Fairness, tested
How It Works: Fair Random Letters
Every letter randomletter.io gives you is picked with secure random numbers from your own browser, and each letter in the set has exactly the same chance. This page explains how the pick works and shows a test of 1,000,000 draws you can check yourself.
1,000,000
Draws per setting in the test
23 of 23
Settings passed in version 2.0
1.1%
Largest gap from the expected count, A to Z
0
Letters sent to our server
01 / What random means
What does “random” mean on this site?
It means every letter you allow has an equal chance on every tap, and no draw depends on the one before. With all 26 letters on, each letter has a 1 in 26 chance each time. That is about 3.85%.
Random does not mean the results look evenly spread in the short run. You can get the same letter twice in a row, or wait a long time for a Q. That is normal for a fair draw. If you want no repeats, turn on “No repeats until all are used” in Settings, or “Unique in set” for different letters within one draw.
02 / The source
Where do the random numbers come from?
They come from the Web Crypto API that is built into your browser. The tool asks it for random numbers with a function called crypto.getRandomValues().
Browsers make these numbers with a cryptographically secure generator, seeded from enough unpredictable data to be safe for security work. It is not a hardware “true random” device, and no website using a browser is. For picking letters it is far more than fair enough.
This feature has worked in all major browsers since 2015. A very old browser without it gets a message asking to update, rather than a weaker random source.
Until 9 November 2026 the tool used Math.random(), a simpler generator that browsers do not promise to be secure. Version 0.3.1 replaced it everywhere.
03 / No bias
How does a random number become a letter, with no bias?
The tool draws a large random whole number, then maps it onto the letters in your set. If it simply took the remainder after dividing, a few letters could come up very slightly more often than others. So it throws away the small band of numbers that would cause that tilt and draws again. Every letter then has exactly the same chance.
This is a standard method for unbiased random ranges, called rejection sampling. A redraw is rare: with 26 letters it happens about once in 195 million draws.
04 / Every setting
What happens with vowels, other alphabets, exclusions, no repeats and the wheel?
The same pick is used everywhere. Only the set of letters changes.
- Vowels: the set is A, E, I, O, U, so each has a 1 in 5 chance.
- Consonants: the other 21 letters, 1 in 21 each.
- Other alphabets: the set is that alphabet’s letters, such as 27 for Spanish or 33 for Russian, each with the same chance. Vowels and Consonants are for English only.
- Custom letters: the set is exactly the letters you type, each counted once, up to 60.
- Exclude letters: removed letters are taken out of the set before the pick, so they can never appear, and the rest stay equal. Since version 1.0 this works with every type, Vowels included: exclude E, and Vowels draws from A, I, O and U only.
- Unique in set: when you draw several letters at once, each letter picked is taken out for the rest of that draw.
- No repeats until all are used: each letter drawn stays out of the set until every letter has come up, and each pick is a fair pick from the letters left. A full round is a random order of the whole set.
- Mixed case: the letter is picked first, then upper or lower case is picked separately, 50/50.
- Wheel: the letter is picked first, then the wheel turns to land on it. The spin is for fun; it does not decide anything. With no repeats on, each drawn letter leaves the wheel.
What each button and setting does is in the user guide, with numbered screenshots.
Nothing you draw is sent to the randomletter.io server. The random numbers are made and used in your browser.
05 / The test
Is it really fair? The 1,000,000-draw test
Yes. I drew 1,000,000 letters with the exact code the site runs and counted how often each letter came up. Every letter landed within 1.1% of its expected count, and a standard statistical test found nothing unusual.
Test details: tool version 2.0, run on 12 November 2026 in a Chromium browser, 1,000,000 draws per setting, 23 settings.
All 26 letters (expected 38,461.5 each)
| Letter | Count | Letter | Count |
|---|---|---|---|
| A | 38,367 | N | 38,676 |
| B | 38,661 | O | 38,496 |
| C | 38,228 | P | 38,245 |
| D | 38,062 | Q | 38,610 |
| E | 38,543 | R | 38,558 |
| F | 38,485 | S | 38,288 |
| G | 38,659 | T | 38,531 |
| H | 38,384 | U | 38,565 |
| I | 38,499 | V | 38,618 |
| J | 38,357 | W | 38,692 |
| K | 38,598 | X | 38,175 |
| L | 38,381 | Y | 38,055 |
| M | 38,524 | Z | 38,743 |
Every setting tested (1,000,000 letters drawn for each setting)
| Setting | Chi-square (df) | p-value | Result |
|---|---|---|---|
| All 26 letters | 24.22 (25) | 0.51 | Fair |
| Vowels | 7.17 (4) | 0.13 | Fair |
| Consonants | 11.39 (20) | 0.94 | Fair |
| All except Q, X, Z | 18.46 (22) | 0.68 | Fair |
| Vowels except E | 0.90 (3) | 0.82 | Fair |
| Mixed case | 48.06 (51) | 0.59 | Fair |
| Spanish | 23.14 (26) | 0.63 | Fair |
| French | 30.24 (25) | 0.22 | Fair |
| German | 22.36 (29) | 0.80 | Fair |
| Italian | 15.09 (20) | 0.77 | Fair |
| Portuguese | 27.46 (25) | 0.33 | Fair |
| Dutch | 27.15 (25) | 0.35 | Fair |
| Greek | 14.94 (23) | 0.90 | Fair |
| Russian | 26.41 (32) | 0.75 | Fair |
| Custom set of 60 letters | 68.86 (59) | 0.18 | Fair |
| No repeats (40,000 rounds of 26) | 23.62 (25) | 0.54 | Fair |
| 100 letters per draw | 35.76 (25) | 0.08 | Fair |
Excluded letters were drawn 0 times (Q, X and Z; E from Vowels). In no-repeat mode, all 40,000 rounds used each letter exactly once; the row shows the first letter of each round. Mixed case counts 52 symbols (26 capitals and 26 small letters). The CSV has all 23 settings and every count.
06 / Reading the numbers
How do you read a chi-square test?
The chi-square test compares the counts you got with the counts you would expect if every letter were equally likely. It adds up the squared gaps, scaled by the expected count.
The p-value says how often a perfectly fair picker would produce a spread at least this uneven by chance. For all 26 letters it was 0.51: a fair picker does that about 51% of the time, so the result is ordinary. I would only worry if the p-value were tiny, below 0.001. None came close.
07 / Letter frequency
Why does Z come up as often as E?
Because a fair letter picker gives every letter the same chance, while English text does not use letters equally. In one study of 40,000 English words, E made up about 12% of letters and Z less than 0.1%.
So if you want letters that feel like English words, such as for a word game, a uniform picker will hand you plenty of Q, X and Z. You can leave them out with “Exclude letters”.
08 / Version history
Changelog
- 14 November 2026, version 2.0: a new look, the Letter Machine, with the wheel or the letters on one side and the controls in a panel beside them (below them on a phone); the homepage opens on the wheel and remembers whether you last used Letters or the wheel; Settings in three tabs; a High contrast wheel theme; Mixed case shows both forms on the wheel, such as Aa; a larger wheel in full screen; clearer steps between Slow, Normal and Fast. No change to how letters are picked.
- 10 November 2026, version 1.1: the letter wheel is now drawn as a sharp vector image (SVG), so it stays crisp on any screen; the version line moved inside the tool for better contrast; each switch has one clear label. No change to how letters are picked.
- 10 November 2026, version 1.0: no repeats until every letter is used; your own custom letters; nine alphabets; 1 to 100 letters per draw in a row or a grid; share link, print and download; a Letters / Wheel switch; an optional timer; every result announced to screen readers. Exclusions now apply to Vowels too. Letters are still picked with the same method.
- 9 November 2026, version 0.3.2: the tool now starts as soon as the page loads, so the first tap always works. No change to how letters are picked.
- 9 November 2026, version 0.3.1: switched every pick to the Web Crypto API with rejection sampling; fixed mixed case so the case of each letter is random (it used to alternate); added anonymous usage counts.
- Before November 2026, version 0.3.0: picks used
Math.random().
I rerun the fairness test with each new version and update this page.