Scan text copied from Llama for hidden Unicode characters, right in your browser. Private and free, since nothing ever leaves your device.
Llama watermark detector
Paste in text copied from Llama and this app scans it for zero-width and non-printing Unicode characters, the kind that take up a position in a string while showing nothing on screen. It's built around a specific search: does Llama hide a watermark in its output.
Everything runs on your device. There's no account, no upload and no cap on how much text you check.
Workflow
A routine check before anything from Llama ships anywhere.
Why it matters
A zero-width space inside a pasted heading splits the keyword, so your slug or anchor stops matching the visible title. A non-breaking space carried over from whichever interface rendered the reply can break a shell command, a config file or a CSV column with no obvious cause.
The same characters cause broken find-and-replace, failed CMS saves, and layout problems in exported PDFs. Checking first turns a category of confusing bugs into a five-second fix.
Where they come from
Because the weights are openly available, dozens of unrelated apps serve Llama: Meta's own consumer apps, developer playgrounds, self-hosted tools and API wrappers with their own front ends. Each renders Markdown and converts it to plain text differently on copy.
That's the most checkable fact here: what a Llama copy carries says far more about which specific interface produced it than about the model underneath. Two people quoting an identical answer through two different hosts can see different invisible residue.
Straight answer
No, and less so for Llama than for most models, since there's no single interface it comes from. The same characters show up in text copied from Word, Google Docs and other tools entirely unrelated to Llama, so a finding is a formatting lead, not proof of source.
It also says nothing about whether the text would pass a separate AI detector, since those tools look at sentence patterns and word choice, not invisible bytes.
Who it's for
Anyone moving Llama output into a place that parses text strictly, publishes it, or judges it.
Not a deliberate hidden marker. Text copied from Llama can carry ordinary invisible characters, and because Llama runs through so many different interfaces, what shows up depends heavily on which one you used.
Paste it above and press Scan text. Every hidden character found is listed along with where it sits.
Yes, with no limit on how many checks you run, since scanning happens on your own device.
No. The scan runs entirely in your browser. Nothing is sent to a server, stored or logged.
No, and less so than for models with one official app. It's a formatting signal tied to the interface, not proof of source.
Because it's open-weight, dozens of unrelated hosts serve it, each with its own Markdown rendering and copy behavior. There is no one canonical Llama interface.
No. It scans characters, not writing style, and makes no claim about which model produced the text.
No tool can promise that. Detectors weigh sentence patterns and word choice far more than invisible characters.
No. This tool only reports findings. Nothing is altered unless you run the separate cleaner.
Yes, the site's watermark remover strips flagged characters while keeping your wording exactly as written.
No. Since the scan runs on your device rather than a server, there's nothing to cap.
Yes. A zero-width space inside one splits the visible text from what a crawler or CMS actually parses.
Yes. Indentation and line breaks are preserved, and any zero-width characters breaking a build are flagged.
Yes, the characters are language-independent, though the zero-width non-joiner is meaningful in Persian and some Indic scripts and worth a closer look before you act on it.
Different hosting interfaces render and convert Markdown differently, which is the main reason the same underlying model can produce text with different hidden-character residue.
Not yet, on this app. For multiple documents, paste, scan and repeat for now.
Yes, since nothing leaves your device there's nothing to worry about sharing.
Yes, the layout and the scan both work the same on a phone browser.
The scan itself is identical. This page is aimed at people specifically searching for Llama, since the many hosting interfaces shape what gets left behind, not the underlying check.
Not for this scan. The characters found are a property of the interface's text handling, not of which Llama model version generated the reply.
No, though you can. The app works fine offline once loaded, which is a quick way to confirm nothing is being sent anywhere while you scan.