DEEP READING · READ 940
Data Bias in a World Designed for Men · Caroline Criado Perez · 2019
This book makes one argument, and that argument explains an enormous amount of daily friction you had filed under "that's just how it is": from drug dosing to crash test dummies, from office thermostats to the order in which snow gets cleared, most of the modern world's defaults were built around a 70-kilogram male body and a male life that goes to work in a straight line and comes home to a dinner someone else cooked — with women treated as a variant of that default. And the crucial point is this: almost none of it is malice. It follows from an assumption nobody ever said out loud — that "person" means male. Because it is never written down, it lives in the blanks in the data, gets cast into products, institutions and algorithms, and reproduces itself for free.
Caroline Criado Perez is a British writer and campaigner, previously known for the campaign that put a woman — Jane Austen — on the Bank of England's ten-pound note. This 2019 book is not theory; it is an evidence dossier: she pulls research out of transport planning, medicine, labour economics, industrial design, disaster relief and tax policy, and threads it onto a single argument. It won both the Financial Times Business Book of the Year and the Royal Society Science Book Prize that year — a feminist book read simultaneously as a business book and a science book, which tells you its method: it appeals not to outrage but to a gap in the data.
Start with the single idea that carries the whole book: Reference Man. In radiation protection, occupational safety and ergonomics there has long been a standardized model of the human body, with height and weight drawn from Western adult men (classically around 70 kg, in his twenties or thirties, Caucasian). Originally it was an engineering convenience: you need a size for "a person" before you can calculate a dose, a load, a safety margin. The trouble is that once the convenience hardens, it stops being one reference point and becomes the human being.
The analogy: it is as if every doorway in a city were cut to one particular gentleman's height, and then, when people kept hitting their heads, we concluded that those people were the wrong height. The body chosen as the default never feels singled out — which is exactly why it is so hard to see. And everyone who departs from it pays the same invisible tax: not only women, but short men, older people, different body types and ethnicities.
How it changes the way you look at things: from now on, any "standard," "default setting" or "universal fit" prompts one extra question — whose body was this measured from? That question almost always turns something up, and far beyond gender.
This is the hinge of the book. The gender data gap means that huge quantities of data are collected without being disaggregated by sex, or from predominantly male samples, so information about women's experience does not exist in the first place.
What makes it insidious is that absence is soundless. You can rebut a bias; you cannot even point at a blank, because there is nothing on the screen. A decision-maker looks at the dashboard, sees no red light, and concludes that all is well. This is the epistemological lesson worth carrying out of the book: silence in data may mean there is no problem, or it may mean nobody took a measurement — and on the report those two look identical.
The classic illustration is snow clearing in Karlskoga, a small Swedish town. Officials running every policy through a gender check joked that snow clearing, at least, could hardly be gendered. Then they looked: the clearing order put main roads first and pavements and cycle paths last — and drivers commuting by car skewed male, while those walking, pushing prams and escorting children skewed female. Then they looked at hospital data: in icy conditions, far more pedestrians were injured than drivers, most of them women, and the medical and lost-work costs of those falls exceeded the cost of the road maintenance itself. So they reversed the order and cleared pavements first. The case gets cited endlessly not because snow matters but because it demonstrates the whole mechanism: nobody discriminated against anyone; it simply never occurred to anyone to count pedestrians separately from drivers — and once you did, even the cost accounting turned out to have been backwards.
The second large gap is in labour. Globally, the overwhelming majority of unpaid care work — childcare, eldercare, housework — is done by women; the International Labour Organization puts it at roughly three quarters. GDP counts only market transactions: pay someone to mind your child and it enters GDP; mind the child yourself and it does not. So an activity that absorbs a vast share of human working hours and largely determines the next generation's human capital is worth exactly zero in the national accounts. The consequence is not abstract: because it does not count, cutting public childcare or eldercare always looks like a saving on paper — when in fact the work has only moved from the budget spreadsheet to somebody's personal calendar. It did not disappear.
That unpaid work also rewrites how people move. Planners model the commute: a peak-hour line between home and office. But people carrying care responsibilities travel by trip-chaining: drop a child at school, go to work, run an errand for an elderly parent at lunch, buy groceries on the way back, collect the child. That is multi-leg, short-distance, off-peak travel with frequent transfers — and almost every fare structure, timetable and route map penalizes it. Per-journey pricing makes chained trips more expensive; radial suburb-to-centre lines make lateral movement absurd; off-peak service is thin. It is not that she dislikes public transport; it is that nobody ever wrote a requirement for that life.
The most alarming chapter, and the one with the clearest policy consequences. Clinical trials were historically dominated by male subjects; even animal and cell studies preferred males, on the grounds that the female hormonal cycle introduced "noise" — a real physiological variable reclassified as an inconvenience to measurement. The US National Institutes of Health has required the inclusion of women in funded clinical research since 1993, and since 2016 has required researchers to account for sex as a biological variable in preclinical work. Those dates are the point: this is not ancient history.
The consequences land in three places. One: dosing. Men and women genuinely differ in metabolic rate, body-fat ratio and enzyme activity, yet doses are often extrapolated from male average weight. The signature case is the sleep drug zolpidem (Ambien): after evidence that women metabolize it more slowly and may still be impaired the next morning, the US Food and Drug Administration halved the recommended dose for women in 2013 — two decades after the drug went to market. Two: symptom recognition. The "classic" heart attack presentation — crushing chest pain radiating down the left arm — was derived from men. Women more often present with abdominal pain, breathlessness, nausea and extreme fatigue, which then get labelled "atypical." Sit with how damaging that word is: atypical relative to whom? A University of Leeds study found women were substantially more likely than men to be initially misdiagnosed after a heart attack (the figure it reported was around fifty per cent higher). Three: pain and self-report. A large body of research suggests women reporting pain are more likely to have it attributed to emotional or psychological causes, wait longer for analgesia, and face longer diagnostic delays — the average time to a diagnosis of endometriosis is measured in years.
How it changes your thinking: apparently neutral clinical vocabulary — "typical," "atypical," "normal range," "outlier" — always has a reference body built into it. And in medicine, a misaligned reference frame is paid for in lives.
The physical world makes the default most visible. Crash test dummies were long built on the 50th-percentile male (about 1.77 m, 76 kg); the so-called female dummy has frequently been a scaled-down male that does not model real differences in fat distribution, muscle mass, neck strength or seating position (shorter people sit further forward and more upright — a posture the standard tests classify as "out of position"). The US only introduced a smaller female dummy into some tests and seating positions around 2011. The book cites research finding women significantly more likely to be seriously injured in comparable crashes — that specific number has since been contested (see the critiques below), but the underlying fact that the test baseline is a male body is not in dispute.
The same logic is everywhere: personal protective equipment (PPE) — body armour, stab vests, safety boots and coveralls patterned on male bodies, so women in policing and firefighting wear ill-fitting gear that both restricts movement and protects worse; tools sized to male hands; the standard piano keyboard width, which raises injury risk for players with smaller hands (narrower keyboards exist as an alternative); mainstream smartphone sizes that exceed comfortable one-handed use for most women; office temperature standards traceable to an old model built on adult male metabolic rates.
The value of this section is not the inventory but the change of eyes: next time something is awkward to use, don't start by assuming you are clumsy — ask whose body it was made for.
The most forward-looking part of the book. Machine learning extracts patterns from historical data, and historical data is full of every gap above — so an algorithm is not neutral; it is the automation of the past's average.
Two concrete mechanisms. First, word embeddings — the technique of turning words into numerical vectors so machines can compute relations between them. A well-known 2016 study found that embeddings trained on news text supported the analogy "man is to computer programmer as woman is to homemaker." Nobody taught the model that; it faithfully learned a statistical regularity present in the corpus. Second, sample bias in the training data itself: early speech recognition was more accurate on male than female voices (one study found such a gap using a video platform's automatic captions), because the training audio skewed male; a résumé-screening model trained on historical hiring records treats "what people who were hired look like" as "what a qualified person looks like."
Why this matters more than the chapters before it: bias in the physical world is slow to fix but visible, while bias in a model is fast to spread and invisible — and it is trusted precisely because it arrives labelled objective, unemotional, data-driven. Once the gap is inside the model, it stops being one engineer's oversight and becomes the system's judgment.
The final layer is institutional. A policy that calls itself gender-neutral has usually just not checked where it lands. Cutting public spending looks even-handed on paper — but when childcare, eldercare and community services are cut, that work does not vanish, it flows back into unpaid household labour, which is overwhelmingly done by women. Tax treatment of the second earner in a household can push effective marginal rates high enough that the lower earner — usually the woman — finds paid work barely worth it. Disaster relief and humanitarian work supply a long list of lessons: temporary toilets and shelters without locks, relief kits with no menstrual supplies, cash assistance paid to the "head of household" (male by default) — none of it malicious, all of it designed by people whose mental image of a disaster victim had no particular body.
The remedy is gender budgeting: a requirement that budgets and policies be assessed against sex-disaggregated data before they are adopted. It sounds like bureaucratic box-ticking, but the Karlskoga snow decision is exactly what such a process produces. That is the book's most practical conclusion: what defeats the data gap is not stronger moral feeling but dumber procedure — mandatory disaggregation, so that somewhere a form has to be filled in for the place where silence would otherwise reign.
Six parts (daily life, the workplace, design, medicine, public life, disaster), but one chain of argument, worth memorizing in one pass:
| 1. Default | "Human" defaults to a male body and a male rhythm of life (Reference Man). |
| 2. Gap | Data is not disaggregated by sex, samples skew male — women's experience is absent from the record. |
| 3. Misreading | Decision-makers read "no data" as "no problem." |
| 4. Casting | The default gets set into doses, cars, protective gear, routes, tax codes, models. |
| 5. Cost | Higher injury rates, later diagnoses, costlier travel, lower income. |
| 6. Loop | The cost is observed as "they're not suited to it / they don't choose it," which re-hardens step 1. |
Hundreds of examples, all running the same loop — and one intervention point: force disaggregation at step 2 and the circuit breaks.
What the book establishes, then, is not the familiar thesis that society discriminates against women. It is something more precise and much harder to argue with: with no one harbouring ill will, a system can still allocate cost reliably to the same group — as long as its defaults are taken from one part of the population and it never measures the rest. That conclusion transfers intact to age, body size, ethnicity, disability and geography. What the book really teaches is a skill — auditing defaults — and gender is simply its most fully documented case.
1. Most of the modern world's defaults — drug doses, crash dummies, protective gear patterns, thermostat settings, bus routes — were taken from a 70-kilogram male body and a life that travels in a straight line.
2. And almost none of it is malice; it follows from an assumption never said aloud: that "person" means male. No perpetrator does not mean no cost.
3. The hinge of the book is the gender data gap: data not disaggregated by sex, samples skewed male, and so half the population's experience simply is not in the record.
4. The deadliest epistemological trap is that "no data" gets read at the decision table as "no problem" — and on a dashboard those two silences look identical.
5. A Swedish town reversed its snow-clearing order to do pavements before roads, purely because icy weather injured far more pedestrians than drivers at a higher medical cost — nobody discriminated; nobody had simply ever counted pedestrians separately.
6. Unpaid housework and care count as zero in GDP, so cutting childcare and eldercare always looks like a saving — the work only moves from a budget line to somebody's calendar. It does not disappear.
7. Planners design for the commute; carers live by trip-chaining — multi-leg, short, off-peak, transfer-heavy. It is not that she dislikes the bus; the bus never had a requirement written for that life.
8. The "classic" heart attack symptoms were derived from men, so women's presentations are called "atypical" — always ask the follow-up question: typical relative to whom?
9. An algorithm is the automation of the past's average: it learns yesterday's defaults verbatim and reissues them as objective, data-driven output — bias in the physical world is slow to fix but visible; bias in a model is fast and invisible.
10. So the antidote is not stronger moral feeling but dumber procedure — mandatory sex-disaggregated data, so a form must be filled in where silence would otherwise rule — and the transferable skill the book actually teaches is this: whenever you meet a "standard," a "default" or a "universal fit," ask whose body it was measured from.