Dunbar's number came back as anything from 2 to 520

One hundred and fifty relationships, supposedly set by the size of the neocortex. A 2021 reanalysis of the same primate data got 16, 42, 69 and 109.

You have met the number 150. It turns up in management books explaining why a company should split at a certain headcount, in essays about why a village feels different from a suburb, and all over the design notes of software built for communities. One hundred and fifty stable relationships, the ceiling supposedly set by the size of the human neocortex.

The Swedish Tax Agency once reorganised its offices to stay under it. That example comes from the paper that later took the number apart.

In 2021 three researchers in Stockholm ran the original calculation again. They got 16. They also got 42, 69 and 109, depending on the statistics used and on which measure of the brain went into them. The confidence intervals ran from about 2 to 520.

Where the number came from

Robin Dunbar's argument, made in 1992 and 1993, was elegant enough to travel further than most findings in his field. Across primate species, the relative size of the neocortex tracks the size of the group the species lives in. Bigger thinking layer, bigger troop. Fit a line through the primates, find where humans sit on it, read off the group size. The answer landed near 150 and it stayed there for thirty years, through Gladwell, through a thousand blog posts, and into the floor plans of a government agency.

Elegance is not evidence, though it is excellent for citation counts. The 2021 reanalysis notes in a footnote that Dunbar's original paper had by then passed 2,500 citations on Google Scholar.

What the reanalysis did

Patrik Lindenfors, Andreas Wartel and Johan Lind published “'Dunbar's number' deconstructed” in Biology Letters in May 2021. No new experiment. They took the same logic, better primate brain data — volumes averaged from three sources, then everything repeated using only the most carefully collected of the three — and current phylogenetic statistics, and asked what number falls out.

Six numbers fell out: two families of statistics, three choices of what to measure in the brain. The Bayesian models gave 69.2, 108.6 and 79.8. The generalised least-squares models gave 16.4, 42.0 and 23.6. The 95 per cent confidence intervals were 3.8–520.0 for the first family and 2.1–336.3 for the second.

One detail in their results section matters more than the headline and gets quoted less. These are estimates of average group size, not of an upper limit — and that was true of Dunbar's 150 as well. The authors note that if you did want a ceiling out of this kind of reasoning, you would read it off the top of the confidence interval, and that those tops “indicate upper limits far exceeding 150 in almost all cases”.

Their closing line is not hedged: “'Dunbar's number' is a concept with limited theoretical foundation lacking empirical support.” A sentence earlier they write that they hope, “though perhaps futile”, that the paper ends the use of the number in science and in popular media.

The part that usually gets lost

A confidence interval running from two to five hundred is not a discovery about human sociability. Nobody's shown you can keep up with 500 friends. What the interval describes is the method: take one species, place it on a regression line built from other species, and the uncertainty you inherit is wide enough that the output is compatible with almost any answer you might have wanted in advance.

That indicts the derivation, not the observation.

Because 150 did not come only from that line. In 2003 Hill and Dunbar counted Christmas cards — who people actually sent them to, and how many people lived in the receiving households. Maximum network size averaged 153.5. Counting only the individuals a sender had explicitly been in contact with, 124.9. Christmas cards are a strange instrument and the study has its own weaknesses, but it measures behaviour instead of extrapolating from brain tissue, and it landed in roughly the same place.

Thirty years of popular writing has fused two questions into one. Does something around 150 describe a real ceiling in human networks? And can that ceiling be calculated from the size of a brain region? The Stockholm paper answers the second. Not by this route. On the first it points at other researchers who have measured human group sizes and reported a wide spread of other numbers, which is weaker ground than the abstract's confidence suggests.

Why this matters if you are trying to build something

The practical use of Dunbar's number has always been permission to stop. A team reaches 150, so you split it. A neighbourhood group gets there and someone starts a second one. The number does the deciding and nobody has to win an argument.

If the number can't be derived, that shortcut is gone. What's left is the older and slower method: watch the thing you built. A group where people have stopped recognising each other's names is too big, whether it has 80 members or 300. A group where the same four people do everything has a different problem entirely, and no headcount will warn you about it.

We keep a map of good deeds; it has no optimal number of dots per town. The neighbourhoods that work are the ones where somebody notices when a stranger needs a hand. Nobody has fitted a regression line to that.

Sources: Lindenfors, P., Wartel, A., & Lind, J. (2021). 'Dunbar's number' deconstructed. Biology Letters, 17(5), 20210158 — figures and quotations taken from the abstract, the results section and table 1. Hill, R. A., & Dunbar, R. I. M. (2003). Social network size in humans. Human Nature, 14(1), 53–72.