At a glance
- On the paste: one tap on the Anti-LLM chip in the preview, or automatic on every ⌘V.
- Visible tells: em and en dashes to hyphens, curly quotes and apostrophes removed, the ellipsis glyph to three dots, emoji and arrows out.
- Structure: markdown flattened, list markers turned into plain bullets, links keep their address.
- Invisible characters: zero-width and non-breaking spaces removed, fullwidth and ligature glyphs normalised (NFKC).
- Your words stay: typography and structure change, wording never does.
What one pass changes
Order of operations matters more than the list. A literal \n or \t from a code or chat copy is restored first, so text that arrived as one escaped line becomes real paragraphs before any other rule runs.
All of it happens on the Mac. Nothing is uploaded, and there is no daily quota, because there is no server.
Punctuation: em and en dashes become hyphens. Every quote and apostrophe glyph is removed, so "don't" comes out as dont. The ellipsis character drops to three dots. Semicolons become commas. Colons are removed, with two exceptions on purpose: clock times like 12:30, and the :// in a URL.
Structure: headings, blockquotes, code fences and table pipes flatten into plain lines. Bold, italic and strikethrough lose the syntax and keep the words. A markdown link keeps its address, so [docs](https://klipto.me) reads as docs klipto.me. Line-start markers, 1. or (a), become bullets. The same markers inline become commas.
Hidden characters: this is the part people miss. AI output carries zero-width spaces, non-breaking spaces and fullwidth forms you cannot see on screen and that survive a copy. NFKC normalisation folds them back to plain letters, ｈｅｌｌｏ to hello and file to file, and the invisibles are dropped outright. Runs of blank lines collapse, and the period at the end of a paragraph goes too.
It does not beat AI detectors
Detectors read statistics in the word choice itself, mostly perplexity and burstiness. Swapping an em dash for a hyphen leaves that fingerprint untouched. I make no claim there. Any tool promising detection-proof output from punctuation edits is overselling. Rewriting sentences is a different product.
The useful job is narrower and far more common. Your Slack message, your email and your commit description should not announce that a model wrote them.
Some tells can be cut upstream. A ChatGPT custom instruction reduces em dashes, and GPT-5.1 added a toggle to switch them off in November 2025. Neither is reliable, and neither touches the invisible characters. For a one-off, try it free in your browser.
Where it sits in Klipto
Anti-LLM is one of more than 20 transforms in the paste preview of a native macOS clipboard manager (macOS 14+, about 11 MB, no Electron). The same panel converts colours, pretty-prints JSON, changes case and counts tokens, over local history with search and the Stack. The app touches the network only to check for updates, which are included.