Scrabble project · 12 November 2026 · revised: usage, not dictionaries, is the test
What the Scrabble Lists Leave Out
Do the official word lists omit words that are as good as the words they include? Here "as good" means one thing only: how often people actually print the word. We measured every listed word, and every unlisted spelling, in twenty years of books.
Short answer. Yes, on a large scale. About 57,000 English words that are in neither NWL2023 nor CSW24 were used in books more often than the typical listed word of the same length. About 5,900 of them were used more often than three quarters of listed words. dataset, runtime, roadmap, lifecycle, plugin and rulemaking are examples.
The lists also err the other way. Of their 284,585 words, 25,939 (9%) were not printed in lower case even once in books of 2000 to 2019, and 64,399 (23%) were printed fewer than 100 times.
The test
A word list for a game could be judged against dictionaries, which is how the lists themselves are made. This article does not do that. It asks what writers do. A spelling counts as a word in use if it is printed, in lower case, in ordinary running text, and it counts for more the more often that happens.
The evidence is the Google Books English corpus, 2020 edition, which gives for every spelling the number of times it was printed in each year, with capital letters kept. We used books published from 2000 to 2019 and three rules:
- Usage is the number of times the spelling was printed entirely in lower case. Turkey the country does not count towards turkey the bird.
- An unlisted spelling is a candidate only if at least 60% of its appearances are in lower case. That removes names and acronyms by how writers treat them, with no dictionary involved.
- Words are compared with listed words of the same length, because long words are rarer than short ones. The yardstick is the median listed word: half of all listed words of that length are used less.
Scrabble's physical limits still apply: 2 to 15 letters, letters only.
The yardstick: how much the listed words are used
| Letters | Listed words | Median uses | Never used | Share never used |
|---|---|---|---|---|
| 2 | 127 | 2,099,900 | 0 | |
| 3 | 1,352 | 228,545 | 0 | |
| 4 | 5,667 | 40,314 | 1 | |
| 5 | 13,030 | 9,906 | 74 | |
| 6 | 23,168 | 4,035 | 623 | |
| 7 | 34,548 | 2,227 | 2,048 | |
| 8 | 42,383 | 1,580 | 3,359 | |
| 9 | 43,207 | 1,290 | 3,974 | |
| 10 | 38,008 | 1,162 | 3,920 | |
| 11 | 29,971 | 990 | 3,573 | |
| 12 | 21,930 | 927 | 2,825 | |
| 13 | 15,073 | 790 | 2,286 | |
| 14 | 9,885 | 518 | 1,857 | |
| 15 | 6,236 | 323 | 1,399 | |
| All | 284,585 | 25,939 | 9.1% |
The median is a low bar. A typical listed nine-letter word was printed 1,290 times in twenty years of books, in a corpus where the appears billions of times. Some famous Scrabble words sit far below it: zyzzyva was printed in lower case 207 times, euouae 150 times, muzjiks 43 times.
"Never used" means no lower-case appearance in books of 2000–2019. Google leaves out any spelling that occurs fewer than 40 times in the whole corpus, so a word used a handful of times also shows as never used.
What is missing
142,344 unlisted spellings pass the lower-case rule and beat the median listed word of their length. Not all of them are words. A corpus of scanned books is full of scanning errors (govemment), pieces of words broken at line ends (tion), foreign words inside English books (und), Latin species names (domesticus) and compounds that lost their hyphen in scanning (wideeyed).
To find out how many are real, we drew two fixed random samples and read every entry. A spelling was counted as a genuine word only if it is an English word that writers print solid. Hyphen-less compounds that are normally hyphenated, foreign and Latin words, misspellings and fragments were all counted against.
| Used more than… | Unlisted spellings | Sample read | Genuine in sample | Estimated genuine words |
|---|---|---|---|---|
| half of listed words of the same length | 142,344 | 582 | 235 (40%) | about 57,000 (52,000–63,000) |
| three quarters of listed words of the same length | 9,386 | 465 | 290 (62%) | about 5,900 (5,400–6,300) |
The ranges are 95% sampling intervals. They do not cover differences of judgment about what counts as a word, which is discussed under Limits.
| Kind | Count | |
|---|---|---|
| Genuine English word, printed solid | 290 | |
| Foreign word, Latin phrase or species name | 71 | |
| Compound that lost its hyphen or space | 58 | |
| Scanning error, fragment, misspelling, abbreviation | 36 | |
| Archaic or dialect spelling (causeth, gettin) | 10 |
Some of the words
Every word in this table was checked against both list files and is in neither.
| Word | Letters | Uses | Median for a listed word of that length |
|---|---|---|---|
| dataset | 7 | 2,455,924 | 2,227 |
| rulemaking | 10 | 1,175,038 | 1,162 |
| runtime | 7 | 1,080,171 | 2,227 |
| namespace | 9 | 888,003 | 1,290 |
| intraoperative | 14 | 620,619 | 518 |
| lifecycle | 9 | 610,846 | 1,290 |
| roadmap | 7 | 472,973 | 2,227 |
| peacebuilding | 13 | 352,575 | 790 |
| plugin | 6 | 326,001 | 4,035 |
| filesystem | 10 | 264,973 | 1,162 |
| offsite | 7 | 228,222 | 2,227 |
| birthweight | 11 | 200,546 | 990 |
| scatterplot | 11 | 189,638 | 990 |
| crewmembers | 11 | 189,198 | 990 |
| backpay | 7 | 189,121 | 2,227 |
| waterbodies | 11 | 80,649 | 990 |
| nociception | 11 | 73,058 | 990 |
| skillset | 8 | 22,838 | 1,580 |
dataset was printed about 2.5 million times, as often as blog, which is listed. It is used more than 95% of all listed seven-letter words.
What kinds of words they are
- Medicine and biology. esophagectomy, neutrophilia, overexpression, upregulation, proteasome, myocytes, and drug names such as propofol, midazolam and carboplatin.
- Computing and mathematics. dataset, runtime, namespace, hostname, subtree, iterator, homotopy, discretized.
- Social science and the humanities. performativity, governmentality, metacognitive, peacebuilding, outmigration.
- Solid compounds. roadmap, lifecycle, skillset, cleanroom, picturebook, lifejackets, trendlines, streamflow.
- Regular derivatives. Words built with non-, pre-, post-, sub- and un-, written solid: nonprocedural, postprocessed, subformula, unplotted, redesignated.
The lists already contain thousands of words of each of these kinds. What separates the listed from the unlisted is not the kind of word or how much it is used. It is whether one of the source dictionaries happened to enter it.
Words removed on purpose
Usage also shows the deliberate removals. gringo (127,705 uses) is used more than three quarters of listed six-letter words and is in neither list. redneck (132,942) is valid in North America and was removed from the Collins list. These were taken out as offensive between 2019 and 2022, a policy about the game and not a finding about the language.
What is included but not used
The same measurement applies to the lists themselves. 25,939 listed words have no lower-case appearance in twenty years of books.
- 8,333 of them are in both lists. Most are inflections that a rule generates and nobody writes: worrisomenesses, tomboyishnesses, squeezabilities, steeplechasings.
- 17,005 are in CSW24 only: disshivers, dragonizes, disembrangling, fustianises.
- 601 are in NWL2023 only.
The two lists differ sharply by this measure. The 87,984 words that only Collins accepts are much rarer in print than the rest.
| Letters | All listed words | CSW24-only words | Their median | NWL2023-only words | Their median |
|---|---|---|---|---|---|
| 5 | 9,906 | 3,554 | 888 | 12 | 12,188 |
| 6 | 4,035 | 6,462 | 282 | 28 | 2,837 |
| 7 | 2,227 | 9,075 | 129 | 42 | 853 |
| 8 | 1,580 | 10,647 | 98 | 42 | 95 |
| 9 | 1,290 | 11,978 | 103 | 27 | 14 |
| 10 | 1,162 | 12,980 | 112 | 574 | 1,750 |
| 11 | 990 | 11,236 | 93 | 796 | 571 |
| 12 | 927 | 8,415 | 86 | 805 | 614 |
| 13 | 790 | 5,750 | 75 | 638 | 517 |
| 14 | 518 | 3,783 | 51 | 448 | 327 |
| 15 | 323 | 2,397 | 44 | 280 | 428 |
A typical Collins-only word of eight letters or more was printed about a hundred times in twenty years, roughly a tenth as often as the typical listed word. The long words that only North America accepts are close to the general level.
Limits of this check
- The genuine-word counts rest on one reader. The samples were classified by Claude, not by a panel, and the line between a technical term and a word is a judgment. A stricter reader who rejected chemical names and non- derivatives would get a smaller number; one who accepted dialect spellings and solid compounds would get a larger one. The raw counts of 142,344 and 9,386 do not depend on judgment.
- Books are not all usage. The corpus is edited print, heavy with academic and technical books. Speech, the web and informal writing are not in it. Slang is undercounted and scientific vocabulary overcounted.
- Lower case is a strict test. A word that usually starts sentences or appears in titles loses some of its count, and a word normally capitalised is excluded altogether even if a dictionary enters it in lower case.
- Scanning joins words. When a hyphenated compound is broken across lines, the scan can print it solid. That is why wellknown looks common. We counted every such compound as not a word, which is cautious: some, like longstanding, are widely printed solid on purpose.
- Inflections are counted one by one. A rare plural of a common word is treated as a rare word, for listed and unlisted words alike.
Conclusion
Measured by what writers print, the official lists are not a picture of English in use. They leave out tens of thousands of words that are used more than half of what they include, several thousand of them heavily, and they include tens of thousands of words that are barely used at all. Both effects have the same cause. The lists are transcriptions of particular dictionaries, and a dictionary's coverage of technical vocabulary, solid compounds and regular derivatives is a matter of editorial policy and date, not of use.
Sources and method
List files: NWL2023 (196,601 words) and CSW24 (280,887 words) as published for tournament software. Usage: Google Books Ngram corpus, English, edition 20200217, 1-grams, restricted to 2000–2019, capitalisation kept; 14.3 million distinct letter-only spellings of 2 to 15 letters. Counts are produced by two C++ programs in the project, ngcase (counts each spelling by capitalisation) and usegap (compares listed and unlisted spellings and draws the samples). The samples are fixed by a hash of the spelling, so they can be drawn again. An earlier version of this article judged omissions against four dictionaries; it was replaced because use, not dictionary entry, is the right test. Background on the removals of offensive words: NASPA Word List and Collins Scrabble Words on Wikipedia.