Every AI Model
Has a Tell.
Each model is trained differently and leaves different patterns. ChatGPT hedges. Claude overuses em dashes. Gemini defaults to lists and headers. Generic humanizing misses them. HumanTone targets the specific signals each model leaves behind.
- 6 models. Pattern removal targeted to each source.
- Real semantic rewriting. No tricks.
- Reads like a person wrote it, not a model.
Humanizer
Source: ChatGPT draft for a B2B blog post
Tone: direct, no hollow affirmations
Avoid: "certainly", bullet defaults
Audience: marketing managers at mid-market
Why Source Matters
The Model That Generated It
Left Its Fingerprints.
Consistent Fingerprints
Each model's training produces the same patterns in every document it generates, regardless of the topic or the writer's intent. ChatGPT hedges in the same way on a product description as on a case study. Claude adds commentary about its own output on a blog post the same way it does on a business proposal. The patterns are reliable because they are structural, not topical.
Pattern Before Content
Editors and AI detectors respond to structure and register before they engage with the content itself. AI-generated text carries recognizable patterns in its opening moves, sentence rhythm, and vocabulary choices. These signals get read before the meaning does. Removing the pattern changes how the document is received before a single claim is evaluated.
Targeted Rewrites
Generic humanizing removes surface markers. Targeted rewriting removes the actual patterns. The difference is in what Custom Instructions specify: hollow affirmations for ChatGPT, em dashes for Claude, header defaults for Gemini. The right instruction set removes what the model actually does, not just what AI writing generally does. The result shows in how the copy reads to editors, clients, and a professional audience.
All Sources
Pick the Model.
Target the Right Patterns.
Each page covers the four patterns specific to that model, with Custom Instructions examples.
Humanize ChatGPT Text
Competent drafts with consistent tells. Hollow affirmations, formulaic openers, and the fingerprint of a helpful assistant trained to please.
Hollow affirmations, formulaic openers, default bullet structure.
Humanize Claude Text
Fluent and carefully constructed, with the most distinctive signature patterns of any AI model currently in wide use.
Em dash overuse, meta-commentary, reflexive both-sides hedging.
Humanize Gemini Text
Carefully structured content that defaults to the same formatting choices regardless of what the piece actually needs.
Header defaults, repetitive summaries, corporate vocabulary patterns.
Humanize DeepSeek Text
Thorough analytical output structured for reasoning tasks, not for the reader receiving the finished document.
Academic preambles, numbered sections, chain-of-thought framing.
Humanize Grok Text
Lively and opinionated, with personality and register shifts that break the frame of professional copy.
Register inconsistencies, forced wit, unexpected tone shifts.
Humanize Perplexity Text
Comprehensive output that covers everything and attributes every claim, reading like a search result rather than a finished piece.
Encyclopedic scope, citation hedging, neutral reference tone.
Model by Model
How to Humanize ChatGPT, Claude,
and Every Other Model.
The question usually arrives as "how to humanize ChatGPT or Claude text", and the honest answer is that those are two different jobs. The process is the same for every model: identify the source, remove that model's specific patterns, then rewrite the register for your audience. What changes is the middle step, because each model gives itself away differently.
ChatGPT's tells sit on the surface. Hollow affirmations, formulaic openers, and bullet structure appear in the first lines, so cleaning ChatGPT text is mostly about openings and formatting. Claude's tells are structural. Em dashes work as sentence architecture, meta-commentary narrates the draft, and every claim gets hedged, so cleaning Claude text is mostly about sentence rhythm. A generic pass tuned for one leaves the other half done. That is why each model here has its own page and its own instruction set.
| Model | Signature patterns | What to remove first |
|---|---|---|
| ChatGPT | Hollow affirmations, formulaic openers, bullet defaults. | The opening moves and list structure. |
| Claude | Em dashes, meta-commentary, both-sides hedging. | The sentence rhythm built around em dashes. |
| Gemini | Header defaults, repetitive summaries, corporate vocabulary. | The formatting applied to prose content. |
| DeepSeek | Academic preambles, numbered sections, reasoning traces. | The scaffold that delays the point. |
| Grok | Register shifts, forced wit, casual asides. | The voice breaks in professional copy. |
| Perplexity | Encyclopedic tone, citation hedging, over-comprehensive scope. | The reference register and excess coverage. |
01
Identify the source.
If you know which model produced the draft, start from its page. If you inherited the text, read the patterns instead: bullet-heavy openings point to ChatGPT, dash-driven sentences point to Claude, headers on short prose point to Gemini.
02
Remove the model's own patterns.
Each source page lists the four patterns that give that model away and shows the Custom Instructions that target them. Removing the right patterns matters more than rewriting everything: the rest of the draft is usually fine.
03
Rewrite the register for your reader.
Pattern removal makes text clean. Register makes it yours. Set the tone, the terms to preserve, and the audience in Custom Instructions, and apply the same set to each draft from that source so every document sounds consistent.
Mixed drafts are common: a team drafts in ChatGPT, edits in Claude, and the result carries both fingerprints. In that case work from the dominant pattern set and name the leftovers from the other model in the same Custom Instructions. HumanTone removes both in one pass. The rewrite does not need to know the model's name, only the patterns you want gone.
Custom Instructions
Same Tool. Different Target
Per Model.
Custom Instructions define the patterns to remove for each source and the register to replace them with. Write them once per model. Apply them to every document from that source.
One instruction set per source. Saved per project. Reapplied automatically on every rewrite.
Custom Instructions
Source: ChatGPT draft for a B2B SaaS blog
Tone: Direct, conversational, no AI phrasing
Avoid: "Certainly", preambles, bullet defaults
Audience: Marketing managers at mid-market
Voice: Practitioner, not commentator
FAQ
Questions About
Rewriting by Source.
How HumanTone targets the patterns each model actually leaves behind.
Content That Sounds Human.
Published Faster.
Start with 1,000 free words. No credit card. See the result in seconds.