FAQ
Questions about AI detection, answered straight
Paste a draft first — then read the answers, including the ones that do not flatter the product.
50 free scans a day · 2,500 words per scan
The basics
What the detector is, and what it hands back when you run it.
What is an AI detector?
An AI detector (sometimes called an AI checker or AI content detector) reads a piece of writing and estimates how likely it is to have been generated by a model such as ChatGPT, Claude, or Gemini. Ours looks at predictability and sentence rhythm, then returns an overall probability plus sentence-level highlights.
It is not a plagiarism checker. Plagiarism matching asks whether the words already exist in a database of sources. AI detection asks whether the texture of the writing looks generated. Fully generated text that matches nothing in a corpus is the normal case, not an edge case, which is why the two checks must be read separately. A report from this site is a probability estimate about a piece of text, not a finding about a person.
Can markhuman detect ChatGPT, Claude, and Gemini writing?
Yes. The AI detector scores any pasted text for AI patterns common in ChatGPT, Claude, Gemini, and similar models. You get an overall AI probability plus sentence-level highlights showing which lines look most machine-written.
You do not have to pick a model first. The box on the detector page scores the writing itself. The ChatGPT, Claude, and Gemini pages exist because those queries are not the same as "AI detector" — each model has a slightly different scaffolding, and spelling that out is useful when you already have a guess. Open-weight families (Llama, Mistral, and similar) share enough instruction-tuned cadence that they sit in the same scan. We do not publish a per-model accuracy table; vendor-measured numbers describe the vendor's test set, not your document.
What do the verdicts mean?
Human-Written means the text reads as composed by a person. AI-Generated means strong machine signals throughout. AI-Assisted means a machine draft that a person has since edited by hand. AI-Humanized means text carrying the fingerprints of a humanizing or paraphrasing tool. Mixed Human & AI means some passages read as machine-written and others as human.
AI-Assisted is a positive finding, not a hedge. On real documents it is the most common truthful answer, and it is the case a single AI-probability score handles worst — the mid-range number that gets argued about instead of acted on. Rounding it into a yes or a no is how both false accusations and missed drafts happen. Read the sentence shading before you act on the headline.
Why doesn't the percentage always match the verdict?
Because the verdict is not the percentage rounded off. Our thresholds are calibrated per length band, so 62% on a 90-word snippet and 62% on a 700-word essay are not the same finding — the shorter one carries far less evidence. The percentage tells you how strong the signal is; the verdict tells you what that strength means for a text of this length.
Do not re-derive a verdict by cutting the percentage at a round number. That throws the length calibration away and is how a confident-looking false accusation gets made. Use the classification the report returned. If the document view and the sentence view seem to disagree, the report says so — that disagreement is information, not a bug to round away.
How do I read the sentence highlights?
Shading marks where the signal is strongest, not which sentences are guilty. On a document judged human or mixed, you get both colours, so you can see which passages read machine-written and which do not. On a document judged AI-authored, every line is shaded — strongly where it gives itself away, softly where it does not.
The useful move in a conversation is to point at two or three lines and ask how they were written, not to read the percentage aloud as an opening. Highlights are a map of attention. They are not a list of charges. Short quoted material, headings, and bullet fragments are weaker evidence than continuous prose; weigh those lines more lightly than a long paragraph that all sits at the same strength.
The verdict says AI but every sentence looks fine. Why?
This is the single case we deliberately refuse to show as green. Text run through a humanizer is built so that no individual sentence looks wrong — the tell only exists across the whole document. We measured a fully machine-written document whose individual sentences scored 2% and 3%; painting those green would have cleared AI text line by line, which is the one direction this tool must never err in. So under an AI verdict the shading grades strength, and never says a line is human.
If that feels unsatisfying, it is supposed to. A line-by-line green light on a document the model already called AI would be the product contradicting itself in the place a reader trusts most. Treat the document verdict as the call, and the shading as where that call was strongest. Then read the words, and talk to the writer if the result matters.
Does it catch text that's been through a humanizer?
That is what the AI-Humanized verdict is for. The detector is trained on AI text pushed through commercial humanizers and paraphrasers, not only on raw model output, because raw output is the easy case and almost nobody submits it. It is still not infallible — see the accuracy answers below.
A synonym pass is much easier to call than a genuine structural rewrite. If someone regenerated the passage — new sentence boundaries, new paragraph mass, the restating close gone — the signal moves, sometimes a lot. That is also what our own humanizer is trying to do, which is why we will not pretend a rewrite is undetectable, including ours. Treat a borderline result on heavily rewritten text with more caution than a clean one.
Accuracy and limits
Where the score is worth trusting, where it isn't, and what it can never settle on its own.
How accurate is the AI detector?
No detector is perfect — ours included. The score is a signal, not a verdict: it reflects how predictable the wording is and how uniform the sentence rhythm feels. Short texts and heavily edited AI text are harder to call, and human writers with a very formal style can trip false positives. We do not publish a headline accuracy percentage because we have no third-party-verified figure to stand behind. Treat the report as one input among many, especially in high-stakes calls.
Vendor-measured numbers describe the vendor's own test set under the vendor's own conditions, not your document. Head-to-head tables against GPTZero, Turnitin, Originality.ai, or anyone else would be the same class of claim if we have not measured them, so we do not print them. Disagreement between tools is normal. Where two scanners split, the honest reading is that the text is genuinely ambiguous and deserves a human look.
My own writing was flagged. How does that happen?
False positives are real and we will not pretend otherwise. The patterns a detector reads as machine-like — even sentence length, plain connectives, low surprise in word choice — are also the patterns of careful formal prose, of technical and legal writing, and of fluent English written by someone who learned it as a second language. If a result on your own work looks wrong, it may simply be wrong. Tell us what you scanned; confident misses are the most useful thing anyone sends us.
A flag on writing you know is yours is a reason to look at the sentence shading, not a reason to panic. Formal lab reports, five-paragraph essays written to a rubric, and translated-then-edited prose all sit close to the decision boundary. If you are a teacher reading this because a student says the same thing, that is possible too — which is why a score alone should never carry a sanction. The accuracy and limits section is the rest of that argument.
How much text do I need to paste?
At least 50 words, and the more the better. Below that the report says so rather than guessing, because a couple of sentences do not carry enough rhythm to judge. A few hundred words of continuous prose gives the clearest read.
Lists, code, poetry, and slide bullets will produce unreliable scores even when they are long, because they do not have the sentence rhythm the engine is built to read. If you are checking a mixed document — some passages written by a person, some by a model — paste enough of both that the sentence-level report has something to point at. A single document score will average them into something less useful.
Why do two AI detectors give different answers?
Different training data, different signals, and different thresholds. Detectors are not measuring one agreed-upon quantity the way a thermometer does — each is a model with its own idea of what machine writing looks like, so disagreement between them is normal rather than a sign that one is broken. It is also a good reason not to treat any single score as settled fact.
When tools disagree, look at the sentences, not the headlines. A scanner that shows you which lines drove the call gives you something to check by eye: even rhythm, missing specifics, a restating close. A scanner that only returns a percentage leaves you arguing about the number. Ours is built for the first case. If you also need a specific institutional tool's read — Turnitin, GPTZero, Copyleaks — run that tool. We will not translate one into the other.
Does editing AI text make it undetectable?
Editing changes the signal, and enough of it changes the verdict — that is exactly what the AI-Assisted class describes. But there is no threshold we can give you, because it depends on what you changed. Rewriting the structure moves the score far more than swapping words does.
A thesaurus pass leaves paragraph mass, signposted transitions, and the restating close in place. Those are the properties a detector can score across a whole document. Regenerating the passage — new sentence boundaries, a short line after a long one, the hedges cut — moves more, and can also move meaning. There is no honest "how many edits until it passes" number. If you need the method without a box, the how-to-humanize guide is five hand edits. If you want a rewrite scored by our detector, that is the humanizer.
Can I use a report as proof that someone used AI?
No, and please don't. A report is a probability estimate about a piece of text, not a finding about a person. Do not use it as the sole basis for an accusation, a grade, a disciplinary outcome, or a hiring decision. Weigh it against drafts, version history, and a conversation with the writer.
The consequences of a false positive fall on a named person. The consequences of a missed draft are usually diffuse. That asymmetry is why process has to carry decisions, not scores. Teachers have a longer version of this on the educator page: triage privately, read against what you already know, ask for process, act only on convergence. The same sequence holds for editors, hiring, and compliance — a report selects where to look, it does not finish the looking.
Your text and your privacy
What happens to what you paste — the short answer is nothing.
Do you store the text I paste?
Text is processed in memory to produce your result and is not saved to a database. Refresh the page and it's gone. If you later want history or saved reports, that will be an explicit, opt-in account feature.
That matters most when the document is not yours — student work, a client draft, something covered by a confidentiality clause. "Uploaded into a platform" is a different privacy proposition than "scored and discarded." We score and discard. The privacy policy in the footer is the longer version of that sentence, including what we do retain (account records, billing, the cookies listed below) and for how long.
Do you train on the text I submit?
No. Submitted text is not used to train, tune, or evaluate any model, it is not read by us, and it is not passed to anyone else for their own purposes. The detection service is our own rather than a third-party API, so there is no vendor in the middle with its own retention policy.
The humanizer is the same rule: text is sent to our inference service to be rewritten and scored, and is not sold, shared, or used to train a model. If that ever changed, it would be an explicit, opt-in change described in the privacy policy — not a quiet reuse of pastes. We would rather lose the training data than train on writing people did not offer for that purpose.
Do I need an account to run a scan?
No. The free detector runs without signing up: 50 scans a day per browser, up to 2,500 words each, no email and no card. An account exists so a monthly allowance can follow you across devices, not as a gate on the tool.
The humanizer is the same shape: 2,000 words a day with no account, 10,000 signed in, no blur and no email gate on the result. Signing in raises a cap. It is not a trial that expires after a paragraph, and it is not a bait for a paid humanizer we have not shipped.
Plans and billing
What the free tier covers and what a paid plan adds.
Is markhuman.ai a free AI detector?
Yes. Free accounts get 20,000 words a month and 2,500 words per scan, plus 50 scans a day per browser — no credit card required. Paid plans unlock higher monthly pools and longer documents per scan.
The daily scan cap and the monthly word pool are different meters. Anonymous use is the daily cap. A signed-up free account adds the monthly pool so the allowance can follow you across devices. Unused words do not roll over. There is no invented rating, review count, or named testimonial on this site — those used to sit on the homepage, they were not real, and they are not coming back without countable consent.
What counts against my allowance?
The words in each scan you run. Sentence-level highlighting is included — it is scored in the same pass, so it does not cost you a second scan. Unused words do not roll over to the next cycle.
The humanizer is metered in words per UTC day, not in requests, because one long rewrite costs the same GPU time as several short ones. If a humanizer run fails the detector gate, those words are refunded rather than spent on a failure. Detector scans that return a report are spent; we do not refund a result you simply disliked.
How do I cancel?
From the billing portal linked in your account menu, at any time. Cancelling stops future charges and returns you to the free plan at the end of the period you have paid for. EU and EEA consumers also have a 14-day right of withdrawal — see the Terms of Service.
There is no phone tree and no "retention offer" flow. The portal is Stripe's. If the portal does not load, the contact page is one inbox, no published SLA, no contact form — including the reason we stopped putting a mailto on marketing CTAs (Cloudflare email obfuscation was 404ing it for crawlers).
Using it for real work
School, clients, and other people's writing.
What kinds of text work best?
Essays, articles, emails, reports, and marketing copy from about 50 words up to your plan’s per-scan limit (free is 2,500 words) give the clearest signal. Very short snippets don't carry enough rhythm to judge, and lists, code, or poetry will produce unreliable scores. If you are checking a mixed document — some passages written by a person, some by a model — the sentence-level report is the part that matters; a single document score will average them into something less useful.
Non-native English, heavily formal prose, and text written to a rigid template sit closer to the decision boundary. That is a property of the whole field, not a quirk of this scanner. Treat those results with more caution, read the highlights, and do not skip the conversation with the writer. Translated text that was then edited in English is a particularly hard case — fluent, even, and often human.
Can I use this for school or client work?
Follow the rules of your school, publisher, or client — some require disclosure of AI assistance regardless of how the final text reads. markhuman.ai is a checking tool: it helps you understand how a piece of writing reads, not misrepresent it.
Cleaner cadence does not make writing yours. If you cannot talk through the claims without the draft in front of you, use the detector — or write the thing. Teachers who need a process around a flag, not another "catch your students" pitch, have a dedicated page. Client work that went through a model and still has to ship is a good use of the sentence report: you can see which passages still read assembled and edit those, rather than rewriting the whole piece on a hunch.
Can I check someone else's writing?
You can, and plenty of people do — but only submit text you have the right to submit, and read the accuracy answers above before you act on the result. A score is a starting point for a conversation with the writer, not a substitute for one.
Student work, unpublished manuscripts, and confidential drafts are the cases where "the right to submit" is doing real work. Text is scored in memory and not stored, which is the privacy half of that problem; the ethics half is still yours. If you are a teacher, do not read a percentage aloud as an opening move. If you are an editor, do not treat AI-Assisted as a kill decision without a house policy that says so. If you are a hiring manager, a cover letter is a short passage — treat the result accordingly.
How to use this page
Start with the doubt you already have, then paste something
People land here because a result looked wrong, two tools disagreed, or someone is about to treat a percentage as a finding about a person. The jump list above is ordered that way: what the detector is, whether to believe it, what happens to the text, what it costs, and whether you are allowed to use it for the job you have in mind. Accuracy and limits is the section to read first if you are here in a hurry.
A report from this site is a probability estimate about a piece of text, not a finding about a person. The percentage on the dial backs the verdict beside it: for a Human-Written verdict it is how human the text reads, for an AI verdict how likely it is that a model wrote it, and it never points against the verdict. The raw, uncalibrated AI probability is in the report’s Details tab. The verdict is calibrated by length — the same score on a short snippet and a long essay is not the same finding. Sentence highlights show where the signal was strongest; they are not a list of guilty lines. Around 50 words is the floor; 50 free scans a day at up to 2,500 words each, no account, no card.
No detector is perfect, and disagreement between detectors is normal. False flags cluster on formal prose, technical writing, and fluent English written as a second language. Missed detections cluster on heavily edited or deliberately rewritten machine drafts. We do not publish a headline accuracy figure, or a head-to-head against GPTZero, Turnitin, Originality.ai, or anyone else, because a number we have not measured would be a claim, not a result.
If you are a teacher, the educator page is the process around a flag — triage privately, read against what you already know, ask for version history, act only on convergence. If you are holding a specific passage and want a check plus signs you can read by eye, that is the is-this-AI-generated page. If you wrote the draft and it still sounds assembled, the humanizer is a structural rewrite scored by the same detector; it is not a promise about anyone else's scanner.
Text is scored in memory and discarded when the response returns. It is not used to train a model. Support is one inbox on the contact page — no published SLA, no contact form. Plans and billing live in the terms. The privacy policy is the longer version of the storage answers above. None of those documents invent a user count, a star rating, or a named testimonial.
Two scans will teach you more than another paragraph: paste something you wrote yourself, then paste something you know a model wrote. The sentence shading is the part you can actually check. If the result on your own writing looks wrong, it may simply be wrong — tell us; confident misses are the most useful mail we get. If the result on the machine draft looks clean, that is the hard case, and the reason we refuse to paint AI-verdict sentences as human line by line.
Why the answer is often uncomfortable
The questions people actually search are the ones a reassuring FAQ would skip
A FAQ that only explains how to paste text ranks for nothing, because nobody searches for reassurance. They search for the doubt they already have. These four are the doubts this page exists to answer in the open.
A flag on writing you know is yours
False positives are real. Careful formal prose, technical writing, and fluent English written as a second language all sit close to the decision boundary. Read the sentence shading, then decide whether the result is evidence or noise. A score is not a character judgement.
Two tools, two answers
Detectors are not thermometers. Different training data, different thresholds, different ideas of machine writing. Disagreement is normal. Where they split, the text is genuinely ambiguous — look at the sentences, and talk to the writer if it matters.
Someone wants to use the report as proof
Do not. A probability estimate about a piece of text is not a finding about a person. Do not use it as the sole basis for a grade, a disciplinary outcome, or a hiring decision. Weigh it against drafts, version history, and a conversation.
The text was edited after a model drafted it
That is the most common real document, and it has its own verdict: AI-Assisted. Rounding it into a yes or a no is how both false accusations and missed drafts happen. The mixed case is a description, not a failure of the tool to make up its mind.
Easier to answer with your own text
Paste something you wrote yourself and something you know a model wrote in the checker at the top. Two scans will tell you more about what the score means than any answer above.
Not answered here? Email us from the contact page. If it is about what happens to your text, the Privacy Policy is more specific than the summary above, and plans and billing are covered in the Terms of Service.