Is Originality.ai accurate?
Originality.ai rarely flags human writing in independent tests (0.11% to 0.25%), but it missed 37% of Claude Opus 4 text in a 2025 study and fell to 9.3% under a simple character-swap attack on RAID.
Short answer
Originality.ai is careful with human writing but misses some AI text. Two large independent tests found it wrongly flagged only 0.25% and 0.11% of human passages. The same 2025 study found it missed 37% of text written by Claude Opus 4, and a 2024 benchmark showed one character-swap trick cut its accuracy to 9.3%.
Is Originality.ai accurate? It is a paid detector aimed at publishers, agencies and SEO teams, and it shows up in more independent studies than most commercial tools. That gives a fuller picture than usual: low false positives in every large test, high accuracy on plain AI text, and clear blind spots. Every number below is dated and linked to its source, and vendor figures are labelled. For how we weigh sources, see how we score AI accuracy.
Originality.ai test results
- Independent study
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| Walters 2023, 126 documents | Documents classified correctly (0 wrong, 2 uncertain) | 98% | Source 1Independent study | |
| RAID benchmark paper (Dugan et al.) | Detection accuracy at a 5% false positive rate, no attack | 85.0% | Source 2Independent study | |
| RAID benchmark paper (Dugan et al.) | Detection accuracy at a 5% false positive rate, homoglyph attack | 9.3% | Source 2Independent study | |
| RAID benchmark paper (Dugan et al.) | False positive rate on human text at the default threshold | 0.25% | Source 2Independent study | |
| Jabarian and Imas 2025 | False positive rate on 1,992 human passages | 0.11% | Source 3Independent study | |
| Jabarian and Imas 2025 | Missed AI text written by Claude Opus 4 | 37% | Source 3Independent study | |
| Hadra and Mesbah 2026 | Macro-average accuracy on human, hybrid and AI texts | 0.69 | Source 4Independent study | |
| Originality.ai accuracy pageVendor's own figure, not independently checked | False positive rate, Lite model | 0.5% | Source 6Vendor claim |
What the Originality.ai accuracy numbers mean
Start with human writing, where the record is strong. The RAID benchmark paper ran commercial detectors on a large human set, and Originality.ai flagged 0.25% at its default threshold. Jabarian and Imas tested 1,992 human passages across six genres in 2025 and found a rate of 0.11%, lower than GPTZero's 0.71% in the same test. In Walters (2023), it made no wrong calls on 126 essays and was unsure about two.
The catch is the domain. When RAID's authors used one threshold for every kind of text, Originality.ai flagged 13% of human Wikipedia text as AI. A low average false positive rate does not hold for every type of writing.
On AI text, results depend on which model wrote it. Originality.ai reached 85.0% accuracy at a 5% false positive rate on RAID, ahead of GPTZero's 66.5% in the same paper. But Jabarian and Imas found it missed 37% of text written by Claude Opus 4, against 2.9% for GPTZero. On mixed human and AI writing, Hadra and Mesbah (2026) scored it 0.69 against Turnitin's 0.61, with near-zero recall on hybrid texts.
Originality.ai's own page claims a 0.5% false positive rate for its Lite model and 99%+ accuracy for its 5.0 model. Those are vendor claims on its own data. The independent tests above used earlier versions.
Where Originality.ai goes wrong
Character tricks
Swapping letters for look-alike characters (a homoglyph attack) cut its RAID accuracy from 85.0% to 9.3%. GPTZero held at 66.2% under the same attack.
Some AI models
It missed 37% of Claude Opus 4 text in 2025, so a low score on a newer model's text means little.
Mixed writing
In Hadra and Mesbah (2026), it almost never identified hybrid human and AI texts as hybrid.
Encyclopedic text
With one threshold for all domains, it flagged 13% of human Wikipedia text.
Non-native writers
It was one of seven detectors in the Liang et al. study, which averaged a 61.22% false positive rate on non-native English essays in 2023. The paper gives per-tool rates only in a chart.
How Originality.ai compares in the same tests
Only compare detectors inside one study.
| Test | Originality.ai | Others in the same test |
|---|---|---|
| Walters 2023, documents classified correctly | 98% | Copyleaks 100%, Turnitin 100%, ZeroGPT 87%, GPTZero 81% |
| RAID paper 2024, accuracy at 5% FPR, no attack | 85.0% | GPTZero 66.5% |
| RAID paper 2024, homoglyph attack | 9.3% | GPTZero 66.2% |
| Jabarian and Imas 2025, false positive rate on human passages | 0.11% | GPTZero 0.71% |
| Jabarian and Imas 2025, missed Claude Opus 4 text | 37% | GPTZero 2.9% |
| Hadra and Mesbah 2026, macro-average accuracy | 0.69 | Turnitin 0.61 |
Against GPTZero, the result depends on the test: Originality.ai flagged fewer human passages and scored higher on plain RAID text, while GPTZero caught far more Claude Opus 4 text and resisted the homoglyph attack. It edged Turnitin on mixed academic texts. Copyleaks was only in the Walters study. For the full picture, see are AI detectors accurate.
What an Originality.ai score can and cannot prove
No detector score should be the only evidence in a decision about a writer.
It can show
- An Originality.ai score can tell you that a text reads like model output to its detector. Because its false positive rate was low in large independent tests, a high score on plain prose is a meaningful signal.
It cannot prove
- It cannot tell you who wrote the work. It also cannot clear a text: it missed more than a third of one current model's writing, and simple character tricks defeat it, so a low score does not prove a human wrote it.
What to do if Originality.ai flags your work
Ask for the full report, including the score and which passages were flagged.
Collect your drafts, outline, notes and version history.
Note the kind of text: encyclopedic, technical or formulaic writing was flagged more often in the RAID tests.
If English is not your first language, cite the Liang et al. findings on non-native writers.
Offer to walk through your research and sources, or to write a short sample.
Ask that any decision rest on more than one piece of evidence.
Another detector to try
Another option is the free AI detector from Walter Writes. Its page says checks are free and unlimited, that it detects text from ChatGPT and Gemini, that it gives a percentage score for every scan and that it works in more than 20 languages. You can try Walter's AI detector here. A second detector's score is one more signal, not proof either way.
Sources
- Independent study
- Independent study
- Independent study
- Independent study
- Independent study
- Vendor claim
Questions
How accurate is Originality.ai?
Accurate on plain AI text, with clear gaps. It scored 98% in a 2023 study of 16 detectors and 85.0% at a 5% false positive rate on RAID. It missed 37% of Claude Opus 4 text in 2025 and fell to 9.3% under a homoglyph attack.
Does Originality.ai give false positives?
Rarely on general prose in independent tests: 0.25% on RAID and 0.11% on 1,992 human passages. The rate rose to 13% on human Wikipedia text when one threshold was used for every domain. Originality.ai itself reports 0.5% for its Lite model.
Is Originality.ai better than GPTZero?
It depends on the test. In the same studies, Originality.ai flagged fewer human passages (0.11% against 0.71%) and scored higher on plain RAID text (85.0% against 66.5%). GPTZero caught far more Claude Opus 4 text (missing 2.9% against 37%) and held up under the homoglyph attack.
Can Originality.ai be fooled?
Yes. Replacing some letters with look-alike characters cut its accuracy to 9.3% on RAID. Paraphrasing did not fool it in that test, where its score rose to 96.7%, but text from some newer models slipped past it 37% of the time.
Should I rely on an Originality.ai score?
Not on its own. A high score on ordinary prose is a stronger signal than with most detectors, because its false positive rate was low in large tests. A low score proves little, because it misses some AI text. Check drafts and sources before deciding anything.
