How to remove Markdown from ChatGPT, Claude, and Gemini output

Why asterisks appear when you paste

You ask an assistant for a summary, copy the tidy formatted answer, paste it into your CMS, and get this:

## Launch plan

**Priority:** Ship the [new page](https://example.com).

The chat window showed a heading and bold text. What you copied was the Markdown that produced them.

Assistants are trained to answer in Markdown because it is a compact way to express structure in plain text, and chat interfaces render it live. The rendering is a display layer. Underneath, the response is characters, and the clipboard gets the characters.

Paste into anything that also renders Markdown — a static site generator, Notion, a developer forum — and it looks right. Paste into an email client, a WYSIWYG CMS field, a Word document, or a plain-text form, and the syntax shows up literally.

What the symbols mean

Knowing which symbol does what makes it obvious which ones are safe to delete and which carry content:

• # through ###### at the start of a line are headings. Delete the hashes, keep the words.
• **text** and __text__ are bold; *text* and _text_ are italic. Delete the markers, keep the text.
• [label](https://url) is a link. Here you have a decision: keep the label, keep the URL, or keep both.
• `code` and fenced ``` blocks mark code. The content usually matters even when the fences do not.
• > at the start of a line is a blockquote. The marker goes, the quotation stays.
• - , * , or + at the start of a line is a bullet. Often worth keeping, since it is real structure rather than decoration.

Why find-and-replace goes wrong

The obvious approach is to search for ** and delete it. This breaks in ways that are easy to miss.

Deleting every asterisk destroys multiplication in code samples and any literal asterisk in the prose. Removing bracket characters mangles arrays and citations. Stripping # removes C# and hex colours along with the headings.

Position is what disambiguates. A # is a heading only at the start of a line. A - is a bullet only at the start of a line. Handling links properly means recognising the whole [label](url) shape and choosing what to keep, not deleting brackets wherever they appear.

The safest order is structural markers first (headings, quotes, bullets), then inline emphasis, then links — because link labels can themselves contain emphasis, and processing in the wrong order leaves orphaned symbols behind.

Stopping it at the source

You can ask for plain text up front. Be specific, because a short instruction tends to decay over a long conversation:

Reply in plain prose. No Markdown, no headings, no bullet points,
no bold, no tables.

This works reasonably well for one answer and less well across a long thread, where the model drifts back to its default formatting. For anything you paste repeatedly, cleaning the output afterwards is more reliable than hoping the instruction holds.

The other half of the problem

Markdown is the visible artefact. AI output usually carries a second, invisible one: em dashes where you would write a comma, curly quotes instead of straight ones, and occasionally zero-width characters.

Those survive a Markdown strip untouched, because they are not Markdown. If your text still looks subtly wrong after removing the syntax — quotes leaning the wrong way, dashes that break a CSV import — that is what you are seeing, and it needs a separate pass.