Image: Steven Lek, CC BY-SA 4.0, via Wikimedia CommonsSeven, plus or minus two, or four
George Miller opened his 1956 paper with a complaint: "My problem is that I have been persecuted by an integer." The integer was seven. It kept appearing in his data and in everyone else's, and the paper, "The Magical Number Seven, Plus or Minus Two", published in Psychological Review and at the journal, became the most cited in the history of the field. Most people who quote it have not read it, which is a pity, because Miller's point was subtler than the title and his final paragraph half retracts it.
Where the seven came from
Miller brought together two lines of evidence. The first was absolute judgement: ask people to sort tones by pitch, or lights by brightness, into categories, and they manage about seven categories before they start to confuse neighbours, whatever the dimension. The second was immediate memory: read someone a list of digits, letters, or words and ask for it back in order, and the longest list most adults manage is about seven items, give or take two. The two sevens looked like the same limit.
They are not, as Miller suspected, and the second seven has a problem that he described himself. The memory span is counted in items, and an item can be anything the person treats as one thing. He called these chunks. A string of 21 binary digits is far beyond anyone's span, but a colleague of Miller's, Sidney Smith, taught himself to recode groups of three binary digits into a single octal digit and could then repeat strings of 40. The information held had grown by a factor of six; the number of chunks had not changed. The span, Miller concluded, is a fixed number of chunks, and the size of a chunk is a matter of learning. That part has held up completely.
Why the number came down
The digit span of seven is measured under conditions that help. The list is spoken, the items are familiar, the person can rehearse silently while it is read, and digits group themselves into pairs and triples as soon as you hear them, the way a telephone number does. Strip those aids away and the span falls. In 2001 Nelson Cowan reviewed the experiments that do strip them away, in a long paper in Behavioral and Brain Sciences, "The magical number 4 in short-term memory", at the journal. When rehearsal is blocked by having the person repeat a word aloud, when the items are presented too fast or too unexpectedly to group, when the task is to notice a change in an array of coloured squares, when the material is a list of tones, the limit that appears is about four chunks, with a range from three to five across people and tasks.
The best known of those experiments is from vision. Steven Luck and Edward Vogel showed people an array of coloured squares for a tenth of a second, blanked it, showed it again with one square possibly changed, and asked whether anything had changed. Performance was almost perfect up to four squares and fell steadily beyond. It made no difference whether each square had one feature to remember or four, which is to say that the limit was four objects, not four features. The paper was published in Nature in 1997 and is at the journal.
So both figures are right. Seven, plus or minus two, is the span you measure for familiar, speakable material that you are allowed to rehearse and group. About four is the number of separate things you can hold in attention when none of that help is available, and it is the better estimate of the core capacity of working memory, the system that holds and manipulates information for the seconds that thinking takes. Unit 5 of the masterclass gives both numbers, with the experiments, and uses Cowan's as the current one.
Time as well as count
There is a third way of stating the limit. Alan Baddeley and colleagues found in 1975 that the span for words depends on how long they take to say: people remember more short words than long ones, and the span comes out at about as many words as they can pronounce in two seconds. Speakers of languages with short number words, such as Chinese, have longer digit spans than speakers of languages with long ones, such as Welsh, and the difference disappears when the speaking rate is taken into account. The phonological loop, the inner voice that rehearses speech, seems to hold about two seconds of it. A chunk that takes long to say costs more of the loop.
What the limit does and does not constrain
The capacity of working memory is one of the most stable facts about the mind, and one of the easiest to work round. The route round it is always the same: make bigger chunks. A chess master shown a position from a real game for five seconds reproduces most of the 25 or so pieces; a beginner places about a quarter of them. Shown a random arrangement of the same pieces, the master does no better than the beginner. The master has tens of thousands of familiar patterns in long-term memory, and a real position is four or five of them. A student who learns the alphabet of a subject, its formulas, names, and standard moves, until they are single chunks, is doing the same thing, and it is why fluency in the basics frees attention for the problem.
The other route is practice of a different kind. A student at Carnegie Mellon known as S.F. spent about 230 hours over two years practising the digit span and raised it from 7 to 79, by recoding groups of digits as running times, which he knew well as a runner, and then chunking the chunks. Asked to do the same with letters, his span was six. The study, by Anders Ericsson, William Chase, and Steve Faloon, was published in Science in 1980 and is at the journal.
Why it matters for teaching
A limit of four has consequences for anyone explaining anything. A sentence with four clauses, an equation with five new symbols, a slide with seven bullet points, a list of instructions with six steps, each asks for more than the listener can hold, and the listener will drop something without knowing which. The practical rules follow: introduce one new thing at a time, name it, use it before the next arrives, and give the listener a structure, such as a numbered list or a diagram, that lets the pieces be grouped into chunks. Good textbooks have always done this by instinct; the figure of four says why it works.
Miller ended his paper by admitting that the seven of absolute judgement and the seven of memory span were probably a coincidence, and that the chunk, not the number, was the finding. Fifty years of work have agreed with him and changed his number. The fact that the famous figure is wrong and the paper is still right is a good lesson in how to read a classic.


