ChatGPT watermarks and hidden characters: what a remover can and can't take out

Does ChatGPT watermark its text? What the April 2025 hidden characters were, how OpenAI's textGrain and Google's SynthID work, and how to clean pasted text.

Summarize with
A type specimen card with baseline, x-height and descender guide lines and one curved brush stroke with a green anchor point

Does ChatGPT watermark its text? Not with hidden characters. The invisible spaces people found in some ChatGPT answers in April 2025 were not a watermark: OpenAI told the company that reported them that they were "a quirk of large-scale reinforcement learning", and they stopped showing up days later. OpenAI's real text watermark, textGrain, announced on 5 October 2026, works another way. It puts a statistical signal into the model's word choices, for eligible ChatGPT and Codex text in the EU and for API customers who turn it on. A "ChatGPT watermark remover" that strips invisible characters cleans your text, but it can't touch a signal that lives in the words.

Hidden characters are still worth removing, for reasons that have nothing to do with detection. A narrow no-break space breaks a search, a zero-width space breaks a URL slug, and a tag character can carry instructions you can't see. I built a free invisible character remover for exactly that: paste the text and every hidden character shows up where it sits.

I'm a design engineer: I design products and build them in React and Next.js, the work behind my design engineering services, and pasted AI text lands in the forms and CMSs I build.

What people found in ChatGPT text in April 2025

On 20 April 2025 the company Rumi published a post saying OpenAI's new o3 and o4-mini models seemed to leave marks in their text. The main character it found was U+202F, the narrow no-break space. On screen it looks like an ordinary space, but it's a different character, so a program comparing text sees a difference a reader can't. Rumi reported it mostly in longer answers, such as full essays, and not in older models such as GPT-4o.

Two days later Rumi added an update: OpenAI had contacted it and said the characters were not a watermark but a side effect of training. On 23 April Rumi reported that the characters no longer appeared in its testing.

The narrow no-break space has a plain typographic job: it keeps a number and its unit on one line. Whatever the cause, a mark that one find and replace removes would be a poor watermark.

What OpenAI has said about watermarking text

OpenAI has worked on text watermarking for years, and its public position has changed.

In August 2024 the Wall Street Journal reported that OpenAI had a working text watermark it hadn't released. OpenAI then updated a blog post from May 2024. As TechCrunch reported on 4 August 2024, the update said the method had been highly accurate and held up against localised edits such as paraphrasing, but was weaker against global changes: running the text through a translation system, rewording it with another model, or asking the model to put a special character between every word and then deleting it. OpenAI also said a watermark could stigmatise AI as a writing tool for people whose first language isn't English.

On 5 October 2026 OpenAI announced textGrain. Its announcement, as posted on the OpenAI developer forum, says the watermark adds a statistical signal through the model's word choices, that eligible ChatGPT and Codex text generated in the EU would carry it, rolling out over the coming weeks, and that API customers worldwide can opt in for selected models, with watermarking off by default in the API. Access to the detector starts with approved researchers and expert organisations. The announcement is framed around the EU AI Act, whose Article 50 asks providers of generative AI to mark synthetic text, images, audio and video in a machine-readable way that can be detected.

The accompanying technical report, dated the same day, describes the method: when the model picks each token, the choice is tied to pseudorandom values derived from a secret key and the preceding text. The detector needs only the text and the key. Nothing in it relies on special characters.

OpenAI is clear about the limits: short passages and later edits make detection less reliable, a positive result can't say who used the system or how much a person contributed, and a missing watermark doesn't prove a person wrote the text.

How SynthID Text works, and why it isn't a character

Google DeepMind's SynthID Text is the best documented watermark of this kind. Its paper, published in Nature on 23 October 2024, says it changes only the sampling step, using an algorithm called Tournament sampling, and leaves the model's training alone. The same paper describes a live experiment on nearly 20 million Gemini responses. Google's developer documentation says a production version is in Hugging Face Transformers from version 4.46.0.

It nudges the probabilities of the next token in a pattern a detector holding the key can score. Google lists the weak spots: factual answers, where there's little room to choose other words, and text that is thoroughly rewritten or translated.

Hidden charactersStatistical watermark (textGrain, SynthID Text)
Where it livesExtra code points between or inside wordsThe pattern of which words the model picked
Can you see itNot on screen, but code and tools canNo, and not by looking at the characters either
Who can check itAnyone, with a tool or a regexWhoever holds the key and the detector
Removed by stripping charactersYesNo
Weakened byA find and replaceHeavy rewriting, translation, short text

So the honest answer to "does a ChatGPT watermark remover work" is: it removes characters. If the text carries a statistical watermark, stripping characters leaves it exactly where it was.

The characters that hide in pasted text

These are the ones I see most when text arrives from ChatGPT, PDFs, Word and web pages. Each name is the character's official Unicode name.

  • U+202F narrow no-break space and U+00A0 no-break space. They look like spaces but aren't matched by a plain space in search, splits or comparisons.
  • U+2009 thin space and the other sized spaces in the General Punctuation block, from U+2000 to U+200A.
  • U+200B zero width space, U+2060 word joiner and U+FEFF zero width no-break space (the byte order mark when it sits at the start of a file).
  • U+00AD soft hyphen. Invisible until a line breaks at it.
  • U+200C zero width non-joiner and U+200D zero width joiner. Careful here: emoji sequences and scripts such as Persian and the Indic scripts need them, so deleting every one breaks family emoji and some words.
  • Direction controls such as U+202E right-to-left override. The Trojan Source research from Cambridge showed how these can make source code display differently from how it runs.
  • Tag characters, U+E0000 to U+E007F. More on these below.

In a product they mean a slug that returns a 404, a search that misses "10 km" typed with a narrow no-break space, and a spreadsheet lookup that fails on a value that looks identical.

How to see and remove hidden characters

Use the tool. Paste text into the invisible character remover and each hidden character is marked in place, grouped by kind. One click cleans it. It keeps the joiners emoji and some scripts need, turns odd spaces into normal ones, decodes hidden text written in tag characters, and runs entirely in your browser, so the text is never uploaded. If the text is Markdown headed for a document, my guide to pasting Markdown into Google Docs covers the formatting side.

Use your editor. VS Code has highlighted uncommon invisible characters in source code by default since version 1.63 in November 2021, with settings under editor.unicodeHighlight, such as invisibleCharacters and ambiguousCharacters. If a string looks right but fails a comparison, open it there first.

Use a regex in code. When I accept pasted text in a form or a CMS, I normalise it on save rather than trusting it. A starting point in JavaScript:

// remove invisible and control characters, keep the zero width joiner for emoji
const hidden = /[­​‌‎‏‪-‮⁠-⁤⁦-⁩]|[\u{E0000}-\u{E007F}]/gu;
// turn unusual spaces into a plain space
const spaces = /[  -    ]/g;

export const cleanText = (s) => s.normalize('NFC').replace(hidden, '').replace(spaces, ' ');

Two cautions. This version removes the zero width non-joiner, which Persian text needs, so leave ‌ out of the pattern if your product serves those languages. And it removes every tag character, including the ones inside subdivision flag emoji such as the Scotland flag, which is usually a fair trade for a text field but not for a chat app. The tool handles both cases; a regex this short can't.

Why cleaning text doesn't make it undetectable

I want to be plain about this, because "ChatGPT watermark remover" is often sold as a way to make AI writing pass as human. Removing hidden characters doesn't do that.

A statistical watermark sits in word choice, so it survives character cleaning, and detectors without a watermark look at the writing, not at spaces. OpenAI also says a missing watermark proves nothing about who wrote a text, so you can't clean your way to proof of authorship.

If a school, a client or an employer asks whether you used AI, the answer belongs to you, not to a tool. Use a cleaner because the text has to behave in code, a CMS, a spreadsheet or a search index. Don't use it, or any tool, to pass off AI writing as your own. My guide on shipping AI features people trust makes the same case from the product side: telling people plainly that something is AI is part of what makes them trust it.

Hidden instructions: tag characters and prompt injection

The more serious reason to look for hidden characters is security. The Unicode Tags block, U+E0000 to U+E007F, mirrors ASCII: each printable letter has an invisible tag twin. Most interfaces don't draw them, but a model can still read them.

In January 2024 the security researcher Johann Rehberger wrote up how this works as ASCII smuggling. He described a proof of concept by Riley Goodside in which pasted text carried hidden instructions that ChatGPT followed, and showed how instructions could hide in web pages, PDFs or documents that a person forwards without seeing them. His advice for builders was to filter tag characters both in prompts and in responses.

For anyone building with AI, that means three checks. Strip or flag tag characters in any text that reaches a model from outside, such as uploads, scraped pages or email. Strip them from model output before it's shown or stored. And when a person approves an AI action, show them what the model actually received. My guide to reviewing AI-generated UI code covers the other security checks I run on interface code before it ships.

Frequently asked questions

Does ChatGPT put a watermark in its text?

Not in characters. On 5 October 2026 OpenAI announced textGrain, a statistical watermark in word choice, for eligible ChatGPT and Codex text in the EU and opt-in API use. The hidden spaces seen in April 2025 were, by OpenAI's account, a training side effect, not a watermark.

What are the hidden characters in ChatGPT text?

Mostly the narrow no-break space, U+202F, which looks like a normal space. Text from any source can also carry zero-width spaces, no-break spaces, soft hyphens, direction marks and tag characters.

Does a ChatGPT watermark remover work?

It removes invisible characters, which is useful for clean text. It doesn't remove a statistical watermark such as textGrain or SynthID Text, because that lives in the choice of words, not in extra characters.

How do I remove invisible characters from text?

Paste the text into a tool such as my invisible character remover, turn on Unicode highlighting in VS Code, or normalise the text in code with a regex like the one above. Keep the joiners that emoji and some scripts need.

Can removing hidden characters stop AI detection?

No. Watermark detectors read word patterns and other detectors read the writing itself. Clean text so it works in your product; be honest about how it was written.

What is ASCII smuggling?

Hiding text in Unicode tag characters, which mirror ASCII but don't display. Some AI models read them, so they can carry instructions a person never sees. Filter them from input and output.

Sources and method

I read every source on 27 October 2026. OpenAI's own announcement page and help article returned an error to my browser that day, so the textGrain details come from OpenAI's announcement as posted on its developer forum and from its technical report on OpenAI's site. OpenAI's reply about the April 2025 characters is reported by Rumi, which received it; I couldn't find it published by OpenAI. The 2024 blog update is quoted from TechCrunch's report. Unicode names were checked against the Unicode Character Database. I give no detection rates for any watermark, because the figures I found weren't on a page I could open.

  • Rumi, New ChatGPT models seem to leave watermarks on text (20 April 2025, updated 22 and 23 April 2025), date read 27 October 2026: rumidocs.com
  • TechCrunch, OpenAI says it's taking a deliberate approach to releasing tools that can detect writing from ChatGPT (4 August 2024), date read 27 October 2026: techcrunch.com
  • OpenAI Developer Community, OpenAI's approach to EU text provenance rules (5 October 2026), date read 27 October 2026: community.openai.com
  • OpenAI, textGrain: Entropy-Calibrated Watermarking for Language Model Text, technical report (5 October 2026), date read 27 October 2026: cdn.openai.com
  • EU AI Act, Article 50, transparency obligations, date read 27 October 2026: artificialintelligenceact.eu
  • Dathathri and others, Scalable watermarking for identifying large language model outputs, Nature 634 (23 October 2024), date read 27 October 2026: pmc.ncbi.nlm.nih.gov
  • Google AI for Developers, SynthID Text (last updated 9 April 2025), date read 27 October 2026: ai.google.dev
  • Google DeepMind, Watermarking AI-generated text and video with SynthID (14 May 2024), date read 27 October 2026: deepmind.google
  • Unicode Consortium, code charts (General Punctuation 2000 to 206F, Tags E0000 to E007F), date read 27 October 2026: unicode.org
  • Visual Studio Code, release notes for version 1.63, Unicode highlighting (November 2021), date read 27 October 2026: code.visualstudio.com
  • Johann Rehberger, Hiding and finding text with Unicode tags (14 January 2024), date read 27 October 2026: embracethered.com
  • Nicholas Boucher and Ross Anderson, Trojan Source, date read 27 October 2026: trojansource.codes

If hidden characters keep breaking your product's search, slugs or AI features, see my design engineering services. A project starts with a free 30 minute intro call, then a written scope and one fixed quote.

Tags: AI, ChatGPT, Unicode, Prompt injection, Design engineering

Keep reading

Let’s look at your product.

A free 30 minute call. You leave with a clear next step, whether we work together or not.

Book a free intro call