Free Llama Watermark Remover & Checker for Text & Code

Paste Llama-generated text or code, see every hidden watermark character exposed in an x-ray view, and copy a clean version in one click. Free, online, no sign-up, and nothing you paste ever leaves your browser.

Paste your text/code

Cleaned Text/Code

More: what was found, x-ray view and options

Normalise unusual spaces Straighten smart quotes Em dashes (—) to hyphens

X-ray view

Your text with every hidden character exposed as a labelled chip. Hover a chip for its Unicode name.

Does Llama watermark the text it generates?

No, not yet. Meta has published real watermarking research, but none of it ships in Llama's text output, not in the downloadable models and not in the Meta AI assistant, a research-to-deployment gap worth watching. It's not the reason invisible characters turn up in Llama text, though: those are picked up on the way out, from terminals, web front-ends and documents, not from the model, and they're exactly what the tool below finds.

Last verified: 12 August 2026

Llama is the vendor page where the gap between research and deployment matters most. Meta's labs have produced some of the best-known academic work on marking model output, which is regularly compressed by commentary into "Meta watermarks Llama". It does not. Tracing where that work actually applies, and where Llama text actually picks up hidden characters, gives a clearer picture than either the rumour or a bare denial.

Research is not deployment

Three strands of Meta work feed the confusion. Its researchers have published on consolidating statistical watermark designs for language models, on "radioactivity", detecting after the fact whether a model was fine-tuned on watermarked synthetic text, and on shipped marking systems for other modalities, such as audio generated by its speech models. All three are real. None of them means a mark is embedded in the tokens a Llama model emits when you prompt it today. The papers describe what a provider could do at sampling time; the released models leave sampling entirely to whoever runs them.

Claim heard about Llama	What the record supports
"Meta watermarks Llama output"	No, no deployed text mark exists in any released Llama model
"Meta has watermarking technology"	Yes, as research, and as shipped systems for audio and imagery rather than text
"Meta AI labels its content"	Platform labels exist on Meta's apps, as interface metadata, nothing is embedded in the text you copy
"Llama text can be identified by a detector"	No such detector exists, and open distribution makes a general one impossible

Weights on a laptop cannot phone home

Llama's defining property is that the weights are handed out. People run them through llama.cpp, Ollama, vLLM and dozens of other stacks, quantised and fine-tuned into thousands of derivatives. A statistical watermark lives in the code that picks each next token, and here that code belongs to the operator, not to Meta. Whatever Meta's hosted assistant might one day apply to its own responses, the same architecture on private hardware produces unmarked text by construction. This is the same structural argument as for other open-weight vendors, but Llama's derivative ecosystem is the largest, so the gap between "the provider" and "the text in front of you" is widest here.

Where hidden characters actually enter Llama text

Because Llama runs everywhere, its output travels stranger routes than most model text, and each route leaves different residue:

Terminals. Output copied from a console session can carry C0 control codes, colour and cursor sequences begin with ESC (U+001B), which the checker flags as high-risk. Interrupted streams also leave stray carriage returns.
Local web front-ends. Interfaces that render markdown reintroduce the usual browser debris: no-break spaces from layout, soft hyphens from justified text, zero-width joiners inside emoji the model produced.
Messaging apps. Text from the Meta AI assistant inside a chat thread usually copies clean, but quoting it through other apps and documents adds the same characters any forwarded text accumulates.
Datasets. Llama output is widely used as training data, where a single BOM or bidirectional mark survives into every downstream copy, worth a check before anything is committed.

What this page's checker settles

Can

Expose control codes from terminal copies, which most cleaners aimed at prose never look for.
Audit a candidate dataset sample so invisible characters do not propagate into a fine-tune.
Show, on any sample you paste, that the model itself emitted no hidden signature.

Cannot

Say whether a passage came from Llama, a derivative, or a person. No text-level evidence distinguishes them.
Detect the statistical schemes described in Meta's papers, which need the provider's key even where they are used.
Clean visible markdown syntax, hash marks and asterisks are formatting, not hidden characters.

Questions about Llama and watermarking

Meta published watermarking papers, does any Llama model use them?

No released Llama model applies a text watermark, and Meta has not announced one for its hosted assistant. The papers analyse sampling-time schemes and post-hoc detection of watermarked training data; publication is not deployment.

Can a fine-tuned Llama derivative carry a watermark from its base model?

Not in its output, because the base model never marked its output. The "radioactivity" line of research concerns detecting watermarked training text inside a model, a forensic tool for providers holding a key, not a property of Llama text you can test for.

Why does my terminal-copied Llama output fail when pasted into a script?

Console sessions embed control sequences that survive some copy paths, an escape character or stray carriage return breaks parsers while remaining invisible in an editor. Paste the text above and the tally will name each control code it finds.

Do the "AI info" labels on Meta's platforms travel with the text?

No. They are interface elements attached by the platform, not characters in the message. Copy the text and the label stays behind, which is exactly why label-based provenance breaks down outside the app it was applied in.