Dunbar’s number and how speaking is 2.8x better than picking fleas

20.56, Tuesday 5 Apr 2022

150, Dunbar’s number, is the natural size of human social groups. Robin Dunbar’s 1993 paper, where he put forward this hypothesis, is a great read – it’s got twists and turns, so much more in it than just the 150 number.

(If you design software for people to socialise or collaborate, like Slack or Google Docs, then what Dunbar says is useful to know! Also true if you build communities in Discords or DAOs, I reckon, good knowledge to have when structuring the spaces and processes for interaction.)

I’ve added a reference to Dunbar’s paper, Coevolution of neocortical size, group size and language in humans, at the bottom of this post. It’s not online but you can snag a pre-print PDF here.

The paper and the number are both super well-known.

BUT - I insist! - still not well-known enough in our software and design circles. Especially given there is a revitalisation and renewed interest in building and innovating with the social internet.

So I figured I would share my favourite bits.


Dunbar lists a bunch of places this 150 group size appears. To pick out a convincing selection…

  • 6500-5500 BC: the size of Neolithic villages in Mesopotamia are of about the same magnitude.
  • In South Dakota: the Hutterites regard 150 individuals as the limiting size for their farming communities: once a community reaches this size, steps are taken to split it into two daughter communities.
  • Academic communities: research specialities in the sciences tend to consist of up to 200 individuals, but rarely more.
  • Professional armies have a basic unit of about 150 men: This was as true of the Roman Army (both before and after the reforms of 104 BC) as of modern armies since the sixteenth century.
  • Work: the likelihood of having friends within the workplace reaches an asymptote at a shop size of 90-150 individuals.

It’s the number of people on your Christmas card list. When AOL Instant Messenger launched, it was the maximum allowable number of buddies.

(And the number of Pokemon in generation one… 151. Huh.)

Pretty compelling. Something to be explained.


The clever bit:

The catarrhine primates: Old World monkeys (baboons, macaques, mandrill, and ~130 other species) plus the apes… tailless simians including gibbons, gorillas, chimpanzees, and humans.

Dunbar’s insight was to look at the catarrhine primates and realise that these three factors are connected:

  • Group size
  • Time devoted to social grooming: a bonding mechanism (if you wanna have friends, you gotta spend time picking fleas)
  • Neocortex size, being the amount of brain available for tracking the social group, i.e. a limit on the number of relationships that an animal can keep track of in a complex, continuously changing social world

Dunbar gives equations that relate these.

Then:

If we extrapolate from the nonhuman primate regression, what group size would we predict for anatomically modern humans, given our current neocortex size?

Oh-ho, a prediction!

Equation (1) yields a predicted group size for humans of 147.8.

So there’s the observed number 150, right there in the size of the brain.

BUT THEN, A TWIST:

The group size predicted for modern humans by equation (1) would require as much as 42% of the total time budget to be devoted to social grooming.

We (humans) clearly don’t spend all that time on social grooming. There’s not the time in the day. It’s incompatible with resting, foraging, and staying in the shade on hot days. Chimpanzees are the most comparable to humans, and they have a social time budget of about only 15%.

So what gives?

Humans, says Dunbar, must have a method of social grooming that is 2.8x more effective than the method used by the nonhuman primates. But what is it?


What is our ultra efficient bonding mechanism, better than caring, grooming, and picking fleas? It is LANGUAGE.

The observed mean group size for chimpanzees (presumably the closest approximation to the ancestral condition for the hominid lineage) is 53.5 (Dunbar 1992a). Since the predicted size for human groups is 147.8, this implies that language (the human bonding mechanism) ought to be 147.8/53.5=2.76 times as efficient as social grooming (the nonhuman primate bonding mechanism).

Speaking is way better than grooming, which requires 100% attention and is one-on-one. But we can talk to more than one person at once! See: not only can speech be combined with almost every other activity (we can forage and talk at the same time), but it can also be used to address several different individuals simultaneously.

Dunbar’s suggestion is that language evolved as a ‘cheap’ form of social grooming, a way to increase group size. And there follows a cascade of consequences and speculations…


The interesting bit, for me, is about the “natural” size of a conversation group.

Dunbar’s prediction, based on the estimated efficiency gain versus chimps: human conversation group sizes should be limited to about 3.8 in size (one speaker plus 2.8 listeners).

And this holds up!

  • Looking at restaurant reservations: the mean size of 3070 groups was 3.8.
  • In a university refectory: the average number of people directly involved in a conversation (as speaker or attentive listener) reached an asymptotic value of about 3.4 (one speaker plus 2.4 listeners) and that groups tended to partition into new conversational cliques at multiples of about four individuals.

Which feels about right, right?

I mean, think of a sitting with friends round a dinner table. Two people, three people, four people, it’s one conversation. Five people, it’s still one conversation – just. At six it’s hard to maintain; the conversation often splits and oscillates between 4/2 and 3/3 modes.

The cognitive limit corresponds to how our ears and voices work.

It turns out that there is, in fact, a psycho-physical limit on the size of conversation groups. Due to the rate at which speech attenuates with the distance between speaker and hearer under normal ambient noise levels, there is a physical limit on the number of individuals that can effectively take part in a conversation. Sommer (1961), for example, found that a nose-to-nose distance of 1.7m was the upper limit for comfortable conversation in dyadic groups; this would yield a maximum conversation group size of five individuals with a shoulder-to-shoulder spacing of 0.5m between adjacent individuals standing around the circumference of a circle.

“Comfortable” conversation means background noise levels typical of both offices and city streets – our normal voice levels, our normal hearing, our normal comfortable personal social distance, our normal width of shoulders all combine to produce conversional groups of… 5 people.

Absolutely wonderful. It makes me laugh every time I read this bit.


Evidence for Dunbar’s Number in the analysis of 6 billion phone calls:

Dunbar actually doesn’t say that we devote “grooming time” to the whole social group of 150. Rather he says that the 150 is made up from welding together much smaller “primary networks”: coalitions, friendship groups. Intensive grooming (language, for humans) is reserved for close friends. Our intimate group is very small, averaging just five.

Dunbar suggests other group sizes too, in papers that follow his 1993 original…

Individuals, he says, generally have up to five people in the closest layer. The next closest layer contains an additional 10, the one beyond that an extra 35, and the final group another 100. So cumulatively, the layers contain five, 15, 50, and 150 people.

And this result is new to me:

Looking at some six billion calls made by 35 million people they did some number crunching and…

the average cumulative layer turns out to hold 4.1, 11.0, 29.8, and 128.9 users.

Ta-da! Dunbar’s number proved, close enough.

You can get the PDF on arXiv: Calling Dunbar’s Numbers (2016). I’ve included the full reference below.

Kinda amazing to have evidence for something that feels so intuitive (the average number of best friends/family) and that lends confidence in the discovery in the data of Dunbar’s number itself too.


BTW I found that second paper via Ethan Mollick (@emollick) on Twitter who daily shares and summarises fascinating papers and is 100% a must-follow.


Again, why this is relevant: if you’re designing systems for working in groups, whether that’s IRL workgroups and committees, or online chat groups, or software, the relevant numbers are 150 people who can be recognised over time, and approx 5 in a simultaneous conversation. That’s what it suggests to me anyway.

The numbers are just averages, of course, and we’re each individuals and you shouldn’t put too much weight on evo psych or be deterministic about this stuff, but what we can do is use these as springboards to provoke new feature ideas. Such as…

  • could a Figma document suggest how to subdivide itself once more than five people are involved? Could my Twitter auto-segment into groups of 150?
  • could we automatically adapt the interface of Zoom at the various Dunbar layers? Can we visually (and with interaction design) represent the structure of a 150 group “welded together” from smaller table-sized conversational groups, the two scales operating simultaneously?
  • in pseudonymous groups where everyone is represented by avatars and obscure names, such as web3 communities on Discords, what is the right level of detail to help our ancient recognition systems to kick in, so we can get to Dunbar’s number and all the helpful formal and informal networks and subgroups that arise?
  • when we’re building software-enabled co-ops or, in new language, creating governance and consensus systems for DAOs, could we optimise around Dunbar’s layers in order to avoid the inevitable bureaucratic requirements when we don’t – bureaucracy which is now perhaps revealed to be a social technology, a kind of relationships prosthetic for when self-organisation caps out.

(Those last two points relevant now the global public timelines of 2010s social media are evaporating into the unindexable Discords and WhatsApp groups of 2020s virtual private neighbourhoods.)

AND SO ON.


References

Dunbar, R.I.M., 1993. Coevolution of neocortical size, group size and language in humans. Behavioral and Brain Sciences 16, 681–694. https://doi.org/10.1017/S0140525X00032325

MacCarron, P., Kaski, K., Dunbar, R., 2016. Calling Dunbar’s Numbers. Social Networks 47, 151–155. https://doi.org/10.1016/j.socnet.2016.06.003

If you enjoyed this post, please consider sharing it by email or on social media. Here’s the link. Thanks, —Matt.

😴