Winston AI reports 99.87% accuracy, but that figure comes from Winston’s own test, run on English documents of at least 500 characters. The independent evidence is thin: the one peer-reviewed comparison that names Winston used 24 texts, and it found the tool unreliable on Croatian. A high vendor number is a reason to try the tool, not a reason to treat its verdict as proof.
Key takeaways
- Winston’s headline figure is an in-house result on 12,000 English documents. It is not an independent audit.
- Independent testing of Winston is limited and small-scale, and its results do not transfer to every language or text type.
- Winston reports a score from 0 to 100 plus sentence-level highlights. Like any detector, it measures resemblance to patterns, not authorship.
- Use the highlights and the writer’s drafts, not the headline number, when a decision affects a person.
What does Winston AI claim about its accuracy?
Winston’s site presents “99.87% accuracy.” Its announcement of Winston 5.0 (published 2026-09-24) describes how that number was produced:
- Data: 12,000 English documents, split evenly into 6,000 human and 6,000 AI-generated texts.
- Human texts: 600 documents from each of ten categories, including books, college essays, news, poetry, recipes, Reddit posts and Wikipedia.
- AI texts: output from 20 generators across OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral and Qwen.
- Rule: a document counts as AI when its AI score is above 0.5. Texts under 500 characters were excluded.
- Outcome: 11,984 correct and 16 errors, which Winston breaks down as 7 false positives and 9 false negatives.
Winston calls this a held-out, in-house evaluation. The announcement does not say how the AI texts were produced, for example whether any were edited or paraphrased, so it tells you little about mixed or revised writing. It also covers only the English model.
Other Winston pages quote different numbers, such as 99.98% and 100% on a benchmark of 8,697 documents. These come from different datasets and conditions, so they are not the same claim stated twice. When two numbers on one site differ, the test conditions matter more than the number.
Is there independent evidence?
Some, but less than the marketing implies. Winston’s own help centre lists third-party studies, and it is worth reading them against the original sources, because a vendor summarising a paper tends to pick the flattering line.
The only study on that list with a direct head-to-head comparison is a pilot paper, Verification of AI-generated content (Information Research, 15 May 2026). It tested four detectors, Winston AI, Originality.ai, ZeroGPT and Smodin, on 24 English and Croatian texts written by three expert authors. Winston’s help page summarises it as a 99% standardized accuracy average with Winston ranked first. The article’s own abstract says something different in emphasis: detection was highly accurate for English but inconsistent or failing for Croatian, Originality.ai showed the highest overall reliability, and the authors conclude that tools need multilingual calibration. With 24 texts, neither ranking should carry weight.
Winston also cites a study of residency personal statements and a large study of consumer complaints that used Winston as a research instrument. Neither is a general accuracy comparison. The crowdsourced DetectArena leaderboard it cites measures which result voters prefer in blind pairings, not whether the label was correct.
Third-party blogs quote lower figures from other studies, but we could not verify their methods, so we do not repeat them. The fair summary is that no large, independent, current benchmark of Winston has been located, and that applies to most commercial detectors.
How does Winston report a result?
According to Winston’s site, a scan returns a score on a 0 to 100 scale, expressed as the odds that the text is human-written or AI-generated, and a sentence-by-sentence “AI prediction map” that colour-codes passages with AI-like characteristics. Winston’s materials for writers refer to a “human score.” Reports can be printed or downloaded.
Winston’s help page on false positives adds three points worth knowing. It publishes no cutoff for calling a score “human” or “AI”. It recommends scanning at least 300 words. And it states that a measured error rate applies only to the tested model, language and sample set, describing a detector score as “evidence” to weigh alongside drafts, revision history and sources. We think that is the right framing.
How should you read a human score?
A human score of, say, 85 is not an 85% chance that a person wrote the text. It is the tool’s output, scaled to 0 to 100, for how closely the text resembles what it learned as human. The same caution applies to every detector, including ours; what an AI detection percentage means explains the distinction.
| If you see | A reasonable reading | What to do next |
|---|---|---|
| High human score, no highlights | No strong AI-like pattern was found. This is not proof of authorship. | Treat as one signal. |
| Mid-range score, scattered highlights | Mixed or uncertain. Formal, translated or edited writing often lands here. | Read the highlighted sentences; ask for drafts. |
| Low human score, many highlights | The text resembles AI output. Could still be a false positive. | Look for other evidence and talk to the writer before concluding anything. |
| Any score on under 300 words | Too little text for a stable result. | Do not rely on it. |
| Any score on a non-English text | Reliability may be lower than the English figures suggest. | Check the vendor’s stated language support and test conditions. |
Winston and Originality.ai: how do their claims compare?
Paid detectors tend to market accuracy the same way: a high percentage, a low false positive rate, and a benchmark the company ran. The details differ, and the details are where you can compare them.
| Winston AI | Originality.ai | |
|---|---|---|
| Headline claim | 99.87% accuracy | 99.4% accuracy at its default 15% AI Allowance threshold |
| Test data | 12,000 English documents, 6,000 human and 6,000 AI | Its “Benchmark V6” slice of 456,872 binary samples |
| Who ran it | Winston, in-house | Originality.ai, in-house |
| Stated rule | AI if score above 0.5; texts under 500 characters excluded | Accuracy depends on the chosen threshold; stricter thresholds trade accuracy for precision |
| Checked | 2026-10-13 | 2026-10-13 |
Originality.ai’s own accuracy page also notes that its older Lite and Turbo models are being retired in September 2026, and that results from its two benchmark datasets are not directly comparable. The two vendors measure different things on different data, so these numbers cannot be ranked against each other, and we do not try to. Compare a tool on your own kind of text, not on a headline.
What do these claims not tell you?
A vendor benchmark can be accurate and still not match your situation. Four gaps recur.
Text type. A test set of essays, news and recipes says little about a technical report, a cover letter or an abstract. Winston’s help centre reports that, for an earlier model, human-text accuracy was lowest on Stack Overflow posts and poems.
Language. The 99.87% figure is English-only. The one independent paper that included Croatian found weak results there.
Edited or mixed text. Real writing is often drafted by a person, tidied by software, or revised from a model’s output. A test of cleanly human versus cleanly machine text does not cover that middle ground, which is where AI-assisted and AI-generated writing differ and where labels are least certain.
Who is affected by an error. Winston publishes no separate false positive rate for students or for non-native English writers. A very low overall rate can hide uneven error rates for a particular group. See AI detection false positives for what research on student writing has found.
So, is Winston AI accurate?
It may well be accurate on the English prose its test set resembles, and the company does describe its test design in some detail. What is not established is how it performs on your text, in your language, with your mix of human and machine input, as measured by someone with no stake in the result. For a general look at the evidence across tools, read how accurate AI detectors are.
What does Winston cost?
Winston’s pricing page, checked on 2026-10-13, lists a free plan with 2,000 credits over a 14-day trial, and paid plans of 100,000, 200,000 and 500,000 credits per month. Monthly list prices run from $18 to $49 before discounts, with annual billing and promotional codes shown on the page. Prices and credit rules change, so check the page before you buy.
Frequently asked questions
Is the 99.87% figure independent? No. Winston ran it internally on its own dataset of 12,000 English documents. It is useful as a statement of what the company measured, but it has not been replicated by an outside party that we could find.
Does Winston show which sentences are AI? Per its site, yes: an “AI prediction map” colour-codes passages with AI-like characteristics. Treat those highlights as places to read closely, not as findings.
Can Winston be wrong about human writing? Yes. Winston’s own test recorded 7 false positives in 6,000 human texts, and error rates are likely to be higher on short, formal, translated or non-native writing. No detector eliminates false positives.
Does Winston work in languages other than English? Winston lists many supported languages, but its headline test covers English only, and the one independent paper that included Croatian found inconsistent results. Check the vendor’s language claims against your own sample.
Before you rely on any score
Whichever tool you use, read the flagged passages, look at drafts and sources, and remember that a detector reports resemblance to patterns, not who wrote the text. Can AI detectors be wrong? covers the failure modes in more detail. If you already have the text, run it through our AI-generated text detector and review the flagged passages rather than relying on the overall score. How we label results, and why we publish no accuracy figure, is described on the accuracy page.