Invisible character remover

This is the draft behind “Try an example”: a LinkedIn post with the Markdown a chatbot’s copy button leaves in, plus the kind of hidden characters that copying through web pages and documents picks up. Each tag marks a character you would not see.

Unicode, the standard behind every character you type, includes characters that take up no space at all, and spaces that look like a normal space but are not one. They exist for good reasons: a no-break space keeps “10 km” on one line, a soft hyphen marks where a long word may break, and a zero-width joiner holds some emoji together. You cannot see any of them, so they come along whenever text is copied.

This is what the remover looks for, and what it does with each one:

Paste the text into the box above. Checking happens as it lands: the number at the top of the results is how many hidden characters the text has, the list names each kind and how often it appears, and the marked copy underneath shows where each one sits. Nothing changes until you press Copy text, so you can use it purely as an invisible character detector and keep your original.

A result of zero means the text has nothing hidden in it. Some characters are listed as kept, such as the joiner inside an emoji like 👩‍💻, because the text needs them to display correctly.

ChatGPT and Claude write in Markdown: **two asterisks** around a phrase make it bold, ### starts a heading, and a dash starts a list item. The chat window shows the formatting, but the copy button hands over the Markdown itself as its plain text version. LinkedIn does not read Markdown, so in a post those symbols stay symbols.

Check the Markdown box under the text field. Bold and italic become the Unicode letters the LinkedIn text formatter uses, which is the only styling a LinkedIn post keeps. Headings become bold lines, list dashes become bullets, links become the text followed by the address, and code marks and dividers are removed. A hashtag is not a heading, so #Hiring stays as it is.

On 23 September 2026 we asked ChatGPT (a temporary chat, default model) and Claude (an incognito chat, Opus 5.5) for the same eight LinkedIn posts, and recorded exactly what each copy button put on the clipboard, along with the text shown on screen.

Neither added a single hidden character, on the clipboard or on screen. The one non-printing character in all sixteen answers was the selector that tells ❤️ to display as an emoji, which belongs to it. What both copy buttons did add was Markdown. Eight answers each is a small sample from one day, so treat it as a snapshot, not a rule.

The question comes from April 2025, when Rumi reported that longer answers from OpenAI’s o3 and o4-mini models contained narrow no-break spaces where ordinary spaces belonged, and suggested they were a watermark. OpenAI said they were not a watermark but a side effect of how the models were trained, and later tests stopped finding them. Our own test above found none, and the full write-up is in Does ChatGPT watermark text?

As of September 2026, OpenAI’s help page on provenance signals lists watermarks for ChatGPT images and audio, not text, and says the goal is to add text under the EU’s Code of Practice on AI-generated content. When that happens, expect it to work the way Claude’s and Gemini’s watermarks do: through which words are picked, not through characters a tool like this could remove.

So a hidden character is not proof that a text came from ChatGPT, and a text without any is not proof that it did not. Hidden characters also come in through ordinary copying: web pages use no-break spaces and soft hyphens to control where lines break, and word processors add their own. Whatever put them there, the list above shows exactly what is in your text.

It does not touch a single visible character unless you check the box for dashes and quotes. The words, the line breaks, the emoji, and the bold or italic letters from the LinkedIn text formatter all stay as they were. That also means it does not make AI text read like a person wrote it: detectors judge the wording, and text watermarks, such as Google’s SynthID in Gemini and the one Anthropic added to Claude in August 2026, are carried in word choice, with no hidden characters at all. The AI content detector shows which lines read as generated. For images, the AI metadata remover does the same job for the Content Credentials that put a CR tag on LinkedIn.

Cleaning a pasted draft fixes the characters. The wording is the harder part, and it is what Depost AI works on: it learns your voice from the posts you have already written, drafts new ones from your analytics, and publishes them through LinkedIn’s official API once you approve them.