8 common ChatGPT formatting artifacts (and how to spot them)
Once you've read a few hundred AI-generated documents you start to recognise the tics. They're not bugs — they're consequences of training to be helpful in a chat — but they don't belong in a finished document. Here are eight of them, with examples and the fix for each.
1. The "Sure! Here's…" opener
Almost every AI response starts with a conversational acknowledgement: "Sure!", "Of course.", "Here's a draft of…", "Below is a memo…". In a chat that's friendly. In a document it's filler. The fix is to delete the first paragraph if it's just acknowledgement.
2. The "Let me know if…" closer
The mirror image of the opener: "Let me know if you'd like any changes.", "Happy to refine further.", "I hope this helps!". Always at the end. Always cut.
3. The em-dash transition
AI prose loves the structure "It's not just X — it's Y." One or two of these is fine; six in two pages becomes a tic readers notice. Vary the sentence structure manually for anything you'll publish under your name.
4. Excessive horizontal rules
--- between every section. --- between every bullet. --- as a decorative bottom border. In the chat UI these read as visual breaks; in a document they read as visual junk. Cut all but the genuine major section breaks.
5. Tables with redundant headers
Two-column tables where the first column is a label and the second is the value — these are almost always better as definition lists or simple bold-then-text paragraphs. AI assistants reach for tables more often than they should.
| Field | Value |
| ----- | ----- |
| Name | Acme Inc. |
| Founded | 2014 |
| Employees | 142 |That same content as **Name:** Acme Inc. on its own line, repeated, reads better and travels better through any export.
6. Indented lists that drift
Multi-level nested lists where the indentation is one space sometimes and three spaces sometimes. The eye reads it as one structure; the parser reads it as a different structure. Normalise indentation — two or four spaces, consistently — and the bullets will render the way you expected.
7. Code blocks without a language tag
Many AI outputs open code blocks with bare ``` instead of ```python or ```bash. The code still renders, but syntax highlighting is off, and any downstream tool that uses the language tag (PDF renderers, technical writing platforms) can't help. Add the language to the opening fence.
8. Smart-quoted strings inside code
The most insidious one. AI assistants emit curly quotes (“ ”, ’) in body prose and the same quotes inside what should be ASCII code samples. Anyone who copies the code into a terminal gets a baffling error because the shell doesn't recognise curly quotes. Inside code blocks and inline code, normalise to ASCII straight quotes.
Spotting them in your own outputs
These eight patterns account for the great majority of "AI text that looks slightly off" in documents. None of them require a human to fix once you've decided the fix should happen. That's exactly what the cleanup pass in Markdown Tidy is for: paste the AI output, hit Clean, and all eight (and several more) are gone in one step. The full list lives in the cleanup checklist.
Related reading: The 12-item AI cleanup checklist · Why ChatGPT Markdown breaks in Google Docs