Originality.ai

AI detection is the most contested category in this entire market, and for good reason. The claims are hard to verify, the consequences of error fall on individuals, and a great deal of nonsense is sold under the label.

Originality.ai is one of the more credible entrants, largely because it does something its competitors mostly avoid: it publishes its own error rates and links to independent studies rather than only to favourable ones.

This review takes the technology seriously and is equally direct about its limits. What follows covers how detection actually works, what the accuracy figures can and cannot support, the one application where these tools should never be used, and why the surrounding fact-checking and plagiarism tooling may matter more than the detector itself.

Detection, Stated Carefully

• Originality.ai is a content checking platform that estimates whether text was AI-generated and checks it for plagiarism, with fact checking, readability and grammar tools around those two core functions.
• It is built for publishers and agencies, not schools — team roles, bulk scanning, shared scan history and QA workflows for high-volume content operations.
• The detector is a modified BERT model covered by a US patent, with stated multilingual support across 30 languages.
• It publishes its own false positive rates and links third-party studies, which is markedly more transparent than the category norm.
• Charging is credit-based — pay-as-you-go credits with a two-year expiry, plus tiered plans; three free daily scans are available without an account.

How AI Detection Works, And Why It Is Probabilistic

Understanding the mechanism explains both the capability and the ceiling, and it is not complicated.

Language models generate text by repeatedly choosing likely next words. That produces prose with a statistical signature: word choices cluster around the probable, sentence lengths even out, and the distinctive unpredictability of human writing is smoothed away. Detectors are classifiers trained to recognise that signature.

This is why the output is a probability rather than a verdict. A detector is not finding a watermark or a hidden marker. It is estimating how much a passage resembles machine-generated text versus human text, based on patterns learned from examples.

Two consequences follow directly, and both matter. Some human writing is genuinely more predictable than average — formal academic prose, technical documentation, writing by non-native speakers working carefully within a limited construction set. That writing scores as machine-like because, statistically, it is.

And some AI text is deliberately made less predictable. Paraphrasing tools — including humanizers built for exactly this — exist precisely to disrupt the signature, which is why Originality.ai states it trains specifically on rewritten content from tools like Quillbot and Grammarly.

That is an arms race with no stable end state, and any detector's accuracy is a snapshot against the models and evasion techniques current at the time it was measured.

A detector does not find a marker. It estimates how predictable your sentences are — which is why careful, formal, non-native writing gets flagged.

The One Use These Tools Must Never Have

This needs stating as plainly as possible, because it is where real harm has already occurred.

An AI detection score must never be the sole basis for accusing a student, an employee or a writer of misconduct.

The arithmetic makes the reason obvious. Even a very low false positive rate produces a substantial number of wrongly flagged people at scale. A detector with a 1% false positive rate, run across 10,000 student essays, wrongly flags roughly 100 innocent students. Those hundred face an accusation that is extremely difficult to disprove — you cannot demonstrate the absence of assistance.

That burden is not distributed evenly, which is the more serious problem. Multiple studies have found that text by non-native English speakers is disproportionately misclassified as AI-generated, because the writing patterns that make prose read as non-native overlap with the patterns that make prose read as machine-generated.

Competing detectors such as GPTZero have published similar cautions. Several universities have withdrawn AI detection from their academic integrity processes for precisely these reasons, and some detection vendors have themselves cautioned against disciplinary use of their scores.

Originality.ai's transparency about false positive rates deserves genuine credit here. It is also exactly why those published rates should be read as a warning about how to use the tool rather than as reassurance.

The defensible use is as a signal that prompts a conversation, alongside drafts, version history and the writer's own account — never as evidence on its own.

• Legitimate: screening freelance submissions before a payment decision, with the writer able to respond
• Legitimate: publishers auditing content quality across a large library
• Legitimate: an agency confirming what it is delivering to a client
• Not legitimate: a disciplinary finding against a student or employee on the score alone
• Not legitimate: any use where the person flagged cannot see and contest the result

The Surrounding Tools May Matter More Than The Detector

The detector gets the attention. For most buyers the rest of the platform is where the recurring value sits.

Plagiarism checking is the mature, uncontroversial half. It matches text against existing published sources, which is a determinate comparison rather than a probabilistic estimate — either the passage appears elsewhere or it does not. For a publisher, that is often the more important check.

Fact checking is the most interesting addition, validating claims against cited sources. Whether content was AI-written matters much less than whether it is true, and confidently stated invented facts are the actual failure mode of generated content. A tool that surfaces unsupported claims addresses the real risk.

Dedicated editing tools like ProWritingAid go deeper here. Readability, grammar and guideline checking handle the editorial baseline, and the AI Allowance setting is a pragmatic touch — letting an organisation define an acceptable threshold rather than pretending a zero-tolerance policy is enforceable.

Deep Scan, which explains why a passage looks generated rather than only scoring it, is the right direction for the whole category. An explanation can be argued with; a number cannot.

The Chrome extension's writing replay, capturing keystroke history, is a genuinely different kind of evidence — it shows a document being composed over time, which is far more probative than any classifier output.

Credits That Expire

This site does not publish plan prices. Originality.ai's structure is worth understanding because it differs from most subscription tools.

Access is credit-based — you buy credits and spend them per scan, with credits stated to expire two years from purchase. There are also tiered plans aimed at different users, from students and individual writers through to agencies, publishers and enterprise.

There is a genuine free entry point: three daily scans without creating an account, with more available once you register. No free trial of the paid tiers is offered.

The credit model suits this category well, because content checking is bursty. A publisher auditing a back catalogue consumes an enormous amount in one week and almost nothing for months, which subscriptions handle badly.

Two things to plan for. The two-year expiry means bulk-buying credits far ahead of use has a real cost. And scans are capped at 2,000 words each, so a long article consumes multiple scans — worth factoring into any estimate of what an audit will cost.

One privacy point specific to the free tier: the vendor states text is not stored by default on free scans but may be used for training, with an opt-out available on paid use. If you are checking confidential or client material, that distinction matters more than the price.

How To Read Any Detection Score

Practical guidance, because the number is routinely over-interpreted in both directions.

A high AI score means the text is statistically similar to machine-generated writing. It does not mean the text was machine-generated, and it certainly does not mean the writer did anything wrong.

A low score means the text does not match the patterns the model was trained on. It does not mean the text is human — a capable writer editing generated output defeats detection almost trivially, which is the most common real-world workflow anyway.

Detection is therefore weakest exactly where it would be most useful: on the hybrid content that now makes up most professional writing. The binary question the tool answers has largely stopped matching how anyone actually works.

Longer passages are more reliable than short ones, because statistical signals need volume. A 200-word sample tells you much less than a 2,000-word one, and confident judgements on short text should be distrusted regardless of vendor.

The most useful framing is that a score is one input among several. Combined with plagiarism results, fact-check flags, editorial judgement and the writer's own account, it contributes to a decision. Alone, it supports nothing.

1. Scan writing you know for certain is human-written, including from non-native speakers on your team.
2. Scan writing you know is generated, including a version rewritten by a paraphrasing tool.
3. Note where the false positives land — that tells you who your policy would harm.
4. Set an AI Allowance threshold your organisation can actually defend rather than aiming for zero.
5. Write down, before deploying it, what happens when someone disputes a score.

Technology, Tools And Assurance

Terms That Decide How You Read A Score

Four definitions that make the difference between using this tool well and using it badly.

False positive

Human-written text wrongly flagged as AI-generated. The error that matters, because it lands on a person who cannot easily disprove it.

Perplexity

A measure of how predictable text is to a language model. Low perplexity reads as machine-like — which is also true of careful formal and non-native writing.

Paraphrase detection

Identifying AI text that has been rewritten to disrupt its statistical signature. An arms race rather than a solved problem.

Plagiarism check

Matching text against existing published sources. Determinate rather than probabilistic, and therefore a much stronger form of evidence.

Should You Be Running Detection At All?

Scan With It If

• You commission content at volume and need to know what you are paying for
• You publish at scale and want plagiarism and fact checking alongside detection
• You run an agency and need to confirm what is being delivered to clients
• You want an explainable result rather than only a percentage
• You will treat a score as a prompt for a conversation, not as proof

Never Use It To

• Discipline a student or employee on the basis of a score alone
• Screen writing from non-native English speakers without accounting for known bias
• Judge short passages, where statistical detection is least reliable
• Enforce a zero-AI policy, which is neither detectable nor realistic
• Check confidential material on a free tier without reading the data terms