AI is generating more text than ever before. ChatGPT has over 200 million weekly active users. Claude, Gemini, Copilot, Perplexity, and dozens of other tools are producing content at a scale that was unimaginable three years ago. Reports, emails, project plans, meeting summaries, proposals, documentation. The output quality is often excellent.
But there is a dirty secret that nobody talks about: almost none of it arrives at its destination looking the way it should.
The content explosion meets the formatting wall
Every week, billions of AI-generated responses flow from chat interfaces into the real world. People copy from ChatGPT and paste into Google Docs. They take Claude's output and drop it into an email. They ask Gemini to write a status update and paste it into Slack.
And every time, something breaks.
Tables become rows of pipe characters. Headings show up as hash marks. Bold text appears as asterisks. Code blocks turn into backtick soup. Links display as bracket syntax instead of clickable hyperlinks.
The content is right. The formatting is wrong. And millions of people spend minutes (or longer) fixing it by hand, every single day.
Why AI outputs markdown
To understand the problem, you need to understand why every AI tool, without exception, outputs markdown.
Markdown is a lightweight formatting language that uses simple text symbols to represent structure. ## means heading, **text** means bold, | separates table cells, triple backticks fence code blocks. It was created in 2004 by John Gruber as a way to write formatted content in plain text.
AI tools use markdown for three compelling reasons:
Token efficiency. Markdown uses 40% or more fewer tokens than HTML for the same formatted content. ## Project Status is 4 tokens. <h2>Project Status</h2> is 9. When you are serving hundreds of millions of users, that difference translates directly into speed and cost.
Training data alignment. Large language models were trained on massive datasets that include enormous amounts of markdown: GitHub repositories, technical documentation, blog posts, README files, Stack Overflow answers. Markdown is the format LLMs know best. It is the natural language of structured text generation.
Human readability. Even without rendering, markdown is scannable. You can understand **deadline is Friday** without a formatting engine. This makes AI responses useful even in plain text contexts, and it makes the raw output easier to review and edit.
From the AI provider's perspective, markdown is the right choice. It is efficient, well-understood by the models, and readable. The problem is not with markdown itself.
The problem: markdown is not a document format
Markdown is an intermediate representation. It describes formatting intent, but it is not the formatting itself. Nobody sends markdown to their boss. Nobody emails a client a message full of hash marks and asterisks. Nobody pastes pipe characters into a presentation.
The world runs on Google Docs, Microsoft Word, Slack, email, OneNote, and plain text. These are the destinations. And none of them speak markdown natively.
This creates a gap in every AI workflow:
AI tool produces markdown. Destination app expects rich text, HTML, or its own proprietary format. The user is stuck in the middle.
Most people fill this gap with the simplest approach available: copy, paste, and manually reformat. Select the heading text, apply Heading 2 style. Find every instance of **text**, select it, press Cmd+B, delete the asterisks. Rebuild the table cell by cell. Remove the backticks from code snippets and switch to a monospace font.
It works. It also takes 3 to 10 minutes per response, depending on length and complexity. Multiply that by every AI response you use in a given week, and the time cost becomes significant.
The last mile problem
The logistics industry has a concept called "the last mile." Getting a package from a warehouse to a distribution center is efficient. Getting it from the distribution center to your front door is the hardest, most expensive part of the entire delivery chain. The infrastructure is optimized for bulk movement, not individual delivery.
AI content has the same problem.
The AI pipeline is optimized for generation. Models are faster, cheaper, and better than ever at producing high-quality text. The prompt goes in, the content comes out, and it is structured with headings, lists, tables, bold text, and code blocks. Beautifully organized. Ready to use.
Except it is not ready to use. It is ready to be converted. And that conversion, the last mile from markdown to destination, is where the entire workflow breaks down.
What breaks (and why)
The failures are consistent and predictable. Here is what happens when you paste AI output directly into common destinations:
Headings become plain text. Your carefully structured response with ## Executive Summary and ### Key Findings arrives as lines of text with hash marks in front of them. Google Docs has minimal markdown detection (bold and italic only, and even that is inconsistent). Word has none.
Tables become unreadable. A clean comparison table in ChatGPT becomes a mess of pipe characters and dashes in your document. No borders, no cells, no alignment. Just characters that look like a broken grid.
Bold and italic show as asterisks. Key terms wrapped in **double asterisks** display literally. Some apps catch bold formatting sometimes. Most do not. The result is unprofessional at best and confusing at worst.
Code blocks lose everything. The monospace font, the background color, the syntax highlighting, the indentation context. All gone. You get backtick characters mixed into regular text.
Links display as bracket syntax. Instead of clickable "see the documentation" hyperlinks, your reader sees [see the documentation](https://example.com/docs). Brackets, parentheses, and bare URLs.
Each destination breaks differently because each one handles formatting in its own way. Slack uses a proprietary format called mrkdwn (not markdown). Email clients only support inline CSS. OneNote has its own rendering quirks. Plain Text has no formatting at all.
A single "markdown to HTML" conversion is not enough. Each destination needs its own conversion.
The scale of the problem
This is not a niche issue. Consider the numbers.
ChatGPT alone has 200 million weekly active users. Assume a conservative average of 5 responses per user per week that get pasted somewhere. That is a billion copy-paste events per week where formatting potentially breaks.
Now add Claude, Gemini, Copilot, Perplexity, and every other AI tool. The actual number is likely several billion paste events per week, globally, where the user has to choose between accepting broken formatting or spending time fixing it.
Most people accept the broken formatting. They send the email with asterisks. They share the doc with pipe characters. They post in Slack with markdown that Slack does not understand. The content gets through, but it looks unprofessional, and it erodes trust in the person who sent it.
The people who do fix it spend real time on a task that adds zero value. Nobody gets promoted for removing asterisks from a document.
The fix: a publishing layer
The solution is a conversion layer that sits between AI output and its destination. Not a generic converter that produces one output, but a system that understands both markdown and the specific requirements of each destination.
Unmarkdown™ converts markdown to 6 destination-specific formats:
- Google Docs: Rich text with proper heading styles and bordered tables
- Word: Native heading styles that work with the navigation pane and table of contents
- Slack: Slack's mrkdwn format with single-asterisk bold and
<url|text>links - Email: Inline-styled HTML that renders correctly across Gmail, Outlook, and Apple Mail
- OneNote: HTML with all 6 heading levels and styled tables
- Plain Text: Clean text with no formatting symbols, headings in ALL CAPS, tables as readable text
Each destination gets output optimized for its specific capabilities and limitations. The conversion takes seconds instead of minutes.
Beyond copy-paste
The formatting problem is solvable at the individual level with conversion tools. But the deeper issue is architectural. The AI industry has built extraordinary generation capabilities and invested almost nothing in the delivery layer.
Every AI tool gives you a "copy" button. None of them give you a "copy for Google Docs" button, a "copy for Slack" button, or a "copy for email" button. The assumption is that copy-paste is sufficient. It is not.
As AI-generated content becomes a larger share of business communication, the formatting gap will become harder to ignore. The last mile needs its own infrastructure, just as package delivery needed its own infrastructure separate from long-haul shipping.
The tools that bridge this gap will become as essential to the AI workflow as the AI tools themselves. The generation side is solved. The delivery side is where the work remains.
Related reading
- The AI Output Problem: Why Every AI Tool Writes in Markdown (And What to Do About It)
- 5 Things That Break When You Paste AI Output (And How to Fix Each One)
- Unmarkdown vs Copy-Paste: Why Raw Pasting Doesn't Work
- Obsidian is Too Complicated: When Simpler is Better
- ChatGPT vs Claude vs Gemini: Which AI Has the Best Formatting?
- 5 Ways to Use AI-Generated Documents in Your Actual Workflow
- How Consultants Are Using AI to Draft Client Reports
