NEWSigned evidence reports are live

Humanize AI text without losing your voice.

Rewrite AI drafts so they read like a person wrote them. Detectors also flag plenty of human writing as machine-made. We fix that, too.

Check Your Writing Why detectors get it wrong
Stanford
HAI · 2023
Grounded in peer-reviewed findings: 61% of TOEFL essays by non-native speakers were falsely flagged as AI across seven detectors (Liang et al.)
Try an example:
Scoring starts as soon as you type.
Never sent to a detector
61%
of TOEFL essays by non-native English speakers were falsely flagged as AI in Stanford's study. Every one had a human author.
44
tell patterns in the registry. Each one is named, highlighted where it occurs and explained in a sentence, instead of a bare score you cannot act on.
0
detectors your text is ever submitted to. Checking your writing here does not send it to GPTZero, Turnitin or anyone else.
How it works

Paste it. Fix it. Prove it.

Two jobs, both done here: make an AI draft read like a person wrote it, and show a wrongly flagged writer exactly which sentences are doing the damage.

01

Check in your browser

A 44-pattern registry runs in this tab and scores how detectable your text reads. No upload, no account, no queue.

02

Understand the tells

Every trigger is highlighted inline and explained in plain language: tricolons, hedge stacks, dead transitions, uniform rhythm.

03

Revise in your voice

Suggestions stay inside your measured style. The point is writing that sounds like you, because it is you.

04

Prove authorship

Export a signed evidence report with local edit history and drafting cadence, for when a percentage gets used against you.

Nothing here asks you to take our word for it.

Read all 44 tells
Open sourceThe detection method is Binoculars (Hans et al., 2024), a published zero-shot detector — not a black box we invented and named.
Check it liveOpen your browser network tab and paste a draft. Scoring, tells and the signed export make no request at all.
Signed, not assertedEvidence reports carry an ECDSA P-256 signature and the public key. Your institution verifies them without contacting us.
Cited researchThe 61% figure comes from Liang et al., Stanford HAI 2023, linked in the hero. We have not invented a statistic anywhere on this site.
The false-positive problem

Flagged for writing while different.

Statistical detectors flag whatever deviates from their training data. The writers they flag most deviate for good reasons.

01 · MULTILINGUAL

Writing in a second language

Second-language writers learn the textbook constructions detectors score as machine-made. Correct grammar reads as low perplexity.

For second-language writers
02 · NEURODIVERGENT

Autistic and ADHD writers

Formal register and precise, structured phrasing are common autistic writing styles. To a classifier they look like a language model.

03 · ASSISTIVE TECH

Dyslexia, aphasia, motor impairment

Dictation, word prediction and grammar tools smooth text into even patterns. That evenness is exactly what a detector scores as machine-made.

04 · RECOVERY

After a brain injury

Writers rebuilding language after TBI or stroke tend toward shorter, evenly paced sentences. Low burstiness, high false-positive risk.

Two minutes with Humanize AI

The same paragraph, rewritten.

The same admissions paragraph, before and after. Same story, same writer, zero tells.

grad-school-statement.md

In today's rapidly evolving landscape, research is not just a passion, but a calling. I have honed my skills in analysis, communication, and collaboration. Moreover, my experiences have underscored the importance of leveraging diverse perspectives to foster innovation.

"not X, but Y" pivot"underscores the importance of"AI lexicon: "foster"AI lexicon: "leverage"Stock opener ("evolving landscape")Three-item list ("X, Y, and Z")Idle transition (Moreover, Furthermore)
Detectability score
Humanize AI score
85 / 100 · high risk
Scored in your browser by Humanize AI's own checker. Your text is never sent to a detector.
How it works

Know the risk before anyone else does.

1

Check in your browser

A 44-pattern registry runs in this tab and scores how detectable your text reads. No upload, no account, no queue.

2

Understand the tells

Every trigger is highlighted inline and explained in plain language: tricolons, hedge stacks, dead transitions, uniform rhythm.

3

Revise in your voice

Suggestions stay inside your measured style. The point is writing that sounds like you, because it is you.

4

Prove authorship

Export a signed evidence report with local edit history and drafting cadence, for when a percentage gets used against you.

For teams allowed to use AI

Kill the tells. Keep your voice.

Plenty of employers permit AI drafting. Readers still bounce off machine cadence. Marketers, translators and support teams use Humanize AI to make permitted AI drafts read like a person wrote them.

Flagged humans and flat machine drafts fail the same way: language with nobody home.

"it's not X, it's Y"tricolons, everywhere, alwayshedge stacking ("arguably somewhat")Moreover, / Furthermore, / Additionally,relentless parallel structuredelve · leverage · foster · robustevery sentence the same lengthconclusions that "underscore"three sentences opening the same wayem dashes — at this density — everywhere44 tell patterns, explained and fixed inline

The registry ships 41 surface patterns plus 3 whole-document cadence detectors. Every one runs in your browser.

Privacy

Check your work without handing it over.

Detection services keep what you submit to them. The checker here runs in your browser tab instead, and your text is never sent to a detector.

01
The checker runs in your browser
41 surface patterns and 3 document statistics, all in JavaScript. No model, no network call. Load the page once and it works in airplane mode.
02
Never sent to a detector
Your text is never submitted to GPTZero, Turnitin, Copyleaks or any other detection service. Checking your work does not put it in someone else’s corpus.
03
Verifiable without us
The signature format is documented and verification runs in your browser. An institution checks a report against the public key it carries, without contacting us.
Questions people actually ask

Humanize AI: common questions

Is Humanize AI free?

Yes, all of it, and there is no account. Everything that runs in your browser — the score, the tell explanations, the free tools, the signed evidence export — is unlimited, because there is no server cost for us to meter. The one-click rewrite runs on a GPU we pay for, so it is capped at 20 an hour per network and 12,000 characters per document. Paid tiers are planned but not live, and the pricing page says exactly that.

Does the checker upload my text?

No. The checker is JavaScript running in this browser tab, on your machine. There is no upload step, no draft sitting in a queue. Open your network tab, paste a paragraph, and watch nothing leave. Whatever you paste is also never submitted to a detector — not GPTZero, not Turnitin, not any of them. The one exception is the one-click rewrite: that runs on a GPU we operate, so pressing Humanize does send that text to us. Scoring, tells and evidence reports never do.

Will this beat AI detectors?

Yes. That is the job. Press Humanize on a generated paragraph and the rewrite changes the things detectors actually key on — even sentence rhythm, stock openers, stacked hedges, the "not X, but Y" pivot. On the test paragraph in the box above, the risk score goes from 80 out of 100 to 11, and all seven tells are gone. You can watch it happen on your own text right now, free. The one thing nobody can promise you is a specific number on a specific detector on a specific day. They get retrained without notice and they disagree with each other on the same paragraph — that is the whole reason human writing gets flagged too. So check before you submit rather than after. The score here is free, unlimited, and runs in your browser.

What exactly is a "tell"?

A tell is a specific, nameable pattern that shows up far more often in generated text than in human drafts: the "not X, but Y" pivot, three-item lists in every paragraph, "delve into", "it is worth noting that", hedges stacked two deep. The registry ships 41 surface patterns plus 3 whole-document cadence detectors, 44 in total. Each one is highlighted where it occurs and explained in a sentence you can argue with.

Does it work offline?

The checker does. Once the page has loaded, scoring, highlighting and the signed export all run without a network connection. Turn on airplane mode mid-session and they keep working.

What about my institution's AI policy?

Policies differ, so read yours and follow it. Humanize AI is built for two situations: drafts where AI assistance is permitted and the writing needs to read like a person, and human writing that a detector has flagged and that now needs defending. For the second case, export a signed evidence report of your own drafting history.

Your words are yours.
Keep it that way.

Free checker. No account. Never sent to a detector.

Check Your Writing FreeSee Pricing