Walter Writes AI detector review
Verdict
A free checker that names its signals and states no accuracy, false positive rate, dataset or test date, from the company that also sells the humanizer it is meant to check
- Best for
- A quick second opinion on a short English text when you already use Walter's humanizer and want the two tools in one workspace
- Main limitation
- No published accuracy or false positive rate, and the vendor's pages disagree with each other on whether checks are free and unlimited or a trial, and on 20 or 80 supported languages, read 9 November 2026
Plans and limits
| Plan | Price | Humanizer words a month | Words per request | Detector | Checked on |
|---|---|---|---|---|---|
| Free | $0, no card | 300 (one-off trial) | Not stated | Free checks on the detector page | 9 November 2026 |
| Starter | $8 a month, billed $96 yearly | 30,000 | 750 | Built-in | 9 November 2026 |
| Pro | $13 a month, billed $156 yearly | 70,000 | 1,500 | Built-in | 9 November 2026 |
| Elite | $26 a month, billed $312 yearly | 200,000 | 2,000 | Built-in | 9 November 2026 |
| Teams | $99 a month, billed $1,188 yearly, 10 members | 500,000 shared | 2,000 | Built-in | 9 November 2026 |
No page read states a word cap, a scan cap or a file size for the detector, which is unusual among the combined tools: Stealthly's pricing page, for comparison, gives its detector a 300-word monthly allowance on its free plan and 50,000 words on Basic, read 9 November 2026. A limit that is not published can change without a trace, so a reader who relies on free unlimited scans should treat that as today's terms.
The detector has no price list of its own. The headline on its page says free and unlimited, the FAQ on the same page says a free trial before subscribing, and the homepage pricing block lists a built-in AI detector on every plan, all read 9 November 2026. The humanizer plans below are the only published figures, and the detector rides on them.
What it did to our text
Does the Walter Writes AI detector work?
Walter Writes AI's detector is free on its own page and bundled into every paid plan from Starter at $8 a month billed $96 yearly, read on 9 November 2026, and it publishes no accuracy figure. This Walter Writes AI detector review covers what the vendor says the checker measures, what the pages leave out, and why a humanizer company's own detector cannot stand as evidence about that company's humanizer. The humanizer half of the product is reviewed separately in the Walter Writes AI review, and the other combined tools sit in the humanizer and detector list.
What does the vendor publish about its method?
The detector page describes three signals, read 9 November 2026: perplexity, which it defines as how predictable the word choices are; burstiness, the variation in sentence length and structure; and semantic flow, whether phrasing and transitions match human rhythm. It states that these feed a human-likeness percentage, that the tool does not look for hidden watermarks, and that the API returns per-paragraph confidence scores plus an overall AI probability. The models it names are ChatGPT, Claude and Gemini.
That is a method description in outline, and it matches what the better-known detectors publish about their own approach. GPTZero's technology page, for example, describes a sentence-level classifier, a "Paraphraser Shield" against paraphrased text, and a false positive rate it keeps at no more than 1 percent when evaluating AI against human text (GPTZero, How GPTZero works, read 9 November 2026). The difference is the last item. GPTZero states a rate; Walter's detector page states that no AI detector can guarantee 100 percent accuracy, that the engine minimises false positives, and gives no number for either.
The page also says the API engine was tested against outputs from several model families across short and long texts. No result of that test is on the page. A test with no published result is a sentence, not a measurement.
What does the research say a reader should expect?
Weber-Wulff and colleagues tested 14 AI text detectors on human, machine, paraphrased and translated text and found every tool below 80 percent on their accuracy measure, with performance falling further after paraphrasing (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Liang and colleagues ran seven detectors on TOEFL essays by non-native English writers and found that the detectors flagged 61 percent of them as AI-written (Liang et al. 2023, Patterns, read 9 November 2026). A checker built on perplexity inherits that risk, because plain, predictable prose scores as predictable whoever wrote it.
Walter's detector page markets the tool to educators for analysing student submissions and to researchers for verifying integrity, read 9 November 2026. With no false positive rate published, an educator has no way to weigh a Walter reading against those studies. No detector score should be the only evidence in a misconduct case.
Why is a vendor's own detector not evidence about its humanizer?
Because the same company built the text and the verdict. Walter's homepage tells users to humanize first and then check the score in the same workspace, read 9 November 2026. A humanizer tuned against its own checker will read as human on that checker whatever an outside detector says. The published method on this site therefore scores every output on five detectors no humanizer vendor owns, GPTZero, Originality.ai, Copyleaks, Pangram and ZeroGPT, and never on the vendor's own tool; the full rule set is on the method page.
| What Walter's detector page states (read 9 November 2026) | What it does not state |
|---|---|
| Perplexity, burstiness and semantic flow as signals | Thresholds, weights or how the three combine |
| A human-likeness percentage per scan | Any accuracy, precision or recall figure |
| Minimises false positives across named model families | A false positive rate, a dataset or a date |
| API tested on short and long texts from several model families | Any result of that test |
| Free and unlimited checks | Whether unlimited applies after the trial the FAQ describes |
What does the Walter Writes AI detector report?
A percentage, a label and, by the vendor's description, a section-by-section view. The detector page states that a scan returns an AI probability score and a detailed section-by-section report, and that uploaded documents get a score per section, read 9 November 2026. The API description adds per-paragraph confidence scores. Nothing on the page states the threshold at which a text is called AI rather than human, so the label cannot be compared with another detector's label on the same text.
| Output described by the vendor | Useful for | Missing |
|---|---|---|
| Overall AI probability percentage | A single number to compare across drafts | The threshold that turns it into a label |
| Section or paragraph scores | Finding the passage that moved the number | A published floor on text length for a stable reading |
| Model family hints (ChatGPT, Claude, Gemini) | Context | Any accuracy per model family |
Who should use the Walter Writes AI detector?
Use it as a free first pass if you already pay for Walter's humanizer and want to see whether a draft reads as predictable prose, knowing that the reading comes from the humanizer's own maker. Teachers and anyone deciding a misconduct case should not rely on it, or on any single detector, and should read do AI detectors work before acting on a score. Writers who want a check the humanizer vendor does not control should run the same text through at least one detector in the five-tool panel above.
Walter's detector is held to the same standard as every other vendor's on this site. Phrasly's detector page, for comparison, also claims a high accuracy figure but at least names a training corpus size, a 30-word minimum and four languages, and Undetectable AI's detector page cites a published study, both read 9 November 2026 and both covered in their own reviews. Walter's page names its signals and nothing else that can be checked.
The humanizer this detector is paired with is reviewed in the Walter Writes AI review, and the eight combined tools are set side by side in the humanizer and detector list. All reviews use the same fields; see the reviews hub.
Questions about Walter Writes AI detector
Is the Walter Writes AI detector free?
Yes on the detector page, which states free and unlimited checks, read 9 November 2026; the same page's FAQ calls it a free trial before subscribing, and every paid plan from Starter at $8 a month billed $96 yearly lists a built-in AI detector.
How accurate is the Walter Writes AI detector?
No figure is published: the detector page states that no AI detector can guarantee 100 percent accuracy and that the tool minimises false positives, with no rate, dataset or date attached, read 9 November 2026.
What does the Walter Writes AI detector measure?
Perplexity, burstiness and what the vendor calls semantic flow, combined into a human-likeness percentage, with per-paragraph confidence scores described for the API, all stated on the detector page read 9 November 2026.
Does the Walter Writes AI detector prove the Walter humanizer works?
No: the detector and the humanizer come from the same company, so a human label from Walter's checker on Walter's output is the vendor grading its own work, and only a detector the vendor does not own can count as evidence.
Which languages does the Walter Writes AI detector support?
The detector page states 20 or more languages, the homepage states more than 80 for humanizing and detecting, and one detector FAQ says any English-language content, all read 9 November 2026, so the figure depends on which page you read.
Related reading
WordAi reviewWordAi reviewed on the same fields as every humanizer we list: a 3-day trial, 50,000 words a month on.
Walter Writes AI reviewWalter Writes AI review on the same fields as every other humanizer: plans and the 750-word request cap.
Best AI detector and humanizer in one toolBest AI detector and humanizer in one tool: nine vendors that ship both, on the same dated fields, and why a.
How we test AI humanizersThe full humanizer test: 24 AI inputs and 6 human controls, five named detectors, a meaning rubric, published.