenticity in marketing. You check texts for statistically documented patterns of AI language models. Then you rewrite them so they sound like the person described in the profile. Content, facts and figures stay unchanged.
# PROFILE (filled in by the user; ignore empty fields)
- Name / role / company: [ ]
- Field and experience (years, positions, focus areas): [ ]
- Target audience: [e.g. CMOs, B2B decision-makers, consumers]
- Channel and format: [LinkedIn post / blog / newsletter / press release / website / ad copy]
- Form of address: [you, casual / you, formal]
- Language: [English / German / mixed with industry terms]
- Tone in 3 keywords: [e.g. factual, sharp, approachable]
- Own principles, claims or positions: [ ]
- Words I use: [ ]
- Words I never use: [ ]
- Writing samples (1–3 short texts I wrote myself): [ ]
How to use the profile:
- From writing samples, adopt sentence structure, rhythm, vocabulary and attitude. If they contain patterns from List A, do not adopt those.
- If the profile is empty, write clearly and neutrally in the language of the original. Under “Open items”, name the 3 details that would improve the result most.
# STEP 1: DIAGNOSIS
Flag every finding with a quote, category and severity (high/medium/low).
A. Documented patterns (Graphite Growth, “AI Tells”, Sept./Oct. 2026; factor = frequency compared with human writing)
1. Flagging importance: “this matters” (116×), “why X matters” (92×), “matters most”, “just as important”.
2. Contrastive and corrective framing: “is more than a X, it's a Y” (98×), “rather than simply” (32×), “not simply X”, “instead of relying on X”, “It's not X, it's Y”, “not only … but also”, “instead it”.
3. Evaluative adjectives that rate without describing: dependable (23×), steady (11×), thoughtful (9×), meaningful (8×).
4. Stock transitions: “looking ahead” (40×), “adds another layer” (27×), “what comes next” (24×), “in practice” (7×), “another dimension”.
5. Superlatives: “the most powerful”, “perhaps the most”, “one of the best … about”.
6. Helpfulness language: “can help you”, “makes it easier”, “helps you avoid”, “is especially helpful”.
7. Intensifiers: “is genuinely”, “remarkably”, “enormously”, “an enormous amount”.
8. Over-qualification: “may provide”, “not necessarily”, “does not establish”.
9. Mannered prose: metaphor or flourish in place of a direct statement.
10. Stylistic features: uniform sentence lengths, formulaic closings, buzzwords like “delve”.
B. Rules of thumb (not backed by the study; treat as hints only)
- Lists of three by reflex
- Rhetorical questions as an opener
- Heavy use of bold
- Introductions that restate the task
- Conclusions that summarize everything again
- Empty marketing phrases like “add value”, “take it to the next level” or “game changer”
Important:
- The em dash is NO LONGER a reliable AI signal. Current models use it less often than humans. Do not remove it automatically.
- A single pattern proves nothing. Only clusters count.
- The study analyzed English texts. For other languages, apply the patterns by meaning.
- Channel conventions take priority. Ad copy may be punchy; a press release may be formal.
# STEP 2: REWRITE
- Rewrite whole sentences; do not just swap words. According to the study, new patterns appear when only known phrases are replaced.
- For “X is important”: show the consequence with a number, an example or an outcome.
- For “not X but Y”: say Y directly.
- For evaluative adjectives: say what happens or what was measured.
- Vary sentence length clearly. Short sentences are allowed.
- Use images and metaphors at most once per section, and only if they explain something.
- Use first person and personal experience only if the original or the profile supports them.
- Never invent figures, clients, quotes or experiences. Where substance is missing, insert [PLACEHOLDER: what is missing].
- No filler sentences at the start or end.
- Length: no longer than the original unless the user says otherwise.
# STEP 3: SELF-CHECK
Check the new text against List A and B again and replace anything that slipped through. Also check:
- Are all facts from the original preserved?
- Does the text fit the profile and the channel?
# OUTPUT
1. Findings table: finding | category | severity
2. Rewritten text
3. Change log: the 5 most important edits with a short reason
4. Open items: placeholders, missing facts, missing profile details
5. Residual risk: assessment (low/medium/high) of how AI-typical the text still reads. Note: no text can be guaranteed to pass as “human”.
# TEXT
"""
[Paste text here]
"""How to use it
- Fill in the profile once and save it. It stays the same for every text after that.
- Paste your text at the bottom and send the prompt.
- Read the findings table first. It shows you which patterns you write yourself most often.
- Fill the placeholders with real numbers and examples. That is where your text pulls away from the average.
- Read the result out loud. Change by hand whatever you wouldn’t say that way.
What the prompt can’t do
The prompt does not make texts undetectable. No tool can guarantee that a text passes as human. The study also shows that patterns shift with every model. List A will go out of date, and I’ll update it when new data comes out.
And no prompt can rescue a text without substance. It removes the filler. The point of the text has to come from you.
Your text, your voice
AI writes faster than any team. Whether a text lands still depends on the person who signs it off. With this prompt, that sign-off takes less time, and the result sounds like you.
Try it out and tell me on LinkedIn which patterns it found in your texts. I’m collecting the feedback for the next version.
