Does Originality.ai Detect AI Writing? What the Score Really Means [2026]

If you write for clients, Originality.ai is probably the detector you will meet first. Students hear about Turnitin. Marketers, agencies and freelance writers hear about this one, usually in an email that starts "quick question about the draft you sent."

So: does it detect AI writing? Often, yes, particularly when the text came straight out of a model with no editing. But the more useful question is what its number actually represents, because a lot of professional arguments happen over a percentage that both sides are reading wrong.

What Originality.ai is

It is a paid AI-content detector built for publishing workflows rather than classrooms. That focus shows up in the feature set: it pairs AI detection with a plagiarism check, it can scan published URLs so you can audit a whole site, it has team seats and shareable reports so an editor can forward a result to a writer, and it offers a browser extension that inspects a Google Doc's revision history to see how a draft was actually produced.

Scans are metered, so unlike some consumer detectors there is no practical way to check a long draft for free. Pricing and plan structure have changed more than once, so check their site rather than trusting a number in a blog post.

One thing worth knowing before you argue about a result: the product has shipped several generations of its detection model, and they do not agree with each other. The same paragraph can score differently depending on which model version an account is set to use. If a client sends you a screenshot, the model matters as much as the number.

How the score is produced

Originality.ai uses a trained classifier rather than a rulebook. Somebody assembled a large corpus of text known to be human-written and text known to be machine-generated, then trained a model to tell the two apart. When you paste a draft in, the classifier estimates which of those two piles your text most resembles.

What it is picking up on is roughly what every detector picks up on:

• Predictability. Language models pick likely next words. Text made of likely next words has a statistical signature that classifiers learn to recognise.
• Uniformity. Machine drafts tend toward sentences of similar length and similar construction. Human writing is lumpier.
• Register and phrasing. The vocabulary and connective tissue that show up disproportionately in model output become features the classifier learns.

Notice what is absent from that list: any knowledge of who typed the words. A detector has no access to authorship. It reads the finished text and reports a resemblance. That distinction is the whole ball game.

What "40% AI" actually means

Here is the misreading that causes most of the trouble. When a report says something like 40% AI, people assume it means two out of five sentences were written by a machine. In most detector reports the headline figure is a confidence estimate about the document, not the share of the text that is machine-written. It is the tool saying how strongly the text resembles its AI training pile, expressed as a percentage.

Read the label in the report you were actually sent before you argue about it. Some views show a document-level confidence, some highlight individual sentences or sections. They answer different questions, and a screenshot of one gets quoted as though it were the other constantly.

The practical consequence: a mid-range score is not "some of this is AI." It is closer to "this tool is not sure." A detector that is not sure has told you almost nothing, and it certainly has not produced evidence of anything.

Why honest writing gets flagged

Every classifier trades false negatives against false positives, and the false positives land on real people.

The best-documented case is the one we have written about before: a Stanford study found that a majority of essays by non-native English speakers were misclassified as AI-generated by common detectors. Second-language writing tends toward simpler vocabulary and steadier sentence structure, which is exactly the pattern detectors read as machine-like.

The same logic reaches into professional writing. If you produce clean, well-organised copy for a living, you have trained yourself into habits that look statistically tidy: consistent paragraph shapes, controlled sentence length, careful transitions. Good technical writing, standard product copy, and any prose written to a house style all carry more false-positive risk than a rambling personal essay does. Being a disciplined writer is a risk factor, which tells you something about what the tool is really measuring.

Vendors publish accuracy figures for their own products. Treat those the way you would treat any self-reported benchmark, and pay more attention to the false-positive rate than to the catch rate, because the false-positive rate is the number that decides whether an honest writer loses a client.

When a client sends you a flagged report

You open your email and there is a screenshot with a red percentage in it. What now?

Do not lead with an apology. An ambiguous score is not an accusation you need to answer for, and treating it like one concedes the point.

Ask which tool and which model. A specific, calm question changes the register of the conversation. Different detectors and different model versions of the same detector disagree on the same text.

Show your process. This is the most effective thing you can do, and it works better than any argument about statistics. Version history in Google Docs or Word, an outline, research notes, a list of the sources you actually read, a first draft that is visibly worse than the final one. Process is evidence. A detector score is not.

Explain what the number is. Politely: the figure is a resemblance estimate, not a measurement of authorship, and mid-range scores mean the tool is uncertain. Most clients have never been told this and are relieved to hear it.

Agree on a policy, not a threshold. The durable fix is a written line in your contract about how AI may be used in the work, and an agreement that no detector score is treated as proof on its own. Getting that into a contract before the dispute is worth more than winning the dispute.

Editing that genuinely moves the number

If your text does resemble model output, whether because you drafted with AI or just write very tidily, the edits that lower a detector score are the same edits that make the writing better to read.

1. Break the rhythm. Uniform sentence length is the single loudest signal. Cut one sentence to four words. Let the next one run. Variation is what the tool is failing to find.
2. Cut the model's vocabulary. Search for delve, crucial, leverage, robust, seamless, navigate, foster, underscore, tapestry. Replace each with the plainer word you would have used in conversation, or delete it.
3. Replace placeholder examples. "A company saw improved engagement" is a slot the model left empty. Fill it with a real client, a real number, a real week when something went wrong.
4. Use contractions where the register allows. Formal-by-default is a machine habit more than a professional one.
5. Break the perfect structure. Not every paragraph needs a topic sentence and a tidy close. Let one start in the middle of a thought.
6. Add something only you know. A specific detail from your own experience is the one thing no model can generate and no classifier expects.

That list is the manual version. It takes real time on a long draft, which is why MakeItHuman exists: it applies the same kind of restructuring in seconds while keeping your meaning and your citations intact. The free tier handles 300 words a day, which is enough to see what it does to a section of your own writing before you decide whether it belongs in your workflow.

We will not tell you it makes anything undetectable, and you should be wary of any tool that does. What we can point to is a benchmark we publish because it can be checked: against Binoculars, a peer-reviewed open-source detector, 88% of humanized texts score on the human side, averaging 82% human on a scale where genuine human writing scores 94%. That is an open detector, not a commercial one, and we quote it precisely because you can verify the methodology.

FAQ

Is Originality.ai accurate?

It performs well on unedited model output and, like every detector, worse on edited, mixed, or unusually clean human writing. Its accuracy is also not one number: it varies by model version, text length, and subject matter. Vendor-published figures are marketing claims, not independent results.

Can Originality.ai detect Claude or Gemini as well as ChatGPT?

It is trained on output from many models, not one, so it is not ChatGPT-specific. What matters more than which model produced a draft is how much editing happened afterwards.

Does Originality.ai detect paraphrased or rewritten AI text?

Sometimes. Light synonym-swapping usually leaves the underlying rhythm and structure intact, which is what the classifier reads. Substantive rewriting that changes sentence shape and adds specific content changes the signal much more.

Is a low score proof that I did not use AI?

No, and a high score is not proof that you did. The tool reports resemblance, not authorship, in both directions. That is why documented process beats any score in a dispute.

Should freelancers run their own work through a detector before delivering?

It is reasonable, if you treat it as a smell test rather than a verdict. A high score on your own unassisted writing is useful information: it usually points at flat rhythm and generic examples, and fixing those makes the piece better regardless of what any detector thinks.

The bottom line

Originality.ai does catch raw AI writing a good deal of the time, and it is a competent tool for the job it was built for. What it cannot do is tell anyone who wrote something. It reports how much a piece of text resembles the machine-written examples it was trained on, which is a genuinely different claim, and one that gets flattened into an accusation somewhere between the report and the client's inbox.

Know what the number means, keep your drafts and your notes, and write with enough rhythm and specificity that the question rarely comes up. If you want the editing pass done faster, try the humanizer or look at what the plans include.