In 2005, Steve Jobs stood on stage at a product event and addressed a complaint that had been accumulating since Apple introduced shuffle play on the iPod. Users were convinced the feature was broken. They'd hear two songs by the same artist back to back, or three tracks from the same album in a row, and conclude that the shuffle wasn't really shuffling. It was favoring certain artists, or playing songs in a predictable pattern, or somehow rigged. Jobs's response was characteristically blunt: "We're making shuffle less random to make it feel more random."
The problem was that Apple's original shuffle algorithm was genuinely random. Each song had an equal probability of being selected next, with no regard for what had just played. And genuine randomness produces clusters. If you have a playlist with ten songs by five artists — two songs each — a truly random sequence will regularly play two songs by the same artist consecutively. It will occasionally play three. It might even, over a long enough session, play all the songs by one artist before touching another. This is the same phenomenon that makes a fair coin produce six heads in a row more often than people expect: random sequences are clumpier than human intuition predicts, and the clumps feel like patterns.
Apple's fix was to introduce what's sometimes called a "spread" or "stratified shuffle." Instead of picking each song independently, the algorithm distributes songs by the same artist or album as evenly as possible across the sequence, then adds a small amount of randomization on top. The result is a playlist that spaces similar songs apart — ensuring variety — while still feeling unpredictable within that constraint. It's less random by every mathematical measure, but it matches what listeners mean when they say they want a "random" mix: variety without repetition, surprise without clustering, the feeling that no song is predictable but no pattern is detectable.
This gap between mathematical randomness and perceived randomness shows up everywhere, not just in music. Spotify encountered the same issue and solved it similarly, eventually publishing research on their shuffling approach. The streaming service found that users expected shuffle to behave like a well-programmed DJ — touching every genre and mood in the library, avoiding back-to-back tracks with similar tempos or keys, and never repeating a song until everything else has played. True randomness does none of these things. True randomness doesn't know what a genre is.
The lesson here extends well beyond playlist design. In any context where randomness is experienced sequentially — raffle drawings, name selections, team assignments — people will perceive genuine random output as biased if it produces clusters. If a teacher uses a random name picker and the same student gets called on twice in a row, the class will suspect the tool is broken even though consecutive selections are a natural feature of independent random draws. If a giveaway spinner lands on adjacent entries in consecutive spins, viewers will question whether the wheel is weighted.
There are two ways to handle this. The first is to educate users about what randomness actually looks like, which is valuable but often insufficient because the perception of unfairness is emotional, not intellectual. A person can understand that consecutive picks are statistically normal and still feel that the process is suspect. The second approach — the one Apple and Spotify adopted — is to design for perceived fairness rather than mathematical purity. In practice, this means adding constraints: don't allow the same name to be picked twice in a row, space out repeated categories, ensure that every entry is selected before any entry repeats. These constraints reduce true randomness but increase the experience of fairness, which is usually what matters.
The irony is that making randomness feel random requires making it less random. Humans want the aesthetic of chance — unpredictability, surprise, the absence of visible control — without its actual statistical properties. This isn't a failure of human cognition. It's a mismatch between what randomness is (a mathematical property of sequences) and what randomness means to us (a feeling of variety and impartiality). The best tools recognize this gap and design for the meaning rather than the math.