Llama invisible character detector

GPTCleanup Pro's Llama detector scans a passage for the zero-width and non-printing Unicode characters that a Llama-based app's copy path can leave behind, without changing how the text renders on screen. It answers a precise version of a common question: not whether Llama hides a watermark, but exactly which invisible codepoints, if any, sit in this specific text.

It's a server tool, since the scan runs against our AI server rather than in the page itself. You get 5 free checks in any rolling 24 hours, no account required, and each result names the exact character and its position rather than a vague warning.

Every character in scope is an ordinary Unicode codepoint with a documented technical purpose, grouped here into seven families for clarity:

Zero-width space (U+200B)

Occupies a character slot and renders nothing. The single most common invisible marker in copied AI output.

Zero-width joiners (U+200C, U+200D)

The non-joiner and joiner, sometimes left over from how rich text renders certain characters and emoji.

Word joiner (U+2060)

An invisible joiner that prevents a line break at that point, often near a bolded phrase.

Byte-order mark (U+FEFF)

A marker meant to sit at the start of a file, occasionally left mid-response by an export or copy path.

Soft hyphen (U+00AD)

Invisible until a line wraps there, then a hyphen appears in a word that never had one.

Non-breaking spaces (U+00A0, U+202F, U+200A)

Look identical to a normal space, behave differently in code, slugs and search.

Bidi controls (U+202A to U+202E)

Five directional embedding and override characters that can quietly reorder mixed-direction text.

One round trip, a precise result. Nothing to configure beforehand.

Step 1

Copy the text using the copy control of whichever app is running Llama, since that path is where these characters are introduced.

Step 2

Run the check. The text goes to our AI server once, gets scanned against the full character list, and the result returns.

Step 3

Review each finding by name and position. Nothing about the text is altered by this scan.

Llama is open-weight and runs inside many different products, Meta's own apps, WhatsApp and Instagram integrations, third-party chatbots and self-hosted playgrounds, each with its own rich-text renderer. Converting that structure to plain text on copy is the primary source of these characters, and exactly which ones appear depends on the specific app, not on Llama itself carrying a fixed signature.

Identical characters appear when copying from a Word document, a PDF, or a translated page, none of which involve Llama. A finding here says something specific about the copy path that produced this text; it does not establish that Llama, rather than some other source, generated the underlying words.

No, and precision matters here. A hidden character is evidence of a formatting artefact in the copy path, not a fingerprint unique to Llama. The same characters appear in human-typed documents exported from Word, in translated text, and in PDF copy-paste.

This should be distinguished from a genuinely different, active research area: statistical, token-level watermarking, where a model's next-word probabilities are subtly biased in a pattern only a statistical test over many tokens can detect, not something implemented as inserted characters. A character-level scan, including this one, has no visibility into that kind of watermark at all.

A range of entirely human workflows produce the same characters as a Llama-app copy-paste, which is why a finding here carries no inference about authorship on its own.

Messaging app formatting

WhatsApp, Instagram and other chat apps render rich text of their own and can leave formatting characters behind on copy.

Translation tools

Machine and professional translation software frequently leaves directional or joining characters behind.

PDF copy-paste

Extracting text from a PDF often introduces stray characters the original document never displayed.

Other chat interfaces

ChatGPT, Claude and other assistants render rich text too, and can leave similar characters behind on copy.

CMS and rich-text editors

Switching between rich text and plain text views in a CMS is a common source of leftover formatting characters.

Regardless of origin, these characters cause concrete problems once text moves somewhere that parses it strictly. A zero-width space inside a heading splits a keyword, so a page's slug and its visible title stop matching. A byte-order mark inside a JSON file breaks the parser, and non-breaking spaces confuse diffing tools.

Pasting directly from a Llama-based app into a CMS, an email or a code editor is exactly where these problems surface. Checking first turns a category of confusing formatting bugs into a non-event.

Writers, editors and developers pasting Llama-based output into a CMS, an email or a codebase use this to catch formatting problems before they surface elsewhere. It's a server tool by necessity, which is why the free allowance is metered.

You get 5 free checks in any rolling 24 hours, tracked without an account. For repeated cleaning with no daily limit at all, the site's zero-width space remover and watermark cleaner run entirely in the browser and never touch a server.