You ask a chatbot for a product description, it looks perfect in the chat window, and you paste it into your CMS. Suddenly the paragraph is littered with double asterisks around random phrases, a line starts with ###, the bullet points have turned into hyphens with uneven spacing, and the apostrophe in “don’t” shows up as a strange box on the published page. Then you read it again and notice the other problem: even with the symbols fixed, it sounds like a brochure written by a committee.
Those are two separate issues, and they need two separate fixes. The first is mechanical. It comes from the way chat tools format text behind the scenes, and a plain text conversion plus a few find-and-replace passes will sort it out. The second is about the words themselves, the phrasing patterns that make a reader’s eyes glaze over, and that takes either a careful edit or an ai humanizer built for rewriting.
Where the stray symbols come from
Most chat assistants write their answers in Markdown, a lightweight formatting language. When the chat window displays the answer, it reads the Markdown and renders it: **bold** becomes bold, ## Heading becomes a heading, a line starting with – becomes a bullet. The text you see is already formatted.
What gets copied depends on how you copy. Some chat interfaces copy the rendered version (rich text with formatting), some copy the raw Markdown, and some give you a “copy” button that hands over the raw source. Where you paste it matters too. A Word document will try to interpret rich text. A plain text field, a form, an email client set to plain text, or a CMS with its own editor might keep the raw symbols, drop them, or convert them into something else entirely.
So the same answer can arrive as:
- Raw Markdown, with visible **, ##, – and backticks.
- Rich text that carries hidden styling (fonts, background colors, line heights) from the chat window.
- A mix, where some formatting survives and some turns into symbols.
The usual suspects
| What you see | What it was | Fix |
|---|---|---|
| **word** | Bold in Markdown | Remove the asterisks, or re-apply bold in your editor |
| *word* or _word_ | Italics in Markdown | Remove the markers |
| ### at line start | A heading | Delete the hashes and set a real heading style |
| Lines starting with – or * | Bullets | Convert to your editor’s list format |
| 1. lists restarting at 1 | Numbered list broken by blank lines | Rebuild as one list |
| Backticks around words | Inline code formatting | Remove unless it’s code |
| [text](url) | A Markdown link | Rebuild the link in your editor |
| Odd boxes or “’” | Curly quotes or apostrophes in the wrong encoding | Convert quotes to straight ones, or fix the encoding |
| Long dashes everywhere | Em-dash characters inserted by the AI | Replace with commas, colons or full stops |
| Huge gaps between paragraphs | Extra line breaks or invisible spacing | Collapse double breaks, strip hidden characters |
Cleaning the mechanical mess
The fastest route is to strip everything down to plain text first, then add formatting back deliberately. Trying to fix rich text in place is slower, because hidden styles keep reappearing in the next paragraph you edit.
Step one is a plain text conversion. Paste the AI output into a plain text converter (or a basic text editor like Notepad or TextEdit in plain text mode). This removes fonts, colors, background highlights and invisible styling from the chat window. What’s left is the characters themselves, including any Markdown symbols.
Step two is removing Markdown markers. A few find-and-replace passes handle most of it:
- Find ** and replace with nothing.
- Find ## and ### at the start of lines and replace with nothing (then set your headings properly in the destination).
- Find lines beginning – and decide whether they’ll become a proper bulleted list or a sentence.
- Find backticks and remove them unless you’re publishing code.
Step three is quotes and dashes. Chat tools often output curly (“smart”) quotation marks and apostrophes. They look fine in a modern browser, but in some older systems, CSV files, code, or databases with mismatched encoding they break into garbage characters. Converting them to straight quotes is safe almost everywhere. For dashes, decide on a house style. The long dash that AI text uses constantly is one of the giveaways readers have learned to spot, so I usually replace each one with a comma, a colon, parentheses or a new sentence, depending on what it was doing.
Step four is spacing. Look for double spaces after full stops, non-breaking spaces (they look like normal spaces but stop text from wrapping properly), and stacks of blank lines. A plain text converter that offers whitespace cleanup takes care of most of this.
A quick before and after
Here’s raw output as it might land in a plain text field:
Our Handmade Candles are crafted with care. ### Why Choose Us – Natural soy wax for a clean burn – Hand-poured in small batches – Long-lasting fragrance that fills the room
After converting to plain text, removing the markers and rebuilding the list:
Our handmade candles are poured in small batches from natural soy wax. They burn cleanly, and the fragrance lasts.
The second version is also shorter. That’s the cue that the problem goes beyond formatting.
The robotic phrasing problem
Once the symbols are gone, many people notice the text still doesn’t sound right. It’s grammatically fine. It’s also oddly uniform, a little too pleased with itself, and full of phrases nobody says out loud. Readers pick up on this faster than they used to, and some clients and teachers run text through AI detectors before they read it.
The patterns I look for first:
- Stock openers. A first line about how busy the modern world is, or “When it comes to,” or “In the realm of.”
- Inflated verbs. Grand, slightly theatrical verbs in places where “improve,” “open,” “look at” or “use” would do.
- The rule of three, every time. “Fast, reliable, and affordable.” “Clear, concise, and compelling.” Once is fine. Every paragraph is a pattern.
- Long dashes in every second sentence.
- Summary sentences at the end of each paragraph that restate what the paragraph just said.
- Even sentence length. Paragraph after paragraph of sentences between 15 and 25 words, with almost nothing very short or very long.
You can fix these by hand. Read the text out loud and you’ll hear most of them. Cut the opener, swap the inflated verbs for plain ones, break one long sentence into two, merge two short ones, and delete any closing line that only repeats the paragraph.
If you’re handling more than a few hundred words, or you’re cleaning text every day, a humanizing tool speeds this up. Quillbot’s version takes the pasted text and rewrites stiff phrasing into something more natural while keeping the meaning, with modes for standard, academic, professional and casual writing. It also gives a human score so you can see roughly how natural the result reads, and you can click individual words to choose alternatives. The free tier handles short texts, with paid plans for longer ones.
One paragraph, rewritten
Here’s a typical AI paragraph for a small bakery’s website, after the symbols have already been stripped:
Finding quality baked goods can be a challenge these days. Our bakery is dedicated to providing fresh, delicious, and wholesome treats that bring joy to every occasion. With a passion for baking and a commitment to excellence, we strive to exceed your expectations every single day.
Nothing in it is wrong. Nothing in it is specific either, and the reader learns almost nothing about the bakery. A rewrite that keeps the meaning but sounds like a person:
We bake bread and pastries every morning from 5 a.m., so what’s on the shelf at opening was made that day. Most people come in for the sourdough. The cinnamon buns sell out by about ten on Saturdays.
The second version is shorter, it uses details the business owner would have to supply (the 5 a.m. start, the sourdough, the Saturday sellout), and it drops the three-adjective list and the “commitment to excellence” line. A humanizer will get you part of the way toward that by smoothing rhythm and swapping stiff phrases. The specifics have to come from you, because no tool knows what time your ovens go on.
A cleanup order that saves time
The order you do things in makes a real difference. Humanizing text that’s still full of Markdown symbols wastes effort, because the rewriting tool may treat ** as content or break the markers in strange places, and you’ll be cleaning up twice.
The order I use:
- Convert to plain text. Strip all hidden formatting.
- Remove Markdown markers and fix quotes and dashes. Mechanical cleanup, done with find-and-replace.
- Rewrite the phrasing. By hand or with a humanizer, working on clean, symbol-free text.
- Read it out loud. Fix anything that still sounds off, and check that facts, names and numbers survived the rewrite.
- Add formatting in the destination. Headings, bold, lists and links, applied with your editor’s own tools so they render correctly.
Step four matters more than people expect. Any rewriting step, human or machine, can change a figure, soften a claim that needed to be firm, or swap a product name for a generic term. Always compare the final version against the original for facts.
Sample prompts that reduce the mess at the source
You can prevent a lot of cleanup by asking for the right output in the first place. These instructions work in most chat tools:
Write this in plain text with no Markdown. No asterisks, no hashes, no bullet symbols. Use straight quotes and no em-dashes.
Write in a plain, friendly tone. Vary sentence length. Don’t open with a general statement about the world. Don’t end paragraphs with a summary sentence.
Give me three short paragraphs, under 60 words each, with no headings.
Even with these instructions, some formatting habits creep back in, especially in longer answers. So the plain text pass is still worth doing, and the ai humanizer step is still useful for anything that reads stiff after you’ve removed the symbols.
Special cases
Emails. If your email client is set to plain text, Markdown symbols will show up literally to the recipient. If it’s set to HTML, pasted rich text can carry the chat window’s font and background color, so the email looks pasted. Convert to plain text first, then format in the email editor.
Spreadsheets and CSV files. Curly quotes and stray line breaks inside cells cause the most trouble here. A single line break inside a cell can split one row into two when the file is opened elsewhere. Clean line breaks and quotes before pasting.
Code and config files. Never paste rich text into code. Smart quotes in a JSON file or a script will break it, often with an error message that doesn’t point to the real cause.
Social posts. Most platforms don’t render Markdown at all, so **bold** appears exactly as typed. Some people use Unicode “bold” characters as a workaround; screen readers handle those badly, so I avoid them.
CMS editors. Some content management systems accept Markdown natively and some don’t. If yours does, you can keep the Markdown and skip the stripping step. If it doesn’t, plain text first.
Mistakes I see often
- Running a humanizer on text that still has Markdown symbols in it.
- Fixing the formatting but leaving every long dash in place.
- Pasting rich text directly into a website editor, which brings along invisible spans and inline styles that break the site’s design.
- Trusting the rewrite without checking facts.
- Using search-and-replace on single asterisks without checking for legitimate uses, like footnote markers or multiplication in a formula.
Quick answers
Why does my apostrophe turn into strange characters?
The text contains a curly apostrophe, and the system displaying it is reading the file with a different character encoding. Converting curly quotes and apostrophes to straight ones before pasting avoids it.
Will removing formatting change my text?
Plain text conversion removes styling and hidden characters, not words. Markdown symbols remain until you remove them, so do that step separately.
Do I need a humanizer if I edit carefully by hand?
No. For a short piece, a careful read-aloud edit does the job. A humanizer helps when the volume is high or the deadline is tight, and you still want to read the result before it goes out.