Clean AI Text. Instantly.

Why AI-generated text needs cleaning before you publish

Every major AI writing tool — ChatGPT, Claude, DeepSeek, Grok, Jasper — outputs text that looks clean in its own interface but breaks the moment you copy it somewhere else. The reason is simple: these tools are built to render Markdown. Their interfaces display **bold** as bold, ## Heading as a heading, and the em-dash (—) as a stylistic separator. When you paste the raw text into Word, Gmail, a CMS, or a plain-text editor, none of that rendering happens — you get the raw symbols instead.

Beyond visible formatting, AI models also embed invisible characters that you cannot see but which cause real problems downstream: broken word counts, failed CMS validation, encoding errors in databases, and search-index anomalies. TextPurify removes all of it — visible formatting noise and invisible Unicode alike — in a single click, entirely in your browser.

What are hidden characters in AI text?

Hidden characters are Unicode code points that have no visible glyph but occupy space in a string. The most common ones in AI output are:

• Zero-width space (U+200B) — Inserted between words; breaks search, word count, and copy-paste into code editors.
• Zero-width non-joiner (U+200C) — Interferes with ligatures and causes unexpected line-break behaviour.
• Byte order mark / FEFF (U+FEFF) — Appears at the start of copied text and corrupts file imports and API inputs.
• Non-breaking space (U+00A0) — Looks like a regular space but is a different character — breaks string comparisons and database queries.
• Soft hyphen (U+00AD) — Invisible in most renderers but treated as a character by search engines and word-count tools.

DeepSeek has been documented embedding zero-width characters as implicit watermarks in generated text. TextPurify strips every invisible Unicode code point regardless of its origin, giving you a clean, portable string.

How to remove markdown from AI text

Markdown is a lightweight formatting syntax designed for plain-text documents that get rendered into HTML. ChatGPT, Claude, and Grok are all trained to format responses in Markdown because their interfaces render it. The problem arises when you copy that text into a destination that does not render Markdown:

• **bold text** appears as literal asterisks in Word, Gmail, and Notion plain mode.
• ## Heading appears as a # symbol at the start of a line in most CMSs and email clients.
• - bullet point appears as a hyphen followed by text, not a formatted list.
• `inline code` appears with backtick characters in spreadsheets and plain-text fields.

TextPurify's Strip Markdown option removes all Markdown syntax while preserving the underlying text. It handles headings, bold, italic, blockquotes, lists, code fences, and links. The Remove Asterisks option is a separate blunt pass that deletes every * character — useful for removing bullet symbols without processing the full Markdown AST.

Em-dashes, smart quotes, and typography normalisation

AI models are trained on professionally typeset text and replicate its typographic conventions. The em-dash (—) appears in ChatGPT and Claude output as a stylistic separator. The Unicode ellipsis (…) replaces three dots. Curly "smart quotes" replace straight ASCII quotes. While these are correct typographically, they create practical problems:

• Em-dashes break rich-text editor layouts in Webflow, Squarespace, and HubSpot.
• Curly quotes fail JSON parsing, corrupt HTML attribute values, and break SQL strings.
• The Unicode ellipsis (U+2026) is treated as one character by CMSs with character-count validation.
• Smart quotes cause find-and-replace failures when searching for a straight quote.

TextPurify normalises all of these to their plain ASCII equivalents: em-dashes to spaced hyphens, curly quotes to straight quotes, and the Unicode ellipsis to three dots — making the text portable across any destination.

DeepSeek R1 <think> blocks and Claude <thinking> output

The newest generation of reasoning AI models — DeepSeek R1 and Anthropic Claude with extended thinking — generate an explicit chain-of-thought before producing their final answer. This reasoning is wrapped in XML-style tags:

• DeepSeek R1 — <think>…</think> Reasoning blocks that can run hundreds of lines before the answer.
• Claude (extended thinking) — <thinking>…</thinking> Internal reasoning generated when extended thinking mode is on.
• DeepSeek API — <reasoning>…</reasoning> Alternate reasoning block tag in some API configurations.

These reasoning blocks appear verbatim when you copy text from the raw API response or from certain third-party interfaces. TextPurify's Remove <think> blocks option deletes the entire reasoning section and leaves only the final answer. The final answer always follows the closing tag and is never affected.

Writer, Developer, and SEO cleaning modes

Different use cases require different cleaning rules. TextPurify ships with three pre-configured modes so you don't have to toggle options manually:

• Writer mode — Full cleanup for prose. Strips all Markdown syntax, normalises punctuation, removes hidden characters, and fixes spacing. Ideal for blog posts, emails, and documents destined for Word or Google Docs.
• Developer mode — Safe for code. Preserves quotes and Markdown structure but removes zero-width characters, normalises whitespace, and strips DeepSeek <think> blocks. Does not touch code-relevant syntax.
• SEO mode — Optimised for web publishing. Removes invisible Unicode and normalises punctuation without stripping headings or structure — preserving SEO-relevant formatting while eliminating hidden-character anomalies that affect crawl and indexing.

You can also switch individual options on or off in the Cleaning Options panel to build a custom rule set for your specific workflow.

Export cleaned text as TXT, Markdown, PDF, or Word

Once your text is clean, TextPurify lets you download it in four formats — no other free AI text cleaner offers multi-format export. Plain text (.txt) is the most portable option for CMS imports. Markdown (.md) preserves heading structure for static site generators like Hugo or Jekyll. PDF produces a formatted document ready to share. Word (.docx) gives you a clean file for Microsoft 365, Google Docs import, or editorial workflows that require a native document format.

All export processing runs in your browser using client-side JavaScript. Your text is never uploaded to a server, stored in a database, or transmitted anywhere. The only data that persists is your cleaning history and personal stats, which are saved to your browser's localStorage and never leave your device.