How Random Are You? The Show HN Digit-Prediction Game That Beats Most Humans

How Random Are You? is a browser game that tries to predict the next digit you say, and it beats most people. Posted to Hacker News as a Show HN by user vunderba, the page at lab.specr.net/rigged-draw reached 158 points and 80 comments. The core finding the game demonstrates is that humans cannot generate independent random digits: they avoid numbers they have recently used, and that avoidance is exactly what the predictor exploits.

What the game asks you to do

The game asks you to say numbers from 1 to 9 out loud, one after another, as randomly as you can, without stopping to think. A typed-input option is also available.

Before each number, the house seals its guess in an envelope. You say the number and the envelope opens. Your first 10 numbers are a warm-up: the house guesses, but nothing counts toward your score. At the end, your ticket shows how you compare with 206 people from a psychology study.

The site's own instruction is explicit that volume matters more than care: "The more numbers you say, the more accurate your result." It also warns that misheard numbers should not interrupt you — "if the house mishears one now and then, don't stop."

How the predictor works

The house runs ten small predictors at once and trusts whichever has been right lately. According to the site's own description, the ensemble includes:

Predictors that count which number tends to follow which number (transition or bigram-style models).
Predictors that track your steps up and down between consecutive digits.
A predictor that looks for the last time you said the same few numbers in a row.
A predictor that notices which numbers you have not said for a while.

The site states that the last of these does most of the work, because "people keep reaching for the numbers they've neglected." The underlying statistical point is stated plainly on the page: "In a truly random sequence, each number is independent of the ones before it."

The model is not trained from scratch on each player. It gets a head start from 418 people who each picked 200 numbers from 1 to 9 in a psychology study, then adjusts to the individual player as they go.

What a good score looks like

The site reports blind guessing as the baseline: a random guess would hit roughly 11% of the time. Commenters reported a wide spread of results against that baseline:

steve_adams_86 automated the game with Math.random() and got 18 of 190 scored numbers correct, or 9%, against a blind-guessing average of about 21.1 (11%). The site told them: "Only 1 of the 206 people in a psychology study was more random than you. The house found you extremely hard to read."
ehwa37 reported 9 of 117 scored numbers, or 8%, with the site stating that none of the 206 study participants were more random.
tchalla reported 1 of 32 scored numbers, or 3%.
kykat reported 10 of 87, or 11%, with only 5 of 206 study participants more random.
riffraff reported 3 of 27 (11%) on a first attempt and 14 of 118 (12%) on a second, and noted that typing may be easier to do well at than speaking.
floorsthelimit reported the house calling 0 of 50.
vorticalbox reported reaching 2 of 30 before the house guessed 7 correctly from numbers 30 through 50.

A lower hit rate is better. The site's framing rewards players whose sequences the house cannot read.

The Aaronson oracle lineage

Commenters immediately identified the game as a variant of the Aaronson oracle, a well-known demonstration that humans cannot type randomly. User costco quoted Scott Aaronson's own account of the original experiment:

In a class I taught at Berkeley, I did an experiment where I wrote a simple little program that would let people type either "f" or "d" and would predict which key they were going to push next. It's actually very easy to write a program that will make the right prediction about 70% of the time. Most people don't really know how to type randomly. They'll have too many alternations and so on. There will be all sorts of patterns, so you just have to build some sort of probabilistic model. Even a very crude one will do well. I couldn't even beat my own program, knowing exactly how it worked. I challenged people to try this and the program was getting between 70% and 80% prediction rates. Then, we found one student that the program predicted exactly 50% of the time. We asked him what his secret was and he responded that he "just used his free will."

User cubefox pointed to a standalone implementation of the same idea at aaronsonoracle.com. User DylanMerigaud summarized the failure mode in one line: "Too many alternations is exactly how people give themselves away."

Strategies commenters found that worked

Several commenters reported beating the house by deliberately modeling the model rather than trying to be random.

Lag the machine's guesses. christophilus wrote: "I started just lagging the guesses of the machine, and it never seemed to catch on." Their result was 7 of 206 study participants more random.
Adversarially avoid a repeated guess. alwa described noticing the house repeat a guess and then running a long streak avoiding that number: "it's kind of fun being on the machine side of the slot-machine-type racket!"
Pick instinctively, then shift by one. maidaman reported doing badly on a first attempt, then doing well by deciding on a number instinctively and picking something next to it: "Maybe we instinctively pick similar numbers, but if we fight our instincts we end up with something truly random."
Alternate and advance from both ends. akshitgaur2005 reported that this called n-1 and n+1 correctly when the previous guess was n, but produced no correct calls on a sequence like 1, 8, 3, 7, 4, 7, 1, 7, 3, 8.
Poison the warm-up. malfist suggested using the unscored warm-up to repeat the same number and then abandon it: "Use your warmup to repeatedly guess the same number, and then don't use it for a while."
Arithmetic mixing (untested). Trusteando proposed thinking of a random integer n1 from 1 to 9, then another n2, and saying (n1 + n2) mod 9 + 1, and invited others to report results.
Cheating with known constants. shric said knowing pi to 100 decimal places made it easy to cheat, and kps reported that the site agreed the digits of pi are random.
Automating it. steve_adams_86 drove the game with a setInterval loop clicking a random ball every second, which produced the 9% result described above.

Critiques, bugs, and open questions

The measurement may not be comparable to the study. kykat raised the sharpest methodological objection: players can see the house's predictions in real time, which the original psychology study participants may not have been able to do. "After my next number was predicted correctly, I changed my 'strategy' to increase randomness (be less predictable). So I don't know if this result is comparable to the one from the study."

A copy bug in the results text. nightpool flagged a sentence that appears to compare a value against itself: "You spread your picks more evenly than a random draw would. Your most-used number made up 23% of your picks, where a random draw's busiest number gets about 23%."

Voice data and privacy. pavel_lishin asked whether the voice mode is used to train a model: "Is anyone sufficiently paranoid that their first thought was, 'Are you going to train a model on my voice?'" The site does offer a typed alternative, which nico called out as a good design choice. coolThingsFirst asked which speech recognition model the site uses.

Provenance of the copy. yesitcan claimed that seeing the word "sealed" was enough to identify the site text as Claude-generated. The site does publish a sources and licences file at lab.specr.net/rigged-draw/licenses.txt.

Related projects and prior art

aaronsonoracle.com, cited by cubefox, is a direct implementation of the f/d prediction experiment.
humanrandomness.com, built by tibbar, inverts the premise: instead of trying to be random, you try to pick the number other people are least likely to pick. tibbar noted the project helped them move from unemployment in the South to working in San Francisco.
xkcd 1210, cited by andai, is the canonical comic about how people who think they are random are not.
The Dilbert strip about Accounting's random number generator was cited independently by both m463 and exabrial.

Bottom line

How Random Are You? is a working demonstration that human digit generation is predictable, not random. The predictor's stated advantage comes from a single strong behavioral regularity — people avoid digits they have recently used — combined with an ensemble of weaker transition and step models. The game is beatable, but the strategies that beat it (lagging the machine, adversarially avoiding repeated guesses, poisoning the warm-up) all require modeling the predictor rather than trying harder to be random. The most credible caveat from the discussion is that live feedback changes player behavior, so scores from the web game are not directly comparable to the 206-person psychology study it cites as a baseline.