Detection Study
How many typos can TypoSweep catch?
Ten complete public-domain novels were seeded with 1,200 errors, then scanned in full.
Abstract
Background: Existing word processing software tools are highly capable of catching misspelled words and simple grammar errors; however, they often fail to catch context-dependent typos (e.g., bare vs bear). They also push prose toward a traditional and plain writing style, sometimes aggressively rewriting in a way that obliterates authorial intent, or even authorship itself. TypoSweep is an AI-assisted proofreader that aims to respect writing style, suggest minimal revisions, and flag only objective writing errors.
Methods: A total of 1,200 errors were randomly planted across ten unabridged classic novels, about 15 from each of eight error types per book. Seven were context errors spellcheck cannot flag: homophones (then for than), real-word substitutions (form for from), tense slips, subject-verb mismatches, apostrophe errors, doubled words, and dropped words. The eighth was plain misspelling, the one type spellcheck already flags. TypoSweep analyzed each book and generated an error report. These findings were compared with the answer key to calculate the accuracy of error detection, including the rate of false positives. Error catch rates were subdivided by type to characterize the tool's strengths and weaknesses as a proofreader.
Results: TypoSweep caught 87.2% of all typos, with clear strengths and weaknesses across error types. The catch rate for context-dependent errors, which spellcheck typically fails to detect, was 91%. Conversely, TypoSweep detected only 60% of misspellings, as it is designed to be used after spellcheck is complete.
Conclusion: TypoSweep is a strong second reader for the subtle, valid-word errors that slip past spellcheck. This tool is not a replacement for spellcheck or a human proofreader, but rather a rapid tool for catching typos frequently undetected by human readers and existing software. The sections below explain why.
1. Background
Spellcheck reliably catches misspelled words and simple grammar errors, but it often misses context-dependent typos: a real word used in the wrong place, a slipped tense, a dropped function word. These errors are now the most common mistakes in finished writing, precisely because spellcheck has all but eliminated plain misspellings.2 They also survive a careful human read, because the sentence still looks correct on the page. TypoSweep is an AI-assisted proofreader built to catch these residual errors while leaving the author's style intact. This study measures how many errors TypoSweep catches, categorizes them by type, and compares the results with published benchmarks for human readers.
2. Methods
2.1 The corpus
Ten complete public-domain novels were tested, 1,407,757 words in total, with every word of each book scanned, including all 560,809 words of War and Peace. From longest to shortest, they were War and Peace (Tolstoy, 1869), Moby-Dick (Melville, 1851), Dracula (Stoker, 1897), Pride and Prejudice (Austen, 1813), The Adventures of Sherlock Holmes (Doyle, 1892), The Age of Innocence (Wharton, 1920), The Picture of Dorian Gray (Wilde, 1890), The Great Gatsby (Fitzgerald, 1925), The Wonderful Wizard of Oz (Baum, 1900), and A Christmas Carol (Dickens, 1843).
Texts came from Project Gutenberg, stripped of front and back matter and scanned in full. For War and Peace, the French-language passages were removed so the test measured English proofreading only.
2.2 Planted errors
Each book received 120 planted errors, about 15 from each of eight categories (roughly 150 per category across the full corpus), spread evenly through the text and logged in advance so detection could be scored against an answer key the tool never saw. The eight categories appear below with an example of each, the incorrect text in red and the correction in green.
| Homophonea great deal more then than that | Real-word swapask him where form from |
| Tense / verb formshe had went had gone down | Subject-verb agreementnow he were he was a sturdy man |
| Apostrophe / contractionin it's its deep gloom | Doubled wordarrange a a a marriage |
| Missing wordreturned to the ranch | Plain misspellingplease recieve receive it |
Proofreading research identifies these eight kinds of typos as distinct error classes, and together they cover the most common mistakes that survive into a finished manuscript. Seven of the eight errors slip past spellcheck because they are spelled correctly. The grammar checkers in Word and Google Docs catch some of them, but inconsistently, leaving many typos behind.
Splitting planted errors evenly across all eight categories keeps any one error type from dominating the analysis, so the overall accuracy rate reflects the full spread of mistakes rather than whichever kind is easiest to catch.
3. Results
Across the full corpus, TypoSweep detected 87.2% of the 1,200 planted errors. By book, detection ranged from 75.0% (A Christmas Carol) to 94.2% (The Age of Innocence). It varied more sharply by error type, as the sections below show.
3.1 Detection by book
| Novel | Author (year) | Words | Errors | Detected |
|---|---|---|---|---|
| War and Peace | Tolstoy, 1869 | 560,809 | 120 | 86.7% |
| Moby-Dick | Melville, 1851 | 164,595 | 120 | 83.3% |
| Dracula | Stoker, 1897 | 160,804 | 120 | 91.7% |
| Pride and Prejudice | Austen, 1813 | 121,566 | 120 | 93.3% |
| The Adventures of Sherlock Holmes | Doyle, 1892 | 104,337 | 120 | 87.5% |
| The Age of Innocence | Wharton, 1920 | 101,338 | 120 | 94.2% |
| The Picture of Dorian Gray | Wilde, 1890 | 78,503 | 120 | 89.2% |
| The Great Gatsby | Fitzgerald, 1925 | 48,143 | 120 | 84.2% |
| The Wonderful Wizard of Oz | Baum, 1900 | 39,260 | 120 | 86.7% |
| A Christmas Carol | Dickens, 1843 | 28,402 | 120 | 75.0% |
| Ten novels | 1,407,757 | 1,200 | 87.2% |
3.2 Detection by error type
Detection rates pooled across all ten novels, grouped by error family. The tool did best where spellcheck and a quick human read do worst.
Share of planted errors detected, by category, across 1.4 million words. Green marks the kinds of error TypoSweep is built to catch. Amber marks the two it catches least often.
3.3 Comparison with human readers
TypoSweep detected 91% of real-word and grammar errors, compared with 66% for human readers under controlled conditions.1 The two metrics are not exactly comparable, since those readers were measured on a single timed pass, and a professional working over days would catch more. But speed is partly the point. TypoSweep does in seconds what would otherwise take days of tedious work, and it catches the errors humans most often miss.
On plain misspellings, the results reverse, because spellcheck already detects those mistakes. TypoSweep expends minimal attention on spelling errors because this improves its ability to detect the kinds of typos that most frequently slip past spellcheck and human proofreaders.
No proofreading process, human or automated, catches everything.1
Detection of each error family, TypoSweep compared with a human proofreader. Human rates are the weighted mean across published proofreading experiments.1
What about Grammarly? One independent peer-reviewed study found that Grammarly flagged only 51% of the errors detected by human readers.3 That test involved a different kind of writing material (i.e., not full-length novels), so it offers context rather than a clear head-to-head comparison. Still, it points the same way: catching errors is harder than correcting them, and no single tool or reader can catch them all.
3.4 Precision and false positives
Detection is only half of proofreading. The other half is precision: leaving correct writing alone. A false positive is a flag on a passage that was already correct. Across all 1,407,757 words, TypoSweep produced only 81 false positives, roughly one for every 17,000 words, or one every forty pages.
Some flags fell outside the planted set. On review, most turned out to be real errors already sitting in these century-old texts: an archaic verb form, or a stray character left behind when the book was digitized. Those are correct catches, not mistakes by the tool. The one recurring source of wrong flags was intentional dialect. In Dracula, for instance, Van Helsing's deliberately broken English was flagged as error. Software cannot reliably tell a character's voice from a slip, so TypoSweep leaves those judgment calls up to the author.
4. Discussion
4.1 Limitations
Missing words (69%). Missing words are challenging for software and humans to detect because nothing on the page appears wrong on first impression, and many human readers will subconsciously supply the missing word, so the sentence retains its intended meaning. This is TypoSweep's largest gap, and the clearest reason a careful human proofreader is still essential.
Plain misspellings (60%). This apparent shortcoming is in fact a design choice. TypoSweep assumes the author has already run spellcheck, so it spends little effort re-finding such errors. In addition, maintaining focus on other error types increases TypoSweep's accuracy for all but misspellings. This tool is built to be used after spellcheck, and is not a substitute for it.
4.2 What this means for your manuscript
TypoSweep catches the errors that escape everything else. The grammatically plausible mistakes that spellcheck cannot see, and that humans often read straight past, were detected at a rate of 91%, well above the 66% catch rate reported for a single human pass.1
TypoSweep does not catch everything, because no reader does. No process, automated or human, is complete,1 and missing words are particularly challenging to detect. Expect TypoSweep to sharply reduce errors but not eliminate them entirely.
TypoSweep is a complement, not a replacement. Spellcheck detects misspellings, TypoSweep catches context errors, and a human handles edge cases and authorial intent. Used in that sequence, TypoSweep does the tedious pass faster and more consistently than a person, allowing a human editor to focus on the work only a human can do.
5. Conclusion
Across 1.4 million words and 1,200 planted errors, TypoSweep caught 87.2% of all errors and 91% of the context errors it is built to catch, at a cost of roughly one false positive every forty pages. It is a complement to spellcheck and a human proofreader, not a replacement, and is most effective when used between spellcheck and human proofreading. This study reflects the detection configuration shipping as of May 2026, and reports the catch rates for all evaluated error types, including the categories where the tool performs worst.
References
- Panko RR. Human error in proofreading for spelling errors. The Human Error Website. 2014. Accessed May 2026. https://panko.com/HumanErr/Proofreading.html. Synthesizes published proofreading experiments (Cohen, 1980; Daneman & Stainton, 1993; Levy, Di Persio & Hollingshead, 1992; Riefer, 1991; Wallace, 1991, among others); weighted-mean detection was 81% for non-word (misspelling) errors and 66% for word errors.
- Lunsford AA, Lunsford KJ. "Mistakes are a fact of life": a national comparative study. Coll Compos Commun. 2008;59(4):781-806.
- Koltovskaia S, Saeli H, Rahmati P. An inquiry into Grammarly's precision and recall: a comparative study with human annotators. J Second Lang Writ. 2026;71:101283. doi:10.1016/j.jslw.2026.101283.
Source texts: Project Gutenberg. The study reflects the shipping detection configuration as of May 2026.