One Hundred and Fifty,
Give or Take Five Hundred

Ground-truth seam · a point estimate, unpinned · verify.mjs 38/38 · every interval cited

You have heard that a human can hold about 150 relationships — "Dunbar's number." It is a single point on a straight line drawn through a few dozen primate species and stretched out to us. The line is real. What almost no retelling draws is the error bar — and once you draw it, the famous number stops being a fact and becomes a guess with a range you could drive a herd through. Draw both here, from the published data.

Where 150 comes from

In 1992 the anthropologist Robin Dunbar noticed that across primates, species with a bigger neocortex ratio — the neocortex's volume divided by the volume of the rest of the brain — tend to live in bigger social groups. He fit a straight line to the logarithms of the two, read it off at the human neocortex ratio (about 4.1), and got a number near 150. That is the whole derivation: a regression, extrapolated. Everything downstream — the "layers," the office-size advice, the Christmas-card studies — hangs on this one extended line.

Instrument 1 — operate Dunbar's own line

The blue dots are the real animals — the 36 primate genera Dunbar fit the line to (from his 1992 Table 1). The gold line is his published equation; our verifier reproduces its slope and intercept from those exact dots, using the reduced-major-axis method he used. Slide the neocortex ratio and read the group size the line predicts. At the human value it lands on ~150 — but watch where that read-off falls: out in the shaded zone, past every animal the line was built from, and put back the confidence interval Dunbar himself reported.

—

Dunbar's fitted line read-off & Dunbar's 95% CI

Log–log axes: neocortex ratio across the bottom, predicted group size up the side. The read-off at the human ratio (4.1) is an extrapolation — it sits past the primates the line was built from, exactly where a straight line is least trustworthy. Toggle the interval and even Dunbar's own figure becomes "100 to 231," not "150."

Instrument 2 — the famous number, next to its error bars

Here is 150 drawn as what it is: one estimate with an interval, beside the estimates a 2021 reanalysis got when it used a larger dataset and corrected for the fact that closely-related species are not independent evidence. Every bar is a published 95% interval. The dashed line is 150.

A log scale across the bottom (10, 100, 1000 people). Dunbar's own point (top, gold) already carries the interval 100–231; the phylogenetic methods below scatter from ~16 to ~109 with intervals reaching past 500. "150" is a dot; the honest answer is the width of the bars around it — verify.mjs confirms seven of the eight reanalysis intervals are wide enough to contain 150, and all eight point estimates sit below it.

But isn't 150 confirmed by real groups?

You will often read that 150 keeps showing up — Neolithic villages, Hutterite colonies that split near 150, a company of soldiers, the number of people you send holiday cards to. These are real observations, and genuinely interesting. But look at what they are: human group sizes that land somewhere in the tens-to-few-hundred range — exactly the range the wide interval already allows. An interval spanning roughly 2 to 520 is "confirmed" by almost any real human grouping you could name. Numbers scattered through a wide band do not narrow it; they sit inside it. And the "confirming" figures scatter themselves — the Hutterite mean Dunbar actually cites is 106.9, not 150; personal-network studies give around 290, nearly double. Compatible with 150 — and equally with 90, or 250.

The check

Every number on this page is recomputed or cited by research/dunbars-number/verify.mjs from data.mjs — Dunbar's own equation reproduced to give 147.8 at the human ratio, his reported 95% interval (100.2–231.1), the eight phylogenetic estimates and their 95% intervals from the 2021 reanalysis (verified against the paper's Table 1), and the numeric core (the Student-t factor pinned against standard statistical tables, ordinary least squares checked against an exact line). The page draws from a byte-identical copy of that data, held in place by a drift guard. 38/38 checks pass, offline.

$ node research/dunbars-number/verify.mjs
# recomputes the read-off, the intervals, and the numeric core
§1 reproduce the line           RMA of the 36 real genera → slope 3.40, intercept 0.087, read-off 147.8
§2 the numeric core            Student-t vs tables; OLS on an exact line; the band balloons
§3 the 2021 reanalysis         8 methods, points 16–109, 95%% CIs ~2–520; 150 swallowed
§4 the "confirmations"         every illustration falls INSIDE the wide interval
§5 drift guard                 the page embeds byte-identical data

What's solid, what's a myth, what's genuinely unsettled

Solid: across primates, social group size really does rise with neocortex ratio — the correlation is there (about three-quarters of the variance in the original sample), and it is a genuine, much-studied pattern, the "social brain hypothesis."

A myth: that 150 is a precise, hard cap — a measured constant of human cognition. It is a point read off an extrapolated line, and lines read that far past their data carry intervals wide enough to swallow most of the answers offered as "confirmation." Even the original author's interval was 100–231.

Genuinely unsettled: the true human figure, if a single number is even the right object. Correcting for shared ancestry, the 2021 reanalysis put the honest range so wide that, in its authors' words, a cognitive limit on human group size "cannot be derived in this manner" and "specifying any one number is futile." Defensible modern point estimates run from about 16 to about 109. That is not a scandal — it is what extrapolating a noisy biological regression to one species honestly looks like. The number stuck because it was memorable, not because the error bars were small.

Sources & honest apparatus

Free choices named: Instrument 1 plots the real 36-genus data from Dunbar's Table 1 and draws his own published line; our verifier confirms that reduced-major-axis regression of those points reproduces his slope (3.40 vs 3.389) and intercept (0.087 vs 0.093) and the ~150 read-off — we reproduce his method, we do not re-tune it. The interval shown is his own reported 95% CI, and the phylogenetic intervals in Instrument 2 are the published ones — we plot uncertainty that was already computed, we do not invent it. "Neocortex ratio" follows Dunbar's definition (neocortex ÷ rest of brain), and we use his printed ratio column; four genera whose printed ratio doesn't match a naive recomputation (his species-averaging, or paper typos) are flagged in research/dunbars-number/facts.md, never rounded away.