Find out what you actually published.
An Originality.ai alternative for editors: three verdicts, every sentence scored.
50 free scans a day · 2,500 words per scan · client copy is scored in memory and never stored
Scope
What this replaces,
and what it does not
Worth settling in the first thirty seconds rather than in the third email. These are two differently shaped products, and which one fits depends entirely on whether plagiarism is part of the same pass for you.
A bundled suite for publishers
It positions itself at web publishers, SEO teams, and content agencies, and it bundles plagiarism checking alongside AI detection — one pass over a document covering both questions.
An AI detector, and nothing else
There is no plagiarism check here, no similarity index, and no source matching. If a single pass over both questions is a requirement of your workflow, this is not a drop-in replacement for that half of it.
Depth on the one question
Three verdicts rather than one number, a score and a shade on every sentence, and a model explicitly trained on AI text pushed through commercial humanizers — which is what arrives when a brief says the copy must pass a checker.
In practice the two questions have come apart. Plagiarism is a question about provenance — has this text appeared somewhere else before — and it is largely solved. AI authorship is a question about production, it is not solved, and a model draft is original by every plagiarism measure ever built. A suite that answers both in one pass is genuinely convenient. It also means the harder of the two questions is competing for roadmap attention with the easier one. Running a dedicated detector for the hard half is a defensible trade, and it is the trade this tool asks you to make.
Editorial policy
Three verdicts map
to three decisions
Most content AI policies are one line long and binary: AI content is not accepted. That line has been unenforceable for a while, because the drafts arriving now are neither purely one thing nor the other. A report with a named middle category gives a policy somewhere to put them.
Human-Written
Nothing in the draft reads as machine-generated. For most contributor agreements this is simply the deliverable arriving in the state it was commissioned in.
AI-Assisted
A model drafted it and a person did real work on top, or the reverse. Most editorial AI policies are binary and have nothing to say about this, which is why they get argued over. Naming it turns an accusation into a conversation about disclosure.
AI-Generated
The draft reads as model output end to end, including copy that has already been run through a humanizer before it reached you. This is the case worth catching before it goes live under someone's byline.
The rewrite most editorial policies need is small: allow assistance, require disclosure, and reject wholesale generation. That version is enforceable because it matches how people are actually working, and because a contributor can comply with it honestly instead of running their draft through a paraphraser to get under a threshold. A three-way verdict is the reporting shape that policy needs — a single percentage forces you back into drawing an arbitrary line and defending it to whoever falls on the wrong side of it.
The audit pass
How to run it over
copy that is already live
Auditing published work is a different job from vetting an incoming draft — you cannot reject it, so the output has to be something you can act on with the writer who filed it.
Pull the body copy, not the page
Strip the nav, the boilerplate, and the CTA blocks. Templated furniture repeats across every post on a site and tells you nothing about who wrote the piece.
Scan section by section on long pieces
Each pass takes up to 2,500 words and needs at least 50. On a long guide, splitting by H2 is more useful anyway — it localises the problem to a section instead of averaging it across the whole article.
Sort by sentence shading, not by score
A piece with a handful of dark sentences in one section is a different editorial problem from a piece that is uniformly grey. The document score cannot tell those apart; the shading can.
Give the writer lines, not a number
A percentage invites a negotiation about the tool. Four specific highlighted sentences invite a rewrite. This is the single biggest practical difference in how a report lands with a contributor.
What we will not claim
No comparison table, no percentages
There are no head-to-head benchmark numbers on this page and there will not be. We have no third-party-verified accuracy figure for markhuman and none for Originality.ai, and a comparison table of invented percentages is worth exactly nothing to an editor making a real decision. No detector is perfect. Every result here is evidence to weigh — most cautiously on short passages, non-native English, and heavily edited copy — and none of it is proof on its own.
What is on this page is either a structural fact about the report — three verdicts, every sentence scored, nothing stored — or a description of what each tool is publicly positioned to do. The AI detector page covers the limits of the engine itself in more detail.
FAQ
Questions from editors
Plagiarism scope, batch auditing, how to raise a flagged draft with a freelancer, and what happens to client copy.
Does markhuman check for plagiarism?
No. It is an AI detector only — there is no similarity index, no source matching, and no plagiarism score. Originality.ai bundles plagiarism checking with its AI detection, so if your process depends on covering both in one pass, you will still need something for that half. We would rather say that on the comparison page than have you find out on your first real audit.
Can I audit a batch of posts or a whole site?
Not in one action. Scanning is document by document in the browser: 50 scans a day at up to 2,500 words each on the free tier, with larger per-scan limits and monthly volumes on the paid plans. For an archive audit, most editors sample rather than sweep — take the pieces from the contributors or the date range you actually have a question about.
What do I say to a freelancer whose draft comes back AI-Generated?
Lead with the sentences, not the verdict. Send the specific lines the report shaded most heavily and ask about those, because that is a question about the writing that a writer can answer. Treat the result as grounds for a conversation and a rewrite, not as proof of misconduct — no detector output is strong enough to carry an accusation on its own, and a contributor who is wrongly accused is a contributor you lose.
Will this tell me whether a search engine can detect the AI in my content?
No, and be wary of anyone who says their tool can. This reports on the text itself. What any search engine does or does not infer about a page, and how that feeds into ranking, is a separate question we have no visibility into. The defensible reason to run this over published copy is that you want to know what you actually published and under whose name.
Does it catch copy that has been through a humanizer or a paraphraser?
That is the specific case the model is trained for. Its training set includes AI text pushed through commercial humanizers and paraphrasers, mirrored against human originals, because a brief that says the copy has to pass a checker is exactly what sends a writer to one of those tools. It remains the hardest category, so treat a borderline result on visibly rewritten copy with more caution than a clean one.
Is client copy stored anywhere?
No. Text is scored in memory and discarded when the response returns — not retained, not logged as content, and not used for training. That matters more than usual in agency work, where the copy you are auditing is frequently under an NDA that was not written with detection tools in mind.
Start with the last piece you filed
Paste the body copy and read the sentence shading. Free — 2,500 words per scan, 50 scans a day, no account.