Can this tool prove that text was written by an AI?

No. It matches a fixed list of phrases that appear often in LLM output. Humans write those phrases too, and any LLM can produce clean text that the list misses. Treat it as an editing aid, never as evidence.

Does the tool send my text anywhere?

Pasted text is analyzed in your browser and saved to localStorage. The tool can also fetch text from a URL, and that fetch goes through the third-party service r.jina.ai — so avoid the URL feature for anything private.

Does it work on text in other languages?

The patterns are English phrases, so it will not detect equivalent clichés in Vietnamese or any other language. Non-English text will simply come back with nothing highlighted.

The LLM Cliché Highlighter Is an Editing Tool, Not an AI Detector

July 18, 2026

ai-writingdev-toolsediting

Simon Willison built a small browser tool that highlights phrases which show up disproportionately in language model output. Paste text in, get the tells marked.

It’s worth understanding precisely what it does, because the obvious use case — catching people — is the one it’s bad at.

What it actually detects

The tool looks for roughly ten recurring patterns in AI-generated writing. Among them:

“no X, no Y” chains
“sit with that”
“you already know”
“is real and”
“worth naming”

If you’ve read much LLM prose, these land immediately. They’re the verbal habits of a model reaching for emphasis: the rhythmic negation, the false intimacy, the pseudo-profound pause.

The tool runs client-side and stores your work in localStorage. It can also pull text from a URL, which it does via the third-party service r.jina.ai — worth knowing before you point it at anything you’d rather not send to a third party. Pasting text directly avoids that entirely.

Note that there’s no public repository or downloadable word list, so you can’t inspect or extend the pattern set. For a tool whose entire value is the list, that’s a real limitation.

Why it can’t detect AI writing

The tool matches fixed strings. That gives it two failure modes that no amount of tuning fixes:

False positives. Human writers use these phrases. “You already know” is ordinary English. A person who happens to write rhythmically will light up the highlighter while having written every word themselves.

False negatives. Ask a model to avoid clichéd phrasing and it largely will. The list catches lazy, default-settings output — exactly the text that was already obvious. It has nothing to say about careful LLM writing, which is the only kind worth detecting.

This is the general problem with AI-detection tools, and it’s why using one to make an accusation is indefensible. A phrase list is not evidence. If you’re tempted to run a colleague’s or a student’s writing through this, don’t — the tool can’t support the conclusion you’d draw from it.

Where it’s genuinely useful

Point it at your own drafts.

If you write with LLM assistance — and if you publish technical content in 2026, you probably do somewhere in the pipeline — the tells creep in without you noticing. You read the draft, it sounds fine, you ship it. Readers who’ve developed an ear for it notice immediately, and it costs you credibility on writing that may be perfectly accurate.

A concrete workflow:

Draft with whatever assistance you use.
Run the result through the highlighter before publishing.
Rewrite every highlighted phrase in your own words — not by swapping synonyms, but by asking what you actually meant.

Step 3 is the valuable one. The clichés usually mark places where the text is padding rather than saying something. Rewriting them tends to make the paragraph shorter and more specific, which is an improvement regardless of who wrote the first draft.

The list is the product

The reason this tool is interesting despite being technically trivial: it’s regex over a curated list, and the curation is all of the value. Anyone can write the matcher in twenty minutes. Knowing which ten phrases to include requires having read enormous amounts of model output and noticed the patterns.

That also explains why the approach doesn’t port. The patterns are English. Vietnamese LLM output has its own tells — different stock phrases, different rhythms — and none of them overlap with this list. Running Vietnamese text through the tool returns nothing highlighted, which means “no clichés found” only in the sense that the tool speaks the wrong language.

Building the equivalent for another language isn’t a coding problem. It’s an editorial one: someone has to read a lot of output and write the list down.

Try it

The tool is at tools.simonwillison.net/llm-cliche-highlighter.

Run something you wrote last month through it. The interesting result isn’t how many phrases light up — it’s whether you agree with the highlights when you reread the sentence.

Source: LLM cliché highlighter, Simon Willison’s blog.