The ninety-second article was about a claim carried by a shape. This one is about a claim carried by a single number, where the uncertainty around that number was published twice, by different people, thirty years apart, and travelled neither time.
Key Takeaway
The 2021 abstract states that two methods produced "wildly different numbers," with approximations of 69 to 109 and 16 to 42, and that "enormous 95% confidence intervals (4–520 and 2–336, respectively) imply that specifying any one number is futile."[1] An encyclopedia records that the 1992 original predicted 148 with an interval of 100 to 230, and that its author considered it exploratory[4].
The Verdict, Stated First
Five claims, in descending order of confidence.
One. The paper's conclusion is unusually blunt. Specifying any one number is described as futile, and a cognitive limit is said to be underivable by this method.
Two. The two 2021 methods do not agree with each other. Their point ranges of 69 to 109 and 16 to 42 do not overlap at all, which is a fact about the paper rather than an inference.
Three. The original already carried a wide interval and an author's own caution, on the encyclopedia's account, and neither survived into circulation.
Four. On our own arithmetic the interval is wide enough to be nearly uninformative, spanning 516 people, and any business decision it supports would be supported equally by its opposite.
Five. One of the authors has commented directly on the organisational use of the number, and we quote him, because his objection is one a business owner would not think of.
Those five are ordered by how much we would stake on them, ours, and the first four rest on published text while the fifth rests on a press release.
Our Grades For These Claims
Applying the scheme from the first article in this series, and the sourcing here is the strongest in some time.
Grade A for the 2021 findings, from an abstract obtained verbatim from the journal's own archived copy and confirmed against two further repository records.
Grade A for the authors' explanation of why the extrapolation fails, quoted directly in their university's own release.
Grade B for the 1992 original's interval and its author's caution, which reach us from an encyclopedia rather than from the paper.
Grade D for the 1992 paper itself, which we did not obtain and about whose contents we say almost nothing.
Grade A for our own arithmetic, with one calculation that rests on an invented baseline, which we flag in the section rather than only in the limits note.
The spread between the first grade and the fourth is the shape of this article, ours. We are confident about the reanalysis and largely ignorant about the original, which means the strongest claims here are about what cannot be known rather than about what was originally done.
A Note On Method
Everything here is verified to August 2026.
We obtained the 2021 abstract verbatim from the journal's archived copy[1] and confirmed it against a research institute's publication record[2] and an open data repository[6].
We obtained direct quotations from two of the three authors in their university's own news release[3], which is an institutional source rather than a peer-reviewed one and is flagged.
The 1992 original's interval and its author's characterisation reach us from an encyclopedia[4], flagged at every use.
We did not obtain the 2021 paper's full text beyond its abstract and opening, nor the 1992 paper at all.
All arithmetic is ours. Every interval quoted is the papers'; every application of them is our own.
This article discusses research on group size and on statistical uncertainty. It is not organisational design advice.
The Claim As Used
What the number does in business, before we look at where it came from.
The abstract describes the belief as: "humans possess a cognitive capacity that is limited to keeping track of and maintaining stable relationships with approximately 150 people."[1]
Four observations, ours.
In business it becomes a cap. Split the office at 150, keep business units below it, cap a team or a community at the figure.
It is attractive because it converts a hard question into an arithmetic one. How large should a unit be is difficult; is it under 150 is not, and a number that dissolves a difficult question will be adopted.
The word "approximately" is in the abstract's description of the belief and does no work in practice, because a cap has to be a specific figure and the specific figure used is 150.
And the claim is about a cognitive limit, which is a much stronger thing than an observed average. A limit says you cannot; an average says people mostly do not.
That last distinction does most of the work and is almost never observed, ours. A limit licenses a cap, because there is no point building past a ceiling.
An average licenses nothing of the kind. A firm could sit above an average deliberately and pay for it, which is an ordinary commercial decision, and the framing as a limit removes that option before it is considered.
Where The Number Comes From
The method, described in the 2021 abstract's own words.
It records that the number "originates from an extrapolation of a regression line describing the relationship between relative neocortex size and group size in primates."[1]
An encyclopedia adds that the original used a regression equation on data for 38 primate genera and predicted a human mean group size from it[4]. Encyclopedia, flagged.
Four observations, ours.
The operative word is "extrapolation." The line was fitted to primates and the human value was read off beyond the data.
That is the same operation the ninety-first article used as a reductio, taking a per-minute figure to fifteen minutes to show the answer was absurd. Here it is the method rather than the test of it.
Extrapolation is not automatically wrong and it is always fragile, because nothing in the data constrains the shape of the line outside the data, and a small change in slope becomes a large change in the extrapolated value.
And the quantity being extrapolated to is a species outside the sample, which is a stronger form of the problem than extending a range within one.
A concrete way to feel the fragility, ours and invented. Suppose the true relationship curves gently downward beyond the primate range rather than continuing straight.
Nothing in a dataset of primates could detect that, because the curvature would occur entirely outside the data, and the fitted line would carry on regardless and produce a confident and wrong answer.
The Original Carried An Interval
The detail that reframes the whole story.
An encyclopedia records that the 1992 work "predicted a human 'mean group size' of 148 (casually rounded to 150), a result he considered exploratory because of the large error measure (a 95% confidence interval of 100 to 230)."[4]
This is an encyclopedia and not the paper, flagged, and we did not obtain the 1992 original to verify it.
Four observations, ours.
The point estimate was 148, and the phrase used for what happened next is "casually rounded."
The interval was 100 to 230, which is more than a factor of two from end to end, and which appears in no business use of the number we have encountered.
And the author himself is recorded as considering the result exploratory because of the large error measure, which is a scientist doing exactly the right thing.
So this is not a case of research being wrong. It is a case of a caveat published alongside a figure and discarded in transit, which is now the fifth consecutive article in this series to find that pattern.
Five in a row is enough to stop calling it a coincidence, ours. The qualification is dropped because it is the part that does not fit on a slide, and every step of transmission has a length limit.
The Word Exploratory
What it means and what it does not. Ours.
Four observations.
Exploratory means a result offered for further investigation rather than for use, which is a standard and honourable category in research.
It is the opposite of what a business needs from a number. A cap on team size is a decision, and a decision cannot be exploratory, so the moment the figure entered management practice it was doing something its author had declined to claim for it.
Nobody is at fault in that sentence, which is worth saying. An author who labels a result exploratory has discharged their obligation, and what happens to the number afterwards is not in their control.
And the practical lesson is available to any reader. When a figure is quoted, ask what the paper called it, because the label travels less well than the number and is more informative.
The 2021 Reanalysis
The source.
Lindenfors, P., Wartel, A., and Lind, J. (2021), 'Dunbar's number' deconstructed, Biology Letters, 17(5), 20210158, 5 May, DOI 10.1098/rsbl.2021.0158[1].
The authors are affiliated with an institute for futures studies and a centre for cultural evolution[1], and the paper's keywords include phylogenetic comparative studies, social evolution, brain evolution, primates and mammals[1].
Four observations, ours.
The paper describes its own aim in its introduction as "attempting to make a decisive deconstruction of the empirical basis" of the number[1], which is a stated intention rather than a neutral posture and should be weighed as such.
Against that, the method is to redo the original analysis with larger datasets and modern statistical technique, which is a straightforward and checkable thing to do.
The underlying data was published openly, which we confirmed from a repository record[6], so the analysis is available for anybody to contest.
And the finding is not that the number is wrong. It is that no number can be obtained this way, which is a different and stronger claim.
Stronger and, oddly, harder to publicise, ours. A refutation gives the public a new number and this gives them none, which is the asymmetry one of the authors names later in his own words.
The Abstract
Quoted at length, because the wording carries the argument.
It states: "Here, we test if there is statistical support for this idea. Our analyses on complementary datasets using different methods yield wildly different numbers. Bayesian and generalized least-squares phylogenetic methods generate approximations of average group sizes between 69–109 and 16–42, respectively. However, enormous 95% confidence intervals (4–520 and 2–336, respectively) imply that specifying any one number is futile. A cognitive limit on human group size cannot be derived in this manner."[1]
Four observations, ours.
"Wildly different" and "enormous" are not hedged academic language, and their presence in an abstract is itself a signal.
The two findings are separable and both matter. The methods disagree with each other, and each is individually uncertain, which are different problems.
The final sentence limits the claim carefully. "Cannot be derived in this manner" says the method fails, not that no limit exists.
And that distinction is one this series has had to make repeatedly. A failed method leaves the question open, which is a less satisfying result than a refutation and is what the evidence supports.
Two consequences of leaving it open, ours. Somebody may yet establish a limit by another route, and nothing here forecloses that.
And in the meantime the honest position is that the question is unanswered, which is a worse basis for a reorganisation than most people realise they are working from.
Wildly Different Numbers
The first finding, examined. Ours.
The two approaches produced approximations of 69 to 109 and 16 to 42[1].
Four observations.
These ranges do not overlap. Forty-two is below sixty-nine, with a gap of twenty-seven people between them.
That is a fact about the paper rather than an inference from it, and it is the more damaging of the two findings in our view, because it concerns reproducibility rather than precision.
Both ranges are also below 150, which is the direction people expect least. The reanalysis did not merely widen the uncertainty; its central estimates came out smaller.
And a firm using the number would find no comfort in either. A cap of 42 and a cap of 109 are different organisational designs, and both are different from 150.
One reason that direction is worth naming explicitly, ours. A firm that adopted 150 believing it conservative was not being conservative on these figures, and would have been running units well above both modern central estimates.
The Intervals
Our own arithmetic on the papers' own figures, laid out for comparison.
The 1992 original: point 148, interval 100 to 230, a ratio from high to low of about 2 times.
The 2021 Bayesian analysis: interval 4 to 520, a ratio of 130 times.
The 2021 least-squares analysis: interval 2 to 336, a ratio of 168 times.
Four observations.
A ratio of 130 means the upper bound is a hundred and thirty times the lower, which is not a margin of error in any ordinary sense.
The intervals include values that would be absurd in either direction. Four people, and five hundred and twenty, are both inside the ninety-five percent range on one method.
And it is worth noting that the newer, better-powered analysis produced far wider intervals than the original, which is the opposite of what more data usually does.
That inversion is informative. It suggests the original's narrower interval came from a method that understated its own uncertainty, which is our inference and not the paper's stated finding.
We hold that inference loosely and would note the alternative, ours. Different datasets can legitimately produce different uncertainty, and without the full text we cannot distinguish a better-calibrated method from a harder dataset.
How Much Was Ruled Out
Our own arithmetic, and the one calculation here that rests on a baseline we invented, which we flag before giving it rather than after.
Suppose that before any analysis you thought a human group limit could plausibly be anywhere from 1 to 1,000 people. That range is ours and arbitrary, and every percentage below moves if you choose a different one.
On that baseline: the 1992 interval of 100 to 230 spans 130 people and rules out 87 percent of the range. The 2021 Bayesian interval spans 516 and rules out 48 percent. The least-squares interval spans 334 and rules out 67 percent.
Four observations.
The objective fact underneath is the width, which needs no baseline. The Bayesian interval spans 516 people, and that is simply what the paper reports.
The percentage is where our invention enters, and it is a real limitation. A reader who thinks group limits could only plausibly range from 50 to 250 would compute completely different figures, and would find the 2021 intervals even less informative rather than more.
What survives any choice of baseline is the comparison between rows. The best available analysis rules out less than the original did, whatever range you start from, because its interval is four times wider.
And that is the honest form of the point. We can tell you the interval is wide and we cannot put a percentage on how uninformative it is without inventing something, which we have done visibly rather than quietly.
Two notes on why we kept the calculation at all, ours. Deleting it would have hidden a useful comparison, since the relative ordering of the three rows holds under any baseline a reader might pick.
And the alternative to an invented baseline here is silence, because no published prior exists for what a human group limit could plausibly be. We would rather show the assumption than omit the section.
Do The Two Methods Agree
The comparison, ours, and it produces an odd result.
The point ranges, 69 to 109 and 16 to 42, do not overlap. The confidence intervals, 4 to 520 and 2 to 336, overlap almost completely.
Four observations.
So the two methods disagree about the answer and agree that neither of them knows it, which is a strange and instructive combination.
If only the point estimates had been published, a reader would conclude the field was in dispute. The intervals show that there is nothing to dispute, because neither estimate is precise enough to conflict with anything.
Which is exactly the argument for publishing intervals. They convert an apparent disagreement into an accurate statement of shared ignorance, and the second is more useful.
And it inverts how these things are usually reported. A news account would lead with the conflicting estimates, and the conflict is the least important thing in the paper.
Which suggests a reading habit worth acquiring, ours. When two studies are reported as disagreeing, check whether either is precise enough to disagree, because a great deal of apparent scientific controversy dissolves at that step.
Why The Extrapolation Fails
The authors' own explanation, quoted from their university's release.
One author is quoted saying: "Other primates' brains do not handle information exactly as human brains do, and primate sociality is primarily explained by other factors than the brain, such as what they eat and who their predators are. Furthermore, humans have a large variation in the size of their social networks."[3]
A co-author is quoted: "It is not possible to make an estimate for humans with any precision using available methods and data."[3]
An institutional news release rather than a peer-reviewed source, flagged, though the speakers are the paper's own authors.
Four observations, ours.
The strongest of the three reasons is the middle one. If primate group size is mainly driven by diet and predation, then the regression is fitting a relationship whose main causes were left out, and extrapolating it carries those omissions along.
The third reason matters commercially and is easy to overlook. Humans vary enormously in network size, so even a correct average would say little about any particular workforce.
And the first reason is the one that undermines the whole approach rather than the estimate. If human brains handle information differently, the line does not extend to us at all, regardless of how precisely it is fitted.
We would note that these are the authors' assertions in a press release rather than findings we obtained from the paper, and a reader should weigh them accordingly.
Though one of the three is checkable without any paper, ours. That humans vary enormously in social network size is something any reader can confirm from their own acquaintance, and it alone is fatal to using a single figure as a design constraint.
What The Authors Say About Business Use
The passage that made us write this article, and an objection a business owner would not think of.
One author is quoted on organisations restructuring around the number: "This reorganization would then be based on the implicit but hopefully unintended assumption that their employees have neither family nor friends outside of work."[3]
Four observations, ours.
The argument is simple once stated and we had not seen it. If the limit is a total across a person's whole life, then work relationships compete with everything else, and a firm sizing a unit at the limit has assumed it gets all of them.
A person with a family, old friends and a neighbourhood has spent much of their allowance before arriving at the office, on the theory's own terms.
Which means that even taking the number at face value, the correct workplace figure would be some unknown fraction of it, and nobody has ever proposed what fraction.
And that is a flaw in the application rather than in the research. The theory was never about workplaces, and the business use added an assumption the theory does not contain.
Two further consequences the objection implies, ours. The workplace share would vary by person, being smaller for somebody with a large family than for somebody without one.
And it would change over a career, which means even a correct fraction would not be a fixed number and could not function as a cap.
Not Quite As Entertaining
A remark from one of the authors that describes this entire series.
He is quoted: "I think Dunbar's number is widely spread, also among researchers, since it's so easy to understand. Our claim that it is not possible to calculate a number is not quite as entertaining."[3]
Four observations, ours.
"Also among researchers" is the part worth pausing on. This is not a story about the public misunderstanding science; the number circulates inside the field too.
The asymmetry he names is the mechanism this series keeps documenting. A number is memorable and an interval is not, and the ninety-second article found the same asymmetry between a diagram and a caveat.
And it explains why corrections lose. "One hundred and fifty" is a fact you can repeat at dinner; "no number can be derived this way" is a methodological objection, and only one of those travels.
Which is, in a sentence, the reason we publish these at eight thousand words rather than as a correction. The interesting content is in why the number cannot be had, and that requires the length.
And it explains a pattern in this series we did not plan, ours. Almost every article here is longer than the finding it examines, because the finding is a sentence and the reason it was misread is not.
What The Interval Does To An Org Chart
Our own arithmetic on an invented firm, to show what the uncertainty means in practice.
Take a company of 600 people capping its business units at the limit. The number of units implied by each value:
At a limit of 4: 150 units. At 29: 21 units. At 89: 7 units. At 150: 4 units. At 336: 2 units. At 520: 2 units.
Four observations.
The interval spans an organisation of 150 units and one of 2. Those are not variations on a plan; they are different companies.
Even confining yourself to the two 2021 point ranges gives 6 to 9 units on one method and 15 to 38 on the other, which is still a fundamental difference in structure.
And on the 1992 figures alone, which are the friendliest available, a cap of 150 gives four units and a cap of 230 gives three, with both inside the published interval.
So the number cannot bear the weight put on it even at its most generous. A decision that changes when you move within the confidence interval was not supported by the estimate, and this is the general test.
Two properties make that test worth adopting permanently, ours. It is arithmetic rather than judgement, so two people applying it to the same figure get the same answer.
And it frequently passes, which is the part people expect least. Many decisions are robust across an interval, and discovering that is as useful as discovering the opposite because it ends an argument about precision that never mattered.
The Rule That Does Not Cite It
A comparison worth making, because one well-known team-size rule is built the opposite way. Ours.
Four observations.
The rule that a team should be small enough to be fed by two pizzas cites no research at all, and is transparently a rule of thumb somebody found useful.
That makes it more honest than a cap of 150, not less, because it claims exactly the authority it has. Nobody defends it by pointing at a regression.
And it is testable in the only way that matters. A firm can try it and see whether meetings improve, which is an experiment available to any owner and does not require a primate dataset.
Though we would apply the eighty-eighth article's caution to that experiment, ours. One team over one quarter tells you very little, and the improvement a firm notices after a reorganisation is the hardest kind of evidence to read.
So the ranking is the reverse of what the sourcing suggests. An unsourced rule of thumb you can test beats a sourced number whose interval spans 516 people, and the second borrows a credibility the first does not claim.
What Would Change Our Mind
Stated in advance, because a critique that nothing could answer is not a critique. Ours.
Four things.
A direct measurement rather than an extrapolation. If somebody counted stable relationships in a large human sample and found a ceiling, that would be evidence of a kind the regression cannot produce.
A reply from the original author addressing the intervals specifically. We did not search for one and cannot say whether it exists, which is a gap in this article.
An error found in the 2021 analysis. Its data was published openly[6], so this is a live possibility rather than a rhetorical concession.
And evidence that the workplace figure behaves differently from the whole-life one, which would rescue the business application even if the underlying number stayed uncertain.
We would note what is not on that list, ours. Popularity is not on it, and neither is how many organisations already use the number, both of which are the arguments most often offered in its defence.
What Actually Survives
Our reading, stated directly.
Five statements.
No single number can be derived from the primate regression, on the 2021 paper's own conclusion.
The two modern methods disagree, with point ranges of 69 to 109 and 16 to 42 that do not overlap.
Both modern intervals are enormous, at 4 to 520 and 2 to 336, and the paper calls specifying any one number futile.
The original carried an interval of 100 to 230 and its own author's label of exploratory, on an encyclopedia's account we did not verify against the paper.
And the workplace application adds an assumption the theory does not make, namely that employees have no relationships outside work.
Those five are what we would defend, ours, and the first three rest on a published abstract we obtained verbatim, which is the strongest sourcing position this series has been in for some weeks.
Group Size Still Matters
Because this could be read as saying team size is arbitrary, which would be wrong. Ours.
Four observations.
Nothing here says group size does not affect how organisations work. It says one particular route to a specific number does not survive.
The lived observation behind the appeal is real. Something changes when an office grows past the point where everyone knows everyone, and most people who have worked in a growing firm have noticed it.
What is not available is a number to put on it, or a claim that the number is a cognitive limit rather than a feature of how a particular business is run.
And the practical alternative is unglamorous and better. Watch for the symptoms you actually care about, which are coordination cost, duplicated work and decisions taking longer, and act when you see them rather than at a threshold.
One advantage of symptoms over thresholds is worth stating plainly, ours. A symptom is measured in your own firm, so it already incorporates your industry, your tooling and how your people work, none of which a primate regression knows about.
Point Estimates And Intervals
The transferable lesson. Ours.
Four observations.
A point estimate without an interval is a claim without a strength, and this article is the clearest case we have found: the same estimate is either useful or futile depending on a number nobody quotes.
The test we would offer is the one from the org chart section. Move to each end of the interval and see whether your decision changes. If it does, the estimate did not support the decision.
That test is cheap and is almost never applied, because the interval usually is not in the room, having been dropped somewhere between the paper and the slide.
And the asymmetry is systematic rather than accidental. A point estimate is a single number and an interval is two, so every simplification drops the interval first, and simplification happens at every step of transmission.
The Rounding
A small detail with a general lesson. Ours.
The encyclopedia records the prediction as 148, "casually rounded to 150"[4].
Four observations.
The rounding is trivially small, at about one percent, and objecting to it would be pedantry.
What it does is not trivial. 148 looks like a measurement and 150 looks like a fact, and the second is easier to repeat and harder to question.
A round number also invites use as a threshold in a way an odd one does not. Nobody caps a business unit at 148, and the roundness is part of why the figure became operational.
And we would generalise it cautiously, as our own observation. Suspicion is warranted when a research figure arrives round, because measurements rarely are, and the rounding usually happened during transmission rather than in the paper.
Two cheap tests follow from that, ours. Ask what the unrounded figure was, and if nobody knows, the number has travelled far enough that the interval is certainly gone too.
And notice whether the rounding went toward a threshold you were already inclined to use, which is where a small rounding does real work.
Where Else This Applies
Other numbers of the same shape. Ours, and offered as reasoning.
Four observations.
Any figure quoted as a threshold has an interval somewhere, and the threshold use is the one most damaged by a wide one.
Business examples are easy to generate and we are not asserting the intervals are wide, only that they exist. Customer acquisition cost benchmarks, industry-standard margins, recommended runway, typical churn, and each is usually quoted as one number.
The right question is always available and rarely asked. What is the range, and does my decision change across it?
And where the interval genuinely is not available, that is itself the answer. A benchmark with no published uncertainty should be treated as an anecdote with a number attached, which is a harsh phrasing and a defensible one.
One qualification we owe that phrasing, ours. Plenty of useful benchmarks have no published interval because nobody computed one, rather than because the underlying data was weak.
The remedy in that case is not to discard the figure. It is to treat it as a starting point rather than a threshold, and to notice that the distinction between the two is exactly what an interval would have told you.
Your Own Numbers
The application to figures a firm produces itself. Ours, and not organisational design advice.
Four points.
Your own averages have intervals too, and the eighty-eighth article's arithmetic showed how wide they are at small sample sizes.
A firm with forty staff computing an average anything has a figure whose interval would embarrass it if computed, and the figure is nonetheless used to make decisions.
The cheap discipline is to state a range whenever you state an average. Reporting "churn was 8 percent" and "churn was somewhere between 4 and 13 percent" are the same information, and only the second is honest about what you know.
And the discipline has a side effect worth having. A range makes it obvious when a decision does not depend on the number at all, which is frequently the case and saves the argument.
Which is the quiet argument for intervals generally, ours. They are usually presented as a way of being cautious and they work at least as often as a way of stopping caution, by showing that a decision holds across everything the data permits.
Bibliographic Note
The series keeps a count, and this article produced two, both from the same source.
The 2021 paper is Biology Letters, 17(5), article 20210158, on the journal's own record[1]. One reference list gives it as 17(3), article 20200748[5], which is a different issue and a different article number.
The same source titles it "The limits of friendship: Dunbar's number deconstructed" where the journal gives "'Dunbar's number' deconstructed"[1][5].
Three observations, ours.
The article number variant is the most consequential we have recorded, because 20200748 would resolve to a different paper or to nothing, rather than merely to a harder search.
The title variant reads like a subtitle that was never used, and we cannot tell whether it comes from a preprint, a press headline or an error.
One practical consequence, ours. A reader chasing that citation would search the wrong issue for the wrong article number under the wrong title, and would likely conclude the paper does not exist rather than that the citation is wrong.
Which is why we now record these. A variant that fails silently is worse than one that fails loudly, and an article number is the quietest failure of the three.
That brings the running count of bibliographic variants across this series to forty-one.
What To Do
Do not cap anything at 150. The 2021 paper concludes that specifying any one number is futile and that a cognitive limit cannot be derived this way.
Ask for the interval before using any figure as a threshold. Here the intervals are 4 to 520 and 2 to 336, and the original's was 100 to 230.
Apply the movement test. Move to each end of the interval and see whether your decision changes; if it does, the estimate did not support the decision.
Note when two methods disagree and both are uncertain. Here the point ranges do not overlap and the intervals almost entirely do, which converts an apparent dispute into shared ignorance.
Remember that the limit, if it existed, would be a whole-life total. A firm sizing a unit at the number has assumed its employees have no relationships outside work.
Be suspicious of round research figures. The prediction was 148 and the rounding to 150 is what made it operational.
Watch the symptoms rather than a threshold. Coordination cost, duplicated work and slower decisions are observable in your own firm and a number derived from primates is not.
And state a range whenever you state an average of your own, which is the same information and is honest about what you know.
The Limits Of This Analysis
Several caveats matter. This article discusses research on group size and on statistical uncertainty and is not organisational design advice; the applications are our own reasoning and untested. Everything is verified to August 2026. We did not obtain the 1992 original at all, and its point estimate of 148, its interval of 100 to 230, its author's characterisation of the result as exploratory, and the detail about 38 primate genera all reach us from an encyclopedia, flagged at every use; if that account is inaccurate, several sections here would need revision, and this is the weakest link in the article. We obtained the 2021 paper's abstract and the opening of its introduction and not its full text, so we cannot describe its datasets, its models, or how it handled the differences between the two methods. The authors' explanations of why the primate extrapolation fails, and the remark about employees having relationships outside work, come from an institutional news release rather than from the paper, flagged; they are direct quotations from the authors and are not peer-reviewed statements. The paper describes its own aim as a decisive deconstruction, which is a stated position rather than a neutral one, and a reader should weigh that alongside the fact that its data was published openly. All arithmetic is ours. Every interval is the papers'; the ratios, the widths and the organisational implications are our own constructions. One calculation rests on a baseline we invented, the assumption that group limits could plausibly range from 1 to 1,000, and every percentage in that section moves if a reader chooses differently; we flagged it in the section rather than only here. The 600-person firm is invented. And our inference that the original's narrower interval reflects a method understating its uncertainty is our reasoning and not a finding of either paper.
Frequently Asked Questions
What did the 2021 reanalysis find?
Did the original claim more certainty than that?
What does an interval that wide actually mean?
Do the two 2021 methods agree with each other?
Why does the primate extrapolation fail?
What is wrong with using it for team size?
So how should a firm decide unit size?
References
- Archived journal copy of Lindenfors, P., Wartel, A., & Lind, J. (2021), 'Dunbar's number' deconstructed, Biology Letters, 17(5), 20210158, published 5 May 2021, DOI 10.1098/rsbl.2021.0158, PMID 33947220, reproducing the abstract in full: that a widespread and popular belief posits humans possess a cognitive capacity limited to keeping track of and maintaining stable relationships with approximately 150 people; that this influential number originates from an extrapolation of a regression line describing the relationship between relative neocortex size and group size in primates; that the authors test whether there is statistical support for the idea; that their analyses on complementary datasets using different methods yield wildly different numbers; that Bayesian and generalised least-squares phylogenetic methods generate approximations of average group sizes between 69 and 109 and between 16 and 42 respectively; that however enormous 95 percent confidence intervals of 4 to 520 and 2 to 336 respectively imply that specifying any one number is futile; and that a cognitive limit on human group size cannot be derived in this manner. The record also carries the authors' affiliations, the paper's keywords, and the opening of its introduction, in which the authors describe themselves as attempting to make a decisive deconstruction of the empirical basis of the number. Note: the journal's archived copy, and our source for every figure attributed to the 2021 paper. We obtained the abstract and the opening of the introduction and not the full text. ncbi.nlm.nih.gov
- Research institute's publication record for the same article, reproducing the abstract identically, including the ranges of 69 to 109 and 16 to 42 and the 95 percent confidence intervals of 4 to 520 and 2 to 336, and the conclusion that a cognitive limit on human group size cannot be derived in this manner. Note: an institutional publication record, used to confirm the abstract independently of the journal copy. iffs.se
- University news release on the study, quoting one author that other primates' brains do not handle information exactly as human brains do, that primate sociality is primarily explained by other factors than the brain such as what they eat and who their predators are, and that humans have a large variation in the size of their social networks; quoting a co-author that it is not possible to make an estimate for humans with any precision using available methods and data; recording that when the researchers repeated the original analyses using modern statistical methods and updated data the results were simultaneously much larger and far lower than 150, that the average maximum group size often turned out to be lower than 150, and that the main problem was 95 percent confidence intervals between 2 and 520 people; quoting an author that a reorganisation around the number would be based on the implicit but hopefully unintended assumption that employees have neither family nor friends outside of work; and quoting the same author that the number is widely spread, also among researchers, since it is so easy to understand, while the claim that it is not possible to calculate a number is not quite as entertaining. Note: a university news release, NOT a peer-reviewed source, flagged at every use, though the statements are direct quotations from the paper's own authors. Our source for the explanations of why the extrapolation fails and for the remark about workplace reorganisation. su.se
- Encyclopedia entry on the number, recording that the original author used the correlation observed for non-human primates to predict a social group size for humans; that using a regression equation on data for 38 primate genera he predicted a human mean group size of 148, casually rounded to 150; and that he considered this result exploratory because of the large error measure, a 95 percent confidence interval of 100 to 230. It cites Dunbar, R. I. M. (1992), Neocortex size as a constraint on group size in primates, Journal of Human Evolution, 22(6), 469–493. Note: an encyclopedia, NOT an academic source, flagged at every use. Our only source for the 1992 paper's point estimate, its interval and its author's own characterisation of it, none of which we verified against the paper. This is the weakest link in the article. en.wikipedia.org
- Educational website article on the number, describing it as based on a correlation between the size of primates' neocortex and the number of individuals with whom they can maintain stable relationships, and citing Dunbar, R. I. M. (1992), Journal of Human Evolution, 22(6), 469–493, and Lindenfors, P., Wartel, A., and Lind, J. (2021), The limits of friendship: Dunbar's number deconstructed, Biology Letters, 17(3), 20200748. Note: an educational website, NOT an academic source, flagged. Recorded principally as the source of two bibliographic variants: it gives the 2021 article as issue 3 with article number 20200748 where the journal gives issue 5 with article number 20210158, and adds a subtitle to the title that the journal does not carry. cursus.edu
- Open data repository record for the study's underlying dataset, published April 2021, reproducing the abstract including the ranges and the 95 percent confidence intervals, and confirming the authors' institutional affiliations. Note: a data repository record, used to confirm the abstract a third time and to establish that the analysis data was published openly, so the work is available for others to contest. We did not examine the dataset. zenodo.org
This article discusses research on group size and statistical uncertainty and is not organisational design advice. The 1992 original was not obtained, and its figures reach this article through an encyclopedia. The 2021 paper's full text was not obtained, only its abstract and the opening of its introduction. The authors' explanatory quotations come from a university news release rather than from the paper. All arithmetic is the authors' own, and one calculation rests on an invented baseline flagged in the section that uses it.