AI Text Cleaner

No model, no server. Each chip above is a regular expression that runs over your text in a fixed order: invisible characters first, then odd spaces, horizontal rules, markdown, dashes, quotes, ellipses, the optional AI phrases and emoji, and finally a whitespace tidy-up. The same input with the same settings always gives the same output.

The counts in the What changed box come from the same pass, so “12 em/en dashes” means twelve dash characters were found and replaced, not an estimate.

Chat windows render markdown, so the model writes it: **bold**, *italic*, # headings, backticks, - and 1. list markers, and [text](url) links. The cleaner removes the symbols and keeps the words. A link keeps its label and drops the URL, so paste the link back in if you need it. Code fences lose their fence lines but keep the code inside. Tables are left alone.

Every em dash and en dash, with any spaces around it, becomes the replacement you pick. The default is a comma, which reads right for most asides: “the plan, as it turns out, failed”. Choose full stop to split the sentence and capitalise the next word, or spaced hyphen if you like the rhythm and only hate the character.

A few cases are handled on their own. An en dash between two numbers (10–15) becomes a hyphen. A dash at the start of a line, such as a dialogue dash or a bullet, is dropped. A dash next to existing punctuation turns into a space so you never get “, ,”.

Curly quotes become straight quotes, the single … character becomes three dots, and non-breaking, narrow, thin and ideographic spaces become an ordinary space. Zero-width spaces, joiners, word joiners, byte-order marks, soft hyphens, bidirectional control marks and Unicode tag characters are deleted. The one exception is the zero-width joiner inside an emoji sequence like a family emoji, which is kept so the emoji does not split apart.

Off by default, because it deletes words. When on, it removes a short list of chatbot habits: openers like “Certainly!” and “Great question.”, lines like “Here's a quick overview:”, closers like “I hope this helps” and “Let me know if…”, summary lead-ins, and the worth-noting and important-to-note hedges. When a lead-in started the sentence, the next word is capitalised. It is a list, not judgement, so read the output before you send it.

Language models learned from books, journalism and edited web writing, where the em dash is common. Then training rewarded answers that read as polished, and the dash does a lot of work: it adds an aside, a pause or a punchline without committing to a full stop. The result is a habit people started calling the “ChatGPT hyphen”. The dash was always fine in human writing. The problem is density: four of them in a paragraph now reads as pasted.

For a long time, asking ChatGPT to stop did not stick. OpenAI said in late 2025 that ChatGPT would finally respect a custom instruction to avoid them, but text you already have still needs cleaning.

The chat window shows bold text and neat headings, but underneath it is plain markdown. Depending on how you copy and where you paste, you get either rich text or the raw source with the asterisks and pound signs still in it. Email clients, LinkedIn, form fields and plain-text editors show the raw version. That is where a ChatGPT text cleaner earns its place.

Two different things get called an AI watermark. The first is character-level: zero-width spaces, unusual Unicode spaces or tag characters sitting inside the text where you cannot see them. Some tools do insert these, and in 2025 people spotted narrow no-break spaces in output from some OpenAI models, which OpenAI described as a quirk rather than a watermark. This cleaner removes all of those.

The second is statistical. Google's SynthID Text, used in Gemini, nudges which words the model picks so a detector can recognise the pattern later. OpenAI has said it built something similar and did not ship it. A statistical watermark has no special characters to strip. It lives in the word choice, and so do the signals that AI detectors like GPTZero and Turnitin score.

So, plainly: this tool removes hidden characters and tidies formatting. It does not humanize text or defeat AI detectors, because it does not rewrite a single word. Anyone selling a character cleaner as a detector bypass is selling you a find-and-replace.

Paste with Ctrl+Shift+V (Keep Text Only) to drop fonts and colours. Markdown symbols survive that, so open Find and Replace with Ctrl+H. Word has codes for the awkward characters: ^+ finds an em dash, ^= an en dash and ^s a non-breaking space. Replace ** with nothing to clear bold markers.

To straighten quotes, go to File, Options, Proofing, AutoCorrect Options, AutoFormat As You Type and untick “Straight quotes with smart quotes”. Then find " and replace with ". Word retypes each one using the new setting. Do the same for the apostrophe.

Ctrl+Shift+V (Cmd+Shift+V on a Mac) pastes without formatting, and Format, Clear formatting fixes text already in the doc. For dashes, open Find and replace (Ctrl+H), paste an em dash into the find box and replace it with a comma and a space.

Tick Match using regular expressions to hunt invisible characters: \x{200B} finds zero-width spaces. Smart quotes are controlled under Tools, Preferences, but that only affects what you type next, not text already pasted. At that point pasting it through the cleaner above is quicker.