Accuracy isn’t about being right. It’s about not being wrong.
Using AI to catch AI is a guess grading a guess. It’s circular, and wrong just often enough to ruin someone. So we took the AI out of the measurement path. The same text gives the same answer on any machine.
Three modes — each card below is one.
Out of 100 samples of AI generated text, it catches 99.9 of them.
*7,400 sample poolOut of 100 samples of human written text, it clears 95 and wrongly flags 5.
Less than 2 in 100 false positives when used together.
*tested on 1,800 documents*All samples drawn from industry-standard testing pools — DMITVA and RAID (ACL 2024).
Every figure here is signed and recomputable — see the proof →
Ask a language model whether a language model wrote something, and you inherit every one of its blind spots. That includes a documented bias against non-native English writers — and drift every time the underlying model updates.
Read the full argument →Don’t take our word for it. Take the math.
Every result is signed and reproducible — same text, same answer, any machine. You can recompute every number in your own browser. We didn’t build it to be right; we built it to not get it wrong.
The number we hold ourselves to.
flagged by mistake
Absolute Verification · both modes agree · 1,858 documents
*All documents from industry-standard testing pools (DMITVA, RAID).
Signed, verifiable results
Every result is Ed25519-signed and reproducible on your own machine — not just downloadable.
Verify it yourself →No AI, no GPU, no data center
Runs on any CPU. Detecting AI text shouldn't require the same infrastructure as generating it.
How it runs →Full methodology
The benchmark, the reference set, the per-configuration breakdown — nothing summarized away.
Read the proof →Built for the person writing, and the institution asking.
For writers & researchers
Check your own work before you submit it anywhere. Proofline never watches how you write — no keystroke logging, no version-history tracking. Free to start.
Get started free →For universities & publishers
Deployed on-prem or in your VPC. We’re ready to work with any institution that needs it — universities, publishers, business, or government.
Talk to us →Try it on your own text
3 free scans. No account needed to start.
Pasted text (800 characters or more) is English-only for now; document upload scores any language.
Frequently asked questions
What does Proofline do?
Proofline uses a deterministic calculation to identify the source of written text — AI-generated or natural authorship.
Can I verify the result of a Proofline detection?
Yes. Every result carries an Ed25519 signature unique to that exact result — change a single character and it no longer matches. On the proof page you re-check the signature against our public key and recompute the figures yourself, in your own browser. Nothing is taken on our word.
Is it accurate?
Proofline’s accuracy comes from a deterministic calculation, cryptographically signed so every result is reproducible and tamper-evident — recompute it yourself, not a marketing number.
Does Proofline track how I write?
No. Proofline was designed to eliminate the invasive methods of other detection and prevention tools. We don’t examine anyone’s writing process — we measure only the finished text you choose to submit.
Does it work on non-English text?
Yes — Document Review mode is language-agnostic; Quick Check is tuned for English.
How is Proofline different from a typical AI detector?
Proofline uses no LLM, no inference, no generative model, no neural network, and no multimodal reasoning in its measurement. It’s a deterministic calculation — the same text always gives the same result. A measurement, not an opinion. See verifiable vs black-box detection.