What's Hiding in AI Text 🫧

Real examples of the invisible watermarks, lookalike letters and machine typography that ride along inside AI-generated text — made visible below, character by character.

1. The narrow no-break space — ChatGPT's calling card

ChatGPT habitually inserts U+202F (a thin, invisible-looking space) before percent signs and units. Nobody types this character by hand. Here's a typical line with the hidden characters exposed:

Revenue grew by 40U+202F% year-over-year a result the team attributes to the new pricing model.

One invisible space plus one em dash: two characters that quietly say "a model wrote this".

2. Zero-width characters — the invisible watermark

Zero-width spaces (U+200B), word joiners (U+2060) and byte-order marks (U+FEFF) render as nothing at all, survive copy-paste, and can be sprinkled through text as a trackable watermark:

This sentence looksU+200B completely normalU+2060, but it carriesU+FEFF three invisible characters.

They also break code, spreadsheet lookups and database keys — the classic "these strings look identical but don't match" bug.

3. Fake Unicode bold — 𝗻𝗼𝘁 𝗿𝗲𝗮𝗹𝗹𝘆 𝗯𝗼𝗹𝗱

That bold text in LinkedIn posts and AI-generated social copy isn't formatting — it's mathematical symbols that only look like letters. Screen readers can't read them and search can't match them:

𝗕𝗶𝗴 𝗮𝗻𝗻𝗼𝘂𝗻𝗰𝗲𝗺𝗲𝗻𝘁: we're hiring! What looks like "Big announcement" is actually 15 math symbols.

4. Markdown residue — the paste that gave it away

Chat interfaces render **bold** and ## headers nicely — but paste that text into an email or a form and the raw symbols come along:

## Key Takeaways 🚀 **Performance** improved across the board ✅ See the [full report](https://example.com) for details

Emoji bullets, bold markers, bracket-links: the instantly recognizable "pasted from ChatGPT" look.

5. The style tells — no hidden characters needed

Even with perfect characters, the phrasing gives it away. These patterns appear in AI text at wildly elevated rates:

In today's fast-paced world, this technology stands as a testament to innovation. It's not just a tool. It's a paradigm shift. Let's delve into the key takeaways.

TextWash flags these but never rewrites them — the fix should be in your voice, not another model's.

Which phrases does TextWash flag?

The engine watches 140+ words and phrases that appear in AI text at far higher rates than in human writing. A sample of the list:

delve · tapestry · testament to · it's important to note · in today's fast-paced world · ever-evolving landscape · game-changer · seamless · leverage · pivotal · unlock the potential · embark on a journey · at its core · gone are the days · key takeaways · actionable insights · low-hanging fruit · move the needle · paradigm · holistic · cutting-edge · meticulous · treasure trove · hidden gem · rich tapestry · dive right in · buckle up · food for thought · elephant in the room · as of my knowledge cutoff — plus Spanish, German and French equivalents.

Flagged phrases are never auto-rewritten: the rewrite should sound like you, not like a different model.

Check your own text

Paste anything into TextWash and get the same x-ray view of every hidden character, a fingerprint count, an AI slop grade from A+ to F, and a one-click wash. It's free, unlimited, and runs entirely in your browser — nothing is uploaded.

🫧 X-ray my text →