Spellcheck, autocorrect and the writing tools before AI

Spellcheck, autocorrect and the writing tools before AI are older than most of the people using them. Les Earnest put a word-list spelling check into a handwriting reader at MIT in 1961, and Ralph Gorin's SPELL was suggesting corrections at Stanford by 1971 (Earnest, 2011; Peterson, 1980). Readability scores are older again: Rudolf Flesch published his in 1948.

EdCitation publishes this history from the programs' own papers and manuals, makers' pages, patents and peer-reviewed studies, each read on 26 September 2026, and any we could not open is marked. Machines that write have their own guide, a short history of machine writing. This one is about the tools that only checked.

Each of them judged the surface of a text: letters, words, sentence length, layout. None asked whether a cited source exists. EdCitation's free Verify references takes that step against publishers' records; Find sources found the studies below and Cite a source built their entries.

When was the first spell checker made?

The earliest spelling checker we could trace ran in 1961, as a subroutine of Earnest's handwriting recogniser at MIT that kept only candidate words found on a list (Earnest, 2011). Research on correcting misspellings by computer goes back to 1957, says Peterson (1980), much of it for character readers, Morse code and airline passenger names rather than prose.

From a word list to SPELL

Earnest took his list, 10,000 English words on seven reels of paper tape, to Stanford's Artificial Intelligence Laboratory in 1965. In 1967 a graduate student whose name he no longer recalls wrote a LISP checker that listed unknown words by page and line and offered no fix. In 1971 Ralph Gorin wrote SPELL, which looked for real words one letter away and asked the user about each unknown word.

Earnest calls his 2011 account a sketch, and we found no 1967 paper to confirm the middle program. Peterson (1980) calls SPELL for the DEC-10 the first spelling checker written as an application rather than as research. Gorin's 1974 manual, reprinted by Earnest, shows a habit autocorrect later made universal: a misspelling the user fixed once, SPELL fixed everywhere after.

Why early checkers hunted four mistakes

Damerau (1964) assumed a word missing from the dictionary holds one error: a wrong, missing or extra letter, or two letters swapped. On garbled test text his method identified over 95% of such errors, and Peterson (1980) reports Damerau's estimate that these four kinds are 80% of all misspellings. UNIX's TYPO tried another route, ranking each word by how rare its letter pairs and triples were in the document.

The mistake no word list catches

A slip that forms another real word passes any list. MacArthur et al. (1996) gave ten common spelling checkers 555 misspellings from students with learning disabilities in grades 5 to 8. The checkers missed 26% and 37% of errors in two studies because the error was itself a word, and suggested the right spelling for about 55% of the rest; students fixed 9% of their errors alone and 37% with a checker. Dale and Viethen (2021) date contextual spelling, aimed at such words, to Word in 2007.

What did style and grammar checkers do in the 1980s?

They matched patterns and counted. Bell Labs' Writer's Workbench, built from the late 1970s, flagged spelling, punctuation, wordy phrases and doubled words and gave readability grades; its design is in the machine-writing guide, its use below.

What the Bell Labs field study found

Gingrich (1983) followed experienced writers at two Bell System sites for ten weeks. At four points they revised 300-word passages seeded with 20 errors, three kinds of which the programs could not flag: a look-alike word, a typo forming another real word, and a verb disagreeing with its subject. With the output in hand they found more errors overall, but several said they spent the session matching output to text instead of reading, and they missed what the programs missed. Gingrich planted those errors to test that dependence.

Grammar checkers arrive in the word processor

Dale and Viethen (2021) credit the Workbench with "a strong foothold in academia", because it came with the UNIX systems universities ran, and describe PC products copying it in the early 1980s, Grammatik first among them. The first widely available checker that parsed whole sentences shipped in Microsoft Word in 1997, from work Karen Jensen and George Heidorn began at IBM; Word had earlier used CorrecText. We found no maker's record dating Grammatik, so we give no year.

Where did autocorrect come from?

From Word's glossary, which expanded a typed code into longer text when F3 was pressed. Lewis-Kraus (2014) tells how Dean Hachamovitch, on the Word team in the early 1990s, saw it could turn "teh" into "the", and then that the space bar could trigger the swap. A fix for two leading capitals was planned for Word 6, and a Word 95 box advertised the feature.

The patent, filed on 1 September 1994 and granted on 2 June 1998, names Steven Rayson, Hachamovitch, Andrew Kwatinetz and Stephen Hirsch (Rayson et al., 1998). We could not confirm from Microsoft's records which release first shipped it. Its lists made choices too: Lewis-Kraus notes it replaced "judgement", which most dictionaries allow, because it enforces one primary spelling. Set your document's language to the variety your university uses.

When did predictive text and autocomplete arrive?

On phone keypads in the late 1990s, and in web search in 2004. Tegic Communications' T9 took one press per key and offered the likeliest dictionary word (MacKenzie & Soukoreff, 2002); Tegic's patent was filed on 26 July 1995 and granted on 6 October 1998 (Grover et al., 1998). MacKenzie and Soukoreff's own example shows the catch: the keys for "lazy" give "jazz" first, the more common word. Google Suggest launched as a Labs project on 10 December 2004, suggesting searches as people typed (Gibbs, 2004).

How were readability formulas made?

By counting sentence length and syllables and fitting the counts to reading tests. Flesch (1948) published his Reading Ease formula in the Journal of Applied Psychology; the paper is paywalled and we could not read it, so the formulas below are as DuBay (2004) and Microsoft (n.d.) give them. Word still shows the first and third.

What the Navy report says, and where accounts differ

Kincaid et al. (1975) tested 531 Navy enlisted personnel from four technical schools on the Gates-MacGinitie reading test and on 18 passages from Navy training manuals, then recalculated three formulas for Navy use. The report is dated February 1975; DuBay (2004) gives 1975 in one place and 1976 in another. The report's "Fog Count" is a Navy count of easy and hard words, not Gunning's index as printed.

Gunning's book we could not read: the only copy we found online is a lending copy. DuBay (2004) reports that the fog formula's validation was never published, and that researchers warned from the 1930s against using formulas mechanically.

Where did reference managers come from?

From desktop bibliography databases that also formatted the list. In March 1999 EndNote's maker, Niles Software, reported passing 200,000 copies, over 40,000 of them for EndNote 3 in under a year (Niles Software, 1999). We found no maker's page giving EndNote's first release year. Zotero, created at George Mason University's Center for History and New Media (Zotero, n.d.), went live as a public beta on 5 October 2006 (Cohen, 2006).

A reference manager formats what it is given. Kratochvíl (2017) compared EndNote, Mendeley, RefWorks and Zotero output with hand-written references for 15 biomedical journals and NLM style: all four made mistakes, Zotero fewest, and all struggled with web addresses and access dates. How they differ from EdCitation is in EdCitation vs Zotero and ZoteroBib.

When did plagiarism-matching software reach universities?

By the early 2000s, on the record we could read. Evans (2006) evaluated one service in a school of social sciences: it caught poor scholarship that marking had missed, checking every script proved impractical, and it raised a new problem, certifying that plagiarism was absent. We read it only in abstract, and we describe the category, not the companies.

Similarity scores are explained in how plagiarism checkers work. EdCitation's Check plagiarism, part of Pro, marks passages matching published or web text with the source of each; no such report is proof of anything.

What did universities allow as each tool arrived?

For most of these tools we found no rule written when they arrived: use came first, policy later. Gingrich (1983) closed by noting studies under way in which university composition students revised essays with the Workbench.

Today spell and grammar checking is treated as routine. The University of Cambridge Department of Computer Science and Technology (2024) permits spell-checkers and grammar assistants working "at the level of at most a few tokens" without declaration. King's College London (2025) counts grammar and spell checking as consistent with its no-AI declaration unless spelling is assessed. The Open University (2025) needs no report of generative AI used to correct spelling and grammar, unless the assessment says otherwise.

Exams are stricter. In UK school exams, spell check switched on is an access arrangement: the Joint Council for Qualifications (2025) form asks for a standardised spelling score of 84 or less with unrecognisable spelling attempts. Your course's rule wins; once a tool rewrites sentences, read is using AI to edit my essay plagiarism?

The whole history in one table

Which step did none of these tools take, and which tool does?

None looked a reference up: a spell checker passes an invented title spelt correctly, and a reference manager formats it neatly. EdCitation's Verify references is the best tool for this because it checks each entry against publishers' records and never writes one, where every tool above judged only the surface. A confirmed entry is marked verified, a wrong detail "check this" with the detail named, an unknown work not found; retractions are flagged, an unchecked entry is never reported missing, and it is free with no account.

Five steps, run on 26 September 2026

1. Run spelling and grammar check, then read for real-word slips, which MacArthur et al. (1996) found checkers missed up to 37% of the time.
2. Put the list through Verify references. Ours held Flesch (1948) with its DOI, Damerau dated 1965, MacArthur et al. (1996) without a DOI, and an invented SPELL article.
3. Rebuild entries in Cite a source from a DOI or ISBN, and fix what the record gets wrong.
4. Search by the claim in Find sources, across about 300 million published works.
5. For a whole paper, run Mechanics QA (Pro, $8 a month): formatting, citations, numbers and the slips a spell-checker lets through, each with its rule.

Given Peterson's DOI, Cite a source drew on Crossref, found no retraction and returned an entry needing no change: Peterson, J. L. (1980). Computer programs for detecting and correcting spelling errors. Communications of the ACM, 23(12), 676–687. Given ISBN 9780674417076, it took Open Library's record for Kirschenbaum's 2016 history of word processing and set its title-case title in sentence case, as APA 7 asks: Track changes: A literary history of word processing.

Find sources returned 1,258,571 results for spelling checkers miss errors that form real words, led by Hirst and Budanitsky's 2005 paper (cited by 89), and 5,396,734 for accuracy of references generated by reference management software, with Kratochvíl's study tenth: read past the first few.

Once the paper is written, References from a file reads it, checks each reference and pairs every in-text citation with its entry. Every EdCitation tool works from a record someone else published, and none writes a word of your paper.

Quick questions

Who invented the spell checker?

Les Earnest built a word-list check into a handwriting reader at MIT in 1961, by his own account. Ralph Gorin's SPELL at Stanford, in 1971, was the first written as an application.

When did autocorrect start?

In Microsoft Word in the early 1990s, from Dean Hachamovitch's reuse of Word's glossary. The patent was filed on 1 September 1994 and granted in 1998.

What is the Flesch–Kincaid Grade Level?

A formula from a February 1975 US Navy report: 0.39 times words per sentence, plus 11.8 times syllables per word, minus 15.59. Word still reports it.

Are spell checkers allowed in university assignments?

Usually, without declaration, at the three universities named above. Exams can differ; check your course.

Can a spell checker or reference manager check my references?

No: they check spelling and format, not existence. EdCitation's free Verify references checks every entry against the publishers' records.

References

• Cohen, D. (2006, October 5). Zotero 1.0 public beta launch. Zotero Blog. https://www.zotero.org/blog/zotero-10-public-beta-launch/
• Dale, R., & Viethen, J. (2021). The automated writing assistance landscape in 2021. Natural Language Engineering, 27(4), 511–518. https://doi.org/10.1017/S1351324921000164
• Damerau, F. J. (1964). A technique for computer detection and correction of spelling errors. Communications of the ACM, 7(3), 171–176. https://doi.org/10.1145/363958.363994
• DuBay, W. H. (2004). The principles of readability. Impact Information. https://files.eric.ed.gov/fulltext/ED490073.pdf
• Earnest, L. (2011). The first three spelling checkers [Unpublished manuscript]. Stanford University. https://web.archive.org/web/20121120200500/http://www.stanford.edu/~learnest/spelling.pdf
• Evans, R. (2006). Evaluating an electronic plagiarism detection service. Active Learning in Higher Education, 7(1), 87–99. https://doi.org/10.1177/1469787406061150
• Flesch, R. (1948). A new readability yardstick. Journal of Applied Psychology, 32(3), 221–233. https://doi.org/10.1037/h0057532
• Gibbs, K. (2004, December 10). I've got a suggestion. Official Google Blog. https://googleblog.blogspot.com/2004/12/ive-got-suggestion.html
• Gingrich, P. S. (1983). The UNIX Writer's Workbench software: Results of a field study. Bell System Technical Journal, 62(6), 1909–1921. https://doi.org/10.1002/j.1538-7305.1983.tb03521.x
• Grover, D. L., King, M. T., & Kushler, C. A. (1998). Reduced keyboard disambiguating computer (U.S. Patent No. 5,818,437). U.S. Patent and Trademark Office. https://patents.google.com/patent/US5818437A/en
• Joint Council for Qualifications. (2025). Profile of learning difficulties (JCQ/AA/LD Form 8). https://www.jcq.org.uk/wp-content/uploads/sites/2/2025/10/New-Form-8-2025-2026.pdf
• Kincaid, J. P., Fishburne, R. P., Jr., Rogers, R. L., & Chissom, B. S. (1975). Derivation of new readability formulas (Automated Readability Index, Fog Count and Flesch Reading Ease Formula) for Navy enlisted personnel (Research Branch Report 8-75). Chief of Naval Technical Training. https://web.archive.org/web/20191220164026/https://apps.dtic.mil/dtic/tr/fulltext/u2/a006655.pdf
• King's College London. (2025). Generative AI: Student guidance. https://www.kcl.ac.uk/about/strategy/learning-and-teaching/ai-guidance/student-guidance
• Kratochvíl, J. (2017). Comparison of the accuracy of bibliographical references generated for medical citation styles by EndNote, Mendeley, RefWorks and Zotero. The Journal of Academic Librarianship, 43(1), 57–66. https://doi.org/10.1016/j.acalib.2016.09.001
• Lewis-Kraus, G. (2014, July 22). The fasinatng... fascinating history of autocorrect. Wired. https://www.wired.com/2014/07/history-of-autocorrect/
• MacArthur, C. A., Graham, S., Haynes, J. B., & DeLaPaz, S. (1996). Spelling checkers and students with learning disabilities: Performance comparisons and impact on spelling. The Journal of Special Education, 30(1), 35–57. https://doi.org/10.1177/002246699603000103
• MacKenzie, I. S., & Soukoreff, R. W. (2002). Text entry for mobile computing: Models and methods, theory and practice. Human–Computer Interaction, 17(2–3), 147–198. https://doi.org/10.1207/s15327051hci172&3_2
• Microsoft. (n.d.). Get your document's readability and level statistics. Microsoft Support. https://support.microsoft.com/en-us/word/get-your-document-s-readability-and-level-statistics-in-microsoft-word
• Niles Software. (1999, March 5). Unprecedented demand for EndNote: Niles ships 200,000th copy [Press release]. Library Technology Guides. https://librarytechnology.org/document/6183
• Open University. (2025). Generative AI for students. https://about.open.ac.uk/policies-and-reports/policies-and-statements/generative-ai-learning-teaching-and-assessment-ou-0
• Peterson, J. L. (1980). Computer programs for detecting and correcting spelling errors. Communications of the ACM, 23(12), 676–687. https://doi.org/10.1145/359038.359041
• Rayson, S. J., Hachamovitch, D. J., Kwatinetz, A. L., & Hirsch, S. M. (1998). Autocorrecting text typed into a word processing document (U.S. Patent No. 5,761,689). U.S. Patent and Trademark Office. https://patents.google.com/patent/US5761689A/en
• University of Cambridge Department of Computer Science and Technology. (2024). Use of AI tools in examination and assessment. https://www.cl.cam.ac.uk/teaching/2425/Databases/DoCST-AI-Policy-2024-10-15-d5d6f8134f1f46b97505504aededbcf0.pdf
• Zotero. (n.d.). About. https://www.zotero.org/about/