How does the AI reply checker work?

It matches your text against two fixed phrase lists: 64 "AI tell" phrases that language models overuse, and 16 outreach spam phrases that get social accounts filtered. Every match costs 15 points off a starting score of 100, and each one comes back with the phrase and a plain instruction for fixing it. No model is involved.

Is my draft sent to a server?

No. The checker is a pure function running in your browser tab. There is no API route, no server action and no analytics call behind it. Open your network tab and type: nothing fires.

Can this prove something was written by AI?

No, and nothing can. This is a phrasing check, not a detector. A person can write every phrase on the list, and a careful model can avoid all of them. What it tells you is narrower and more useful: whether your reply is built from the stock phrases readers have learned to scroll past.

Why does it flag phrases I use normally?

Because a lot of AI phrasing is ordinary phrasing used too often. The lists are kept high precision, but the flag is a prompt to look, not a verdict. If the phrase is genuinely how you talk, leave it in.

Is this the same check Quillen runs on its own drafts?

Yes, on the two universal passes. Inside the product the same function also enforces your persona: your own banned words, your formatting rules, your length limit. Those are switched off here because a visitor has no persona to enforce.

Free tool

Does this reply sound like a bot wrote it?

Paste a reply. You get back the exact phrases that read as machine-written, highlighted in place, each with what to say instead. This is the same linter every draft inside Quillen has to clear, running in your browser, on your text, for free.

Your reply or post10 flagged spans196 / 3000

That is a sample, already checked. Click into the box and type to replace it with your own.

Human score

0/100

Reads as machine-written

This is dense with the phrases models reach for. Cutting the highlighted spans is usually enough to fix it.

Outreach spam“would love to connect”
drop the connection pitch and react to what they actually said.
Outreach spam“love to connect”
drop the connection pitch and react to what they actually said.
AI tell“dive deep”
cut "dive deep" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“let's unpack”
cut "let's unpack" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“unpack this”
cut "unpack this" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“a testament to”
cut "a testament to" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“the power of”
cut "the power of" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“ever-evolving”
cut "ever-evolving" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“navigating the”
cut "navigating the" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“synergy”
cut "synergy" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“couldn't agree more”
cut "couldn't agree more" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.
AI tell“really resonates”
cut "really resonates" and say it the way you would text a friend. Use shorter words and a concrete detail from their post.

What it actually checks

Two phrase lists and one subtraction. Read out of the shipping code on 1 September 2026.

64 AI tell phrases

Hedge openers ("it is worth noting"), LLM buzzwords ("delve into", "seamless", "a testament to"), and reply-bot agreement ("couldn’t agree more", "this resonates"). The phrasing a model reaches for and a person almost never does.

16 outreach spam phrases

"Would love to connect", "feel free to reach out", "DM me". Classic pitch language that gets replies filtered and accounts quietly down-ranked. A persona can extend this list in the product; it can never opt out of it.

One flat penalty

Fifteen points per match, floored at zero. Deliberately dumb arithmetic, because a score you cannot reconstruct by hand is a score you should not trust.

What this is not

It is not an AI detector. Nothing is. It cannot tell you who or what wrote a piece of text, and any tool that says it can is selling you a coin flip with a percentage sign on it. All this does is match a fixed list of phrases and hand back the matches.

It is also not a style guide. A high score means you avoided the stock phrases. It says nothing about whether you had anything worth saying, which is still the part that decides whether anyone replies.

Common questions

How does the AI reply checker work?: It matches your text against two fixed phrase lists: 64 "AI tell" phrases that language models overuse, and 16 outreach spam phrases that get social accounts filtered. Every match costs 15 points off a starting score of 100, and each one comes back with the phrase and a plain instruction for fixing it. No model is involved.
Is my draft sent to a server?: No. The checker is a pure function running in your browser tab. There is no API route, no server action and no analytics call behind it. Open your network tab and type: nothing fires.
Can this prove something was written by AI?: No, and nothing can. This is a phrasing check, not a detector. A person can write every phrase on the list, and a careful model can avoid all of them. What it tells you is narrower and more useful: whether your reply is built from the stock phrases readers have learned to scroll past.
Why does it flag phrases I use normally?: Because a lot of AI phrasing is ordinary phrasing used too often. The lists are kept high precision, but the flag is a prompt to look, not a verdict. If the phrase is genuinely how you talk, leave it in.
Is this the same check Quillen runs on its own drafts?: Yes, on the two universal passes. Inside the product the same function also enforces your persona: your own banned words, your formatting rules, your length limit. Those are switched off here because a visitor has no persona to enforce.

You just ran the last step of the pipeline

Inside Quillen this check runs on every draft before you ever see it, against a persona built from your own writing, on conversations we surfaced for you. You approve; then it posts. Same linter, none of the typing.

Start free Try the thread splitter