The average person can hold only 7 (plus or minus 2) items in working memory at once. Go beyond that limit and things start to fall through the cracks.
In 1956, psychologist George Miller published a landmark paper titled “The Magical Number Seven, Plus or Minus Two.” He had noticed something consistent across dozens of experiments: when people were asked to hold items in short-term memory — numbers, letters, words, tones — performance reliably fell off a cliff at around seven items. Below seven, recall was near-perfect. Above it, things disappeared.
The implication for design is straightforward and unavoidable: every time you present more than seven things at once without grouping or structure, you are asking users to hold more than working memory can reliably contain. Some of it will drop. The question is which part — and in a form, a navigation, or an AI response, that's rarely the part you'd choose.
Miller ran participants through a series of tasks involving sequences of digits, letters, and words presented at speed. He wasn't studying how much people could memorise — he was studying something more precise: the channel capacity of human short-term cognition. At what point does information simply not get processed?
The answer, replicated across many conditions, was consistently around seven. But Miller's more important observation was that this limit applied to chunks of information, not to raw data. A trained musician can hold a complex chord in working memory as a single chunk. A novice holds each note separately. The capacity is the same — what differs is the chunking ability of the person and, crucially for designers, how the interface structures the information before it reaches the user.
“The span of absolute judgement and the span of immediate memory impose severe limitations on the amount of information we can receive, process, and remember.”
— George Miller, 1956
The 16-digit credit card number is one of the oldest and most universal chunking problems in interface design. As a flat string of digits, it requires users to hold all 16 in working memory to spot an error, verify what they've entered, or match it against their physical card. As four groups of four, it requires holding four chunks — well within Miller's limit — and aligns with how the number is physically printed on the card.
Every major payment processor, bank, and e-commerce platform chunks card numbers into four groups. This isn't a visual preference — it's a direct application of Miller's Law that reduces transcription errors and lets users visually verify their entry at a glance.
16 raw digits in one field. No structure to verify against.
4 chunks of 4 digits. Matches the physical card. Errors jump out instantly.
The chunked version does something the flat version cannot: it lets users verify each group of four independently against their physical card. Instead of a single 16-item memory task, the verification becomes four 4-item tasks — each well within working memory's reliable range. Error rates drop, and so does the anxiety of entering a number that controls access to money.
One of the most consistent cognitive load failures in AI-generated content is the flat bullet list. When a model responds with fourteen bullets in a single undifferentiated list, it is presenting fourteen items to working memory simultaneously. Users must scan the entire list before they can begin to identify what's relevant, group related items mentally, and retain what they've read.
The same information organised under three labelled headings becomes three chunks. Each section is a group with a clear label — the label itself tells users whether they need to read the items beneath it before they start reading them.
14 items, no structure. Working memory fills before you reach the end.
3 sections of 4–5 items. The label tells you whether to read the group. Same 14 points — completely different cognitive cost.
This is why prompting an AI to “organise your response into sections” consistently produces more usable output than asking for a list. The model contains the same information either way. The structure determines whether working memory can process it. Good AI interfaces are beginning to apply this automatically — recognising response length and complexity, and grouping output into labelled sections before presenting it. This is Miller's Law applied at the level of content generation, not just interface layout.
Miller's Law offers a clear diagnostic question for any interface: how many separate items is the user holding in working memory at this moment? If the answer is more than five to seven, the design is doing cognitive work that it should be doing for the user. Grouping, chunking, progressive disclosure, and visual hierarchy all have the same effect: they reduce the number of things the brain must hold simultaneously, freeing capacity for the task itself.