How Originality.ai detects AI writing, what its score means, and what to do if it flags you

What Originality.ai is, and who uses it

Originality.ai was founded by Jon Gillham after he sold a content-marketing agency, and it launched in November 2022 to detect GPT-3 text, shortly before ChatGPT appeared. The company is based in Collingwood, Ontario. It sells to "Website Owners, Content Marketers, Writers and Publishers", and its Site Scan feature, which crawls a whole website on a schedule, says plainly that it is "Not for Students".

That has changed at the edges. Originality.ai has since added an education page with per-word pricing for schools, a Moodle plugin and, in August 2026, Google Classroom support through its Chrome extension, so students now meet it too. Its product menu covers an AI checker, a plagiarism checker, a fact and hallucination checker, grammar and readability checks, bulk scans and an API on its Enterprise plan. It does not offer a humanizer.

If you are a freelance writer, the practical point is that the person running Originality.ai on your work is usually a client or an editor, not a teacher. That shapes what you can do when it flags you, which is covered below.

Its detection models in 2026

In September 2026 Originality.ai retired its classic models, Lite, Turbo and Academic, which most older guides still describe. Two remain: Multi Language, for 30 languages, and AI Allowance, which became the default in mid-2026.

AI Allowance asks whoever runs the scan how much AI use is acceptable, then judges the text against that line. The levels, in Originality.ai's words:

• 0%, "Very Strict".
• 5%, "Strict": "Highly sensitive. Will flag even minor AI edits."
• 15%, "Moderate": "light AI editing allowed, like grammar or spelling". Originality.ai says this level is "best for most use cases" and uses it as the default.
• 25%, "Flexible": "only flags heavy AI usage."
• 40%, "Generous": "Hybrid AI writing is permitted (but not fully AI-generated content)", described as "near-zero false positives."

Why the allowance level matters to you

Originality.ai's help centre explains that the allowance "does not change the underlying detection technology. Instead, it adjusts how scan results are interpreted." The same draft can therefore get a different verdict at a different level. If a client tells you a piece "failed Originality", the first question to ask is which allowance level they used, and whether the two of you ever agreed one.

What the score means, and what it does not

This is the most misread number in AI detection. Originality.ai's own example: "Likely Original, 60% Confident" means the model thinks the content is human-written and is 60% confident in that prediction. It "DOES NOT mean that 60% of it was Original and 40% was AI-generated." In the same way, a 90% AI score means the tool is 90% confident AI was used in some part of the text, not that AI wrote 90% of it.

The report also highlights passages: dark green for sections that look original, dark red for sections that look AI-generated. Originality.ai says the patterns behind the highlighting "cannot be translated into a simple list", so treat the colours as a pointer to where the model reacted, not a list of sentences to fix. For how scores and thresholds work across detectors in general, see what a detector score means.

Scans on the website need at least 100 words (50 through the API) and accept up to 50,000. You can paste text, upload Word or PDF files, or scan a URL.

Why it flags human writing

Originality.ai is unusually open about its false positives. Its help centre puts them at "about 1.56% of cases" (a figure from before AI Allowance) and lists the usual causes: using editing tools, short content, formulaic content, academic content, and public-domain works. On editing tools it goes further than most: "Using any tool whether it's Grammarly, ChatGPT, Quillbot, or others will raise your AI score", and "even Microsoft Word Editor has become powerful enough in recent years to trigger our AI detector." In its view, human text that was heavily edited with AI counts as AI, not as a false positive.

Independent tests found a wide range, depending on the text and the model version. A 2023 Stanford study tested seven detectors, Originality.ai among them, on essays that non-native English speakers wrote for the TOEFL: on average the detectors labelled 61% of those essays as AI, while classifying essays by US eighth-graders almost perfectly. The 2024 RAID benchmark found a 13% false-positive rate on Wikipedia text once Originality.ai's threshold was set for 5% overall. A 2025 working paper by University of Chicago researchers found rates below 1% for the then-current Turbo model, and Epoch AI in July 2026 found 19 of 495 human passages flagged (3.84%), most of them fiction and all written before ChatGPT existed. None of these studies tested the current AI Allowance model.

Originality.ai names public-domain works among its own causes of false positives. Our weekly benchmark shows the same failure in a different detector: famous human passages, written long before AI, scored as likely AI. Why real writing gets flagged explains the mechanism.

If a client or a teacher flags your work

Originality.ai publishes advice for this situation, and it is better than most. Its central line, for anyone holding the score: "No AI detector should be used as the sole basis for a disciplinary, employment, publishing, or academic decision." It also tells clients not to have "a HARD rule that every piece of content MUST meet a certain threshold." Quote both if you need to.

• Ask which model and allowance level was used, and for the full report, not a screenshot of the number.
• Check what your contract or brief said. Originality.ai suggests confirming whether AI was prohibited and whether a specific score and tool were agreed; it calls agreeing a score "not a recommended way to manage the misuse of AI".
• Show how the piece was made. A Google Docs or Word version history is the strongest evidence you have, and Originality.ai's free Chrome extension can replay a document's creation character by character. Its help centre notes that an incomplete history "does not prove AI use".
• Offer earlier work. Originality.ai suggests comparing three to five writing samples by the same author, and running previously submitted work through the detector too.
• If you used Grammarly or a similar tool, say so. By Originality.ai's own account, that alone can raise the score.
• Keep citations and quotations. Originality.ai tells reviewers to check citations separately: "Do not remove citations simply to obtain a lower score."

The cost of a false flag is real

In 2024 Gizmodo reported freelance writers who lost work after Originality.ai flagged writing they said was their own. Gillham told the reporter: "We hear these stories more than we wish we did, and we understand the pain that false positives cause writers." In September 2026 the company started paying $100 in credits to paying customers who submit verifiable human writing that it misclassifies. For students, if you are accused of using AI sets out the steps in an academic setting.

Where AntiGPT fits, honestly

Originality.ai and AntiGPT get compared because both show an AI score, but they do different jobs. Originality.ai checks content you receive, for AI, plagiarism and facts, and costs about $0.75 per 10,000 words scanned on its Pro plan; it does not rewrite anything. AntiGPT rewrites drafts you write and checks them as you go, from $11.99 for 10,000 words, payable by card or, in Ghana, by mobile money.

AntiGPT is a rewriting tool with its own detector, Hello-SimpleAI's open RoBERTa classifier. It is not Originality.ai, its score will not match Originality.ai's, and Originality.ai says it actively tests tools that rewrite text and treats heavily AI-edited writing as AI. If your contract says no AI, a rewriting tool does not change that: agree terms with your client and disclose what you used.

Where AntiGPT helps is the writing itself. If you draft with AI help where that is allowed, it rewrites the draft in your voice, shows a before-and-after estimate, and lets you redo single sentences that still read like a template. The habits that make writing sound like a person (specific detail, varied sentence length, an actual opinion) are the same ones editors pay for; our guides for blog posts and marketing copy go through them, and the free sentence variety checker shows one of the signals in your own text.