humanflow

Free AI detector · with its reasoning

Know why it scored that way.

Most detectors hand you one scary percentage. HumanFlow shows how confident it is, the signals that drove the number, and how your sentences split across human, mixed and AI-like — so you can judge the text rather than take a figure on trust.

Check text freeHumanize instead
  • Signals, not just a score
  • 1,500 words per scan, free plan
  • Honest verdicts
  • No ads

Check a draft

A free account includes 10,000 detection words a month — a full essay in one scan. No card.

Checking text needs a free account — create one or sign in. No card. Free accounts get the score, its confidence, the headline signals and the band counts; the sentence map is on the paid plans.

Your text
0 words

Model-agnostic by nature

A detector reads the writing, not the logo.

There's no "ChatGPT mode" to pick and no model list to keep updated. Predictable rhythm — even sentence lengths, borrowed transitions — looks the same whichever assistant produced it, so one scan covers them all, including whatever ships next.

  • No model selection — paste and check
  • Works on edited and mixed drafts too
  • New models need no update from you

Same flat sentence, three models — same read

"In today's fast-paced world, effective communication is essential."

ChatGPT
0.86
Claude
0.84
Gemini
0.85
A person
0.14

Illustrative scores, not measurements. The signal is the flatness of the sentence — its author's logo never enters the math.

What a detection score really measures

A detection score measures how predictable writing is — not who wrote it. Language models pick the most probable next word over and over, which produces even sentence lengths, borrowed transitions, and smoothed-out specifics. Detectors, ours included, look for exactly that flatness.

That's also why every detector can be wrong in both directions. Careful human writing — an exam essay, a second-language writer, a well-edited memo — often reads as "too even". A number alone can't tell you which case you're in. The sentences can.

What the detector gives you

Sentence-level view

Every sentence gets its own read, so you see exactly which lines sound machine-written — usually the flat ones worth rewriting anyway.

Three-way verdict

Human, AI, or mixed — with the emphasis on mixed, because most real documents are. No single scary percentage pretending to be proof.

Built for false-positive reality

Research shows detectors misfire on real people. Ours tells you it's an estimate, shows its reasoning, and never claims to be a verdict on you.

Private and simple

Text is processed over encrypted connections solely to produce your result. No ads, no data sales, delete your account any time.

Works with every AI model

If you can paste it, HumanFlow can humanize it.

HumanFlow rewrites the text, not the model behind it — so drafts from any AI assistant read naturally afterwards.

ChatGPTClaudeGeminiCopilotLlamaDeepSeekGrokMistralPerplexity

How it works

01

Paste your text

An essay, an article, a submission you're reviewing — up to 1,500 words in one scan on the free plan.

02

Read the verdict

A human / AI / mixed verdict with its confidence, the signals behind it, and how many sentences fell in each band. On a paid plan, every one of those sentences is marked up in your text.

03

Fix or verify

Rewrite the flagged lines yourself, send them through the humanizer, or use the read as evidence that your writing is yours.

One number vs. a sentence map

The same paragraph, two ways of reporting it. Only one tells you what to do next. The free plan reports the score and the band counts; the map below — your own text, line by line — is what a paid plan adds.

A typical detector says

78% AI

Which sentences? Why? What now? The number doesn't say — and numbers like it have flagged plenty of fully human writing.

HumanFlow says

  • AI-like

    "In today's fast-paced world, effective communication is essential."

  • Human

    "My manager reads email on her phone, between meetings, usually annoyed."

  • AI-like

    "It is important to note that clarity fosters engagement."

Two flat lines to fix, one that's fine. That's actionable.

How accurate are AI detectors?

Honest answer: imperfect, everywhere, ours included. Seven widely used detectors misclassified 61.22% of TOEFL essays — written by real people sitting an English exam — as AI-generated, while handling essays by US 8th-graders almost perfectly. OpenAI shut down its own detector in July 2023 after it correctly identified just 26% of AI text.

That's why we report sentences, not verdicts about people. If you've been flagged for work you wrote, the sentence map plus your draft history is a far stronger response than any counter-percentage.

Go deeper

We build a humanizer and a detector, so we have an obvious interest here. These pages are written so that someone who never installs our app still leaves with a correct understanding — and they say plainly where the evidence runs out.

Turnitin AI detection, in full

The pillar for everything we have written on Turnitin: what its AI indicator measures, every figure the company publishes, and the independent research that disputes them.

Read →

GPTZero, reviewed

The consumer detector an individual instructor is most likely to run — what it costs, and why it recorded the highest false-positive rate of fourteen tools in peer-reviewed testing.

Read →

Originality.ai, reviewed

Built for publishers, not schools — and until 2024 its own CEO said so, advising against academic use entirely. What changed, and what happened to writers it got wrong.

Read →

Copyleaks, reviewed

The most accurate detector in peer-reviewed testing, and the one that produced the most false accusations in the same study. Why both are true.

Read →

Sapling, reviewed

One of very few detectors that publishes a false positive rate — under 3%. What that means across a cohort, and their own statement that no detector including theirs should be a standalone check.

Read →

QuillBot's detector, reviewed

Claims a 99% detection rate, requires 80 words, and states outright that when a result is unclear its model leans towards calling text human. That last disclosure changes how both verdicts should be read.

Read →

Grammarly Authorship, reviewed

Not a detector. It records where your text came from while you write instead of guessing afterwards — the one shipping product built on the same reasoning this cluster argues from, and what it still cannot cover.

Read →

Pangram, reviewed

The one detector with materially better independent evidence than the rest of the category — a claimed 1-in-10,000 false positive rate, an NBER working paper that broadly supports it, and the reason none of that helps if you were flagged by something else.

Read →

ZeroGPT, reviewed

The free detector most people meet first. Its false-positive rate ranged from 3.7% to 39.4% across the same corpus of human essays, depending only on the subject they were written for.

Read →

Scribbr, reviewed

Free, unlimited and decent — and the source of the detector ranking half the internet quotes, in which the tester's own brands take the top three places, undisclosed.

Read →

SafeAssign, reviewed

It does not detect AI — and the company that makes it tested the idea across 65 institutions, decided against shipping it, and published why.

Read →

Compilatio, reviewed

94–99% on its own test, under 78% on the peer-reviewed one — and the company publishes the methodology detail that explains the gap.

Read →

What a score actually means

There is no passing score, and both Turnitin and GPTZero say so in their own documentation. What the asterisk means, and why 20% is a display setting rather than a limit.

Read →

AI detection and PDFs

Turnitin accepts PDFs, within limits it publishes: 300 words of prose minimum, 30,000 maximum, and English, Spanish or Japanese only.

Read →

AI detection and PowerPoint

It does not run. PowerPoint is not among the four file types the AI writing report accepts — and slides rarely clear the 300-word floor anyway.

Read →

For teachers: reading a score

What a detection score can and cannot establish, the false-positive research in plain terms, and a checklist to work through before raising it with a student.

Read →

For students: what a score means

Why two tools give two answers, how to self-check honestly before submitting, and the steps that actually help if your own writing gets flagged.

Read →

Are AI detectors accurate?

What peer-reviewed testing found, including a 61.22% false positive rate on essays by non-native English speakers — and why OpenAI withdrew its own detector.

Read →

Why human writing gets flagged

Who false positives happen to, a worked before/after example, and the steps that actually help if your own work has been questioned.

Read →

Falsely accused of using AI?

What to do today, the evidence that rebuts a score, what to say in the meeting — and the universities that stopped using these tools, including one that cancelled its contract.

Read →

How to prove you wrote it

Version history, timestamps and process evidence — where to find them, what they actually establish, and what to do if you have none.

Read →

Appealing an AI detection finding

Once a written decision exists: the stages and what each decides, the evidence pack in order of weight, and the deadline that ends more appeals than any argument about detector accuracy.

Read →

How AI detectors work

Perplexity, burstiness and thresholds, explained plainly — and the structural ceiling the method runs into.

Read →

The short version

  • A detection score measures how predictable text is, not who wrote it.
  • False positives are common and fall hardest on people writing in a second language.
  • Different detectors regularly disagree about identical text, because each picks its own threshold.
  • Draft history — Google Docs version history, Word AutoSave — is the strongest response to an accusation.
  • No tool can promise a particular score from a particular detector. Anything claiming otherwise is selling you something.

Common questions

Is the detector free?

Yes — 10,000 detection words a month on the free plan, up to 1,500 words in a single scan. Every scan returns an overall score, a confidence read, the headline signals behind it, and how many sentences landed AI-like, mixed and human-like. The sentence map, which highlights those lines in your own text, is on the paid plans.

Can a detection score prove who wrote something?

No. It measures how predictable the writing is, which correlates with AI authorship but isn't proof in either direction. Treat every detector's output — including ours — as a signal to investigate, never a verdict.

Why was my own writing flagged?

Careful, uniform writing reads as predictable — the same signal detectors flag. It happens most to non-native English writers and heavily edited text. Every scan names the signals that pushed the score, and on a paid plan the sentence map shows which lines they came from. Either way, draft history is your strongest evidence.

Does checking text use my humanizer words?

No — detection has its own, much larger allowance on every plan, because it costs us far less to run than humanizing.

What happens to text I check?

It's processed over encrypted connections solely to produce your result. We show no ads, we don't sell personal data, and you can delete your account and its data at any time.

Read the reasoning, not just the score.

Free plan, 10,000 detection words a month, no card, no ads.

Check text free

Pages in this section last reviewed 27 July 2026.