Guest Post by Milan Adams
Four hundred times a shift, a cashier hears the same flat slap of plastic landing on the scanner glass. Nobody looks up. A woman in her forties digs the card out of a purse fat with others, lays it down, and the till begins writing.
Bread, milk, frozen peas, two tins of tomatoes, a reduced chicken, forty-two pence off a chocolate bar. Then the payment, on a card ending in digits this shop has seen every second Thursday for six years. Attach an address and the basket has a household. Multiply by four hundred thousand similar baskets and the shop holds a working model of how a city eats in the week before payday.
Ninety seconds of ordinary shopping, and one woman has been converted from a customer to a profile with a value attached. Companies she will never hear of buy that profile before the peas are in the cupboard. Most of them want to sell her something. A few want to decide whether she is worth lending to, insuring, employing, or treating.
Speed is the part worth holding on to. Fifteen years ago money was spent and the state caught up afterwards, on paper, slowly, by post. Watching was an event, announced by knocking. Bookkeeping now runs the whole time, at the till, at the meter, on the gantry over the motorway, in the app, in the doorbell camera two doors down that happens to face a drive. The resulting judgment reaches the landlord’s screen before a viewing and the insurer’s pricing engine before a quote.
Six threads run through this piece, and they knot together: what a till records and who buys it; the tens of millions of adults whose credit records are too thin to be read; the point where an income model stops watching and starts accusing; the geography that fixes a life before its occupants arrive; the small surcharges on people who cannot shop around; and the shelf that will reprice itself while you stand at it.
Loyalty schemes were sold as a favour and built as a research asset, which is worth remembering when the discount gets waved in your face. Tesco introduced the Clubcard in 1995 and spent twenty years working out what the data could do before attaching the discount to the card itself through Clubcard Prices, where the shelf label carries two figures and one belongs only to the identified. Shopping anonymously now costs more for the identical loaf, so the card has stopped being a perk and become the entry fee for a fair price.
Analytics arms like Tesco’s dunnhumby, since largely sold off, made their reputation by turning baskets into households and households into segments. Kroger runs the same operation in America under different branding. Walmart began fitting electronic shelf labels at scale in 2024, reaching thousands of stores — a stationery upgrade on the surface, groundwork for the last section of this article underneath, because a price that lives on a screen needs nobody to walk the aisle with a labelling gun.
Pack sizes betray how many mouths are being fed. The own-brand ratio betrays how tight the month has gone. A slide from fresh to frozen, coffee dropping out, value lines appearing in week four and everything returning in week two, draws a pay cycle and the margin left at the end of it. Nobody has to be told the salary; the trolley says it.
Heavier plumbing sits under the shops. Payment networks see a transaction even when the loyalty card stays in the purse, and each one carries a merchant category code that sketches a life with uncomfortable clarity — payday lender, pawnbroker, betting shop, storage unit, launderette, cheque casher, the petrol station at eleven at night. Open banking, compulsory in Britain and patchy across Europe, hands over the statement itself: salary credits, direct debits, overdraft use, the exact day the balance hits zero. Pay-in-four credit occupies the newest and least policed seam of the same rock, and plenty of providers now report to credit agencies, so forty pounds of trainers split across a fortnight can dent a record a mortgage underwriter reads four years later.
Two decades of regulatory chasing have mostly ended in the regulator losing ground. A Federal Trade Commission study in 2014 found firms holding dossiers on hundreds of millions of consumers, largely unvisited by the people inside them, and recommended legislation that has since arrived in fragments. Chris Jay Hoofnagle, then at Berkeley, put the distributive problem plainly: lower-income consumers take the sharp end of data-driven discrimination, because the products pitched at them cost more and carry worse terms, and because the data gathered about them is collected for somebody else’s benefit. Rich households generate rich records too. What differs is the use — a fat record earns better offers, a thin one justifies worse.
Target’s pregnancy model, described by Charles Duhigg in The New York Times Magazine in 2012, remains the anecdote everyone repeats, and it holds up because a statistician spotted that certain basket combinations clustered around early pregnancy and coupons began arriving at a teenager’s family home before her father had been told. Aim the same technique at income and the household gets approached in its worst week: the week a second wage vanishes, the fortnight after a split, the month the car fails inspection and the commute starts to wobble.
Accusation is where the models turn dangerous. Australia’s Robodebt, from 2015, averaged annual income across fortnights, compared the average with what claimants had reported and billed them for the difference. An accounting shortcut wore an auditor’s uniform. Around 433,000 people were pursued and some A$1.76 billion in debts raised before a royal commission in 2023 found the scheme unlawful at its base; two suicides were directly linked in evidence.
Debts raised under Robodebt, cumulative, A$ million
1,760 ┤ ╭────
1,320 ┤ ╭───────────╯
880 ┤ ╭─────────────╯
440 ┤ ╭───────╯
0 ┼──────╯
2015 2016 2017 2018 2019 → collapse
Michigan’s MiDAS flagged roughly 40,000 unemployment claimants in 2013 on a software mismatch and the state refunded about twenty million dollars. Between 2013 and 2019 the Dutch childcare benefits affair caught an estimated 26,000 families, dual nationals out of all proportion, on risk flags and impossible paperwork, and brought down the Rutte government in January 2021. Three countries, three political traditions, one shape: a score built from income data, the burden of proof reversed, and an appeal too slow to reach anyone in time.
Mathematician Cathy O’Neil gave the pattern a name in Weapons of Math Destruction in 2016, and one line from it belongs screwed to the wall above every procurement desk in every ministry: “Big Data processes codify the past. They do not invent the future.” A model trained on who has historically paid, defaulted and been flagged will keep producing the silhouette of the past and calling it prophecy. That past was sorted by race and class and street long before spreadsheets existed, so the codified version arrives pre-loaded with the old unfairness and wearing a lab coat.
Twenty-six million American adults held no credit record at all in 2015, by the Consumer Financial Protection Bureau’s estimate, with another 19 million carrying files too thin to score. Pay rent in cash for fifteen years, hold a job, raise children, and the apparatus still treats you as a rumour. Experian’s Boost lets people volunteer utility and phone payments into a file, which is generous in the way a toll road is generous. Meanwhile the same trade that cannot see a tenant with no history sees a landlord’s holding company perfectly.
Invisibility gets answered with more instrumentation, never less: rent reporting, phone contracts, streaming subscriptions, gym memberships, insurance direct debits, prepaid cards dressed as accounts. Everyone acquires a record in the end. The question designers skip concerns the people whose record turns out accurate and still ruinous, since accuracy without context is only a better-aimed insult.
Brussels has begun to legislate. The EU’s AI Act, in force since 2024, classes creditworthiness assessment as high-risk and switches the obligations on in August 2026, obliging lenders to document their models, keep a human in the loop and explain the score to its subject. Britain’s Data (Use and Access) Act, with royal assent in June 2025, rewrote the consent architecture underneath much of this traffic. Whether either reaches the Leeds supermarket depends on enforcement budgets nobody has voted through.
Four miles across almost any American city, insuring the same car against the same hazards can cost hundreds of dollars more. Same driver, same vehicle, same mileage. The trade calls it actuarial honesty. Anyone who has been poor recognises a verdict reached before the evidence. Territory pricing stands in for whatever the industry would rather describe obliquely — race, density, housing age, how fast the police come — all of it crushed into a postcode and presented as a fact about a vehicle.
Home Owners’ Loan Corporation surveyors drew colour-coded maps of American cities in the 1930s and shaded the districts they judged risky in red, and that shading governed lending for forty years. The practice died. Its geography did not. Wealth settled where the maps were green, disinvestment where they were red, the actuarial tables inherited the result, and the result now gets read back to us as though the streets had done it to themselves.
Compounding works in both directions. A cheap postcode means a low valuation, a smaller loan, a slightly worse house on a slightly worse street, a worse school, a slower sale, and the whole arrangement handed to the next family. Two cars and a drive in a leafy district insure cheaply and the saving goes into a deposit. Add the small fees sprinkled along the route — higher utility deposit, loading for paying monthly instead of annually, short-term cover bought by the week because the annual lump sum sits out of reach — and the wrong street stops being a bill and becomes a levy.
Wildfire made the sentence explicit. Insurers pulled out of Californian fire country through 2023 and 2024, State Farm halting new business and later moving against roughly 72,000 existing policies, while the FAIR Plan, the state’s insurer of last resort, swelled past 450,000 policies and became load-bearing by accident. Florida’s Citizens Property Insurance arrived at the same position by the same route, growing into one of the state’s largest carriers because the private market left first. Retreat by insurers throws the exposure onto the residual pool, and the pool bills everyone still standing.
FAIR Plan policies in force, California, thousands
450 ┤ ╭─────
360 ┤ ╭───────╯
270 ┤ ╭───────╯
180 ┤ ╭───────╯
90 ┤ ╭───────╯
0 ┼──╯
2019 2021 2022 2023 2024 2026
Renewal letters arrive on Saturdays and carry the news years before the fire does.
Property portals went further into the same terrain and got burned. Zillow’s Zestimate put a number on nearly every roof in America, and for a season the company believed its own number hard enough to start buying houses against it, until the Offers division collapsed in one quarter in November 2021 with an $881 million writedown and a quarter of the staff gone. Business pages filed it as an expensive misjudgment. The interesting failure was epistemological: the model guessed well what a house was worth and badly what it would be worth in eight weeks, and the gap between those two skills is where the money sat. Millions still read a Zestimate as a valuation rather than a guess with a decimal point.
Catchment areas in Britain and attendance zones in America are priced the way everything else is priced. Estate agents print the Ofsted grade or the district rating in the same typeface as the bedroom count, and traffic on a particular crescent thickens at ten to nine. Choosing a house means choosing a peer group for a six-year-old and, without anyone saying it aloud, a set of future earnings for that six-year-old.
Emergencies are where geography bites hardest. Response times, fire station closures, distance to a trauma centre, the number of crews stacked outside a hospital waiting to hand over a patient — all distributed unevenly, all tracking deprivation with dreary loyalty. British figures have shown the slowest arrivals in the poorest and most rural districts for years, and pressure widens the gap, because the nearest available crew is a fiction once every crew is queuing. Cardiac arrest in a comfortable suburb and cardiac arrest on a hard estate are one clinical event with two different clocks. Ten minutes is survivable. Twenty-two often is not.
Broadband follows, and with it the chance to work from home, claim online, attend an appointment by video. Branches close, the post office counter becomes the last place to pay a bill in cash, then the counter shuts. Somebody reviews the bus route. A neighbourhood can be de-funded by a run of individually sensible decisions taken in offices that overlook somewhere else, and nobody in the chain needs to be hostile for the result to be.
Amazon’s HQ2 contest through 2017 and 2018 turned the whole thing into theatre — 238 bids, abatements offered up, waterfront campuses rendered for a prize of 25,000 jobs split between two winners. What the spectacle hid is that every city runs a low-grade version of the contest permanently, and the winners already had the airport and the university. Money moves toward money, one lease at a time, and the losers are told they were brave.
Cameras have become street furniture. Automatic number plate recognition is dense enough in some British counties to follow a car across a whole commute, and in America the private network Flock Safety has planted readers in thousands of communities, paid for by residents’ associations and police foundations, which puts coverage precisely where people can afford to buy it. Plate readers are sold as a crime tool and function as a traffic census. An estate that cannot fund cameras gets documented by the cameras of the estate next door.
Rent has been priced by yield-management software for years, and in some metros a handful of systems set whole districts at once, which raises the awkward question of whether competing landlords are competing or consulting one oracle. Tenants bid against an expectation trained on what the last tenant paid, and the ratchet only turns one way.
Move to a cheaper street and insurance climbs, the commute lengthens, childcare gets harder to reach, broadband drops out mid-interview. Stay and the rent rises. Nobody built the loop end to end, which is why it is so hard to argue with — no author to complain to, only reasonable decisions made by reasonable people in separate rooms.
Push the curve forward and the sums turn nasty within a paragraph.
Electronic shelf labels are the hinge. Time-of-day pricing comes first, ordinary in electricity and airline seats and creeping into groceries, where the same chicken costs more at six in the evening than at nine in the morning because the six o’clock shoppers could not shop at nine. Store-level pricing follows, with the wealthy branch and the poor branch running different price files for identical goods. Last comes personalisation keyed to the loyalty card, where the label is a display and the real figure is computed from your record — the discount depends on how much the shop thinks you can be made to pay before you walk.
Shopping unidentified already carries a surcharge, visible in the gap between card price and till price. Paying monthly instead of annually carries a loading. Short-term insurance, bought by people who cannot find a year’s premium in one movement, costs several times the annual equivalent. Overdraft fees, late fees, cash machine charges, prepaid top-up fees, interest on the rent-to-own washing machine at an effective rate that would embarrass a loan shark. Virginia Eubanks wrote in Automating Inequality that “The poor are the testing ground for new technology,” and they are the testing ground for new pricing too, absorbing the failure modes of both without being asked.
Mobility will be priced by the hour and by the metre of behaviour. Usage-based cover started as a careful-driver discount when Progressive launched Snapshot in 2008 and Allstate followed with Drivewise in 2011, and has matured into underwriting that scores braking, cornering, phone handling, time of day and, in some versions, location. The category sounds fair and distributes unfairly: a care worker doing 30,000 miles a year earns a worse rate than a consultant doing 4,000. Within the decade the pricing unit will be shorter than a year, possibly shorter than a week, with a premium that resets and a phone that serves as the meter, so a bad month raises the cost of getting to work inside that month.
Wearables will do the same to life cover. John Hancock stopped selling traditional life policies in 2018 and moved its book onto a Vitality model where activity tracking earns discounts, which works as a surcharge on the sedentary — a device counting steps, resting heart rate and sleep hands an insurer a rolling clinical impression it never had to commission. Underwriting shifts from what happened to you toward what your wrist reports, and that report is continuous, so the price can be continuous too.
Health records have already been consolidated nationally. England’s NHS signed a seven-year, £330 million contract with Palantir in November 2023 for a Federated Data Platform built to let separate trusts see the same operational picture — sensible on a ward, permanently contentious everywhere else, since the platform spans records for tens of millions of people and the terms for secondary use are still being negotiated in public. In America the Change Healthcare breach of 2024 exposed data on the order of 190 million people, showing that consolidation concentrates risk as well as records, one contractor’s bad night becoming a continent’s bad year. Equifax had taught the same lesson about credit bureaus in 2017 at 147 million records, and the lesson failed to stick.
Continuous underwriting is the likely endpoint. Mortgage approvals refreshed against live income feeds. Rents quoted off a tenant’s present cash position rather than last year’s payslips. Premiums adjusting after a bad fortnight. Little of it needs new bureaucracy: the plumbing exists, the consent hides in terms nobody reads, and the commercial incentive runs one way. Better risk pricing gets passed along as segmentation, accurate segmentation keeps good risks cheap and prices bad ones out of cover, and the efficient outcome is a population split into insured and uninsured halves.
Queues for treatment are rationed by clinical priority and, increasingly, by whether an appointment stands a chance of being attended — and attendance depends on transport, childcare and shift patterns, which depend on income. Triaging by likelihood of turning up favours the comfortable without meaning to. Remote-first consulting hardens it further, since a video call needs a room, a connection and the privacy to describe a symptom without a landlord walking through the flat. At the front of the queue the practical question becomes who is easiest to process, and the answer matches the postcode at the top of this article.
Climate accelerates the sorting. Insurance retreat from fire and flood country will spread past California, and where cover vanishes the lenders pull back next, because nobody lends against an uninsurable asset for long. Lending gone, the market for the house dies. Districts enter managed decline through a stack of renewals and refusals rather than a single decision. Households with equity leave early and take it. Households without stay in houses nobody will advance money against.
Counter-currents exist and deserve their space. Pricing behaviour rather than the demographic someone was born into helps the young, the self-employed and the migrant with a short history. Open banking lets thin-file borrowers evidence income payslips never captured, and rent reporting drags tenants toward visibility. Aadhaar in India and M-Pesa in Kenya pulled hundreds of millions of adults into formal finance within a generation, with real gains. Estonia’s digital state does in an afternoon what other countries take a fortnight and three stamped forms.
Every one of those is also a register, and a register that includes you retains the power to exclude. The identity rails that let a farmer in Maharashtra draw a payment also let a clerk in a district office rule him ineligible. Shoshana Zuboff’s phrase in The Age of Surveillance Capitalism — human experience claimed as free raw material for translation into behavioural data — describes the direction of travel well enough, though the commodity refined here is solvency rather than attention. We send the ore of our own household budgets out to systems we cannot inspect and get a price back.
Prediction is where it turns strange rather than merely dear. Models already forecast default, divorce, job change, pregnancy, illness and moving house from transaction patterns, so a lender that expects you to move in eleven months prices differently from one that thinks you are settled, and a retailer that knows about a separation sends single portions and cheaper wine. By 2030 the system’s opinion of your next three years will be formed and sold before you have formed it, and it will stick, because a score that follows you is a reputation you never wrote and cannot appeal to a person.
The raw material was never meant as evidence. Gym check-ins. Subscription churn. Phone unlocks at three in the morning. Steps. Whether the heating came on. Whether the car moved. Each is a weak signal; the composite is not weak, because the composite gets read against millions of comparable households and the deviation is what draws notice. A household quietly coming apart leaves a signature — cheaper cuts, more frozen and fewer fresh, more cash withdrawals, cancelled subscriptions, a cluster of pharmacy visits — and it reaches the market weeks before anyone in the house says the word struggling out loud. Being seen early is a mercy. Somebody acting on it rather than selling against it would be a bigger one.
Employers, landlords and, in places, benefit offices read the same telemetry. Warehouse productivity runs on scan rates and idle time, remote work normalised keystroke counters and screenshot agents, delivery routing scores the stop rather than the shift. Eric Schmidt’s boast from the middle of the last decade, that the platforms know where you are and where you have been, was a remark about a phone in a pocket and has since become a fair description of a payroll.
Come 2030 the picture looks undramatic. Nothing explodes, the traffic moves, the shops open. What has changed is that the price of everything necessary depends on what the machinery has concluded about you, drawn from loyalty records, meter readings, telematics, statement feeds and the phone in your hand. Two neighbours in identical houses doing identical jobs get charged differently for the same basket, the same cover, the same credit, and neither learns the other’s number.
Asymmetry matters more than the cameras. A system that prices you well knows more about your circumstances than you do, and will use the advantage by default, because nobody has written a rule obliging it to do otherwise. The statements are in the drawer. Somewhere in a building you will never visit, a model has an opinion about your next three years and has held it a while.
Friction at the right joints is the counterweight: scores that travel with the person rather than the address, a right to read the record, an appeal that reaches a human with power to reverse the machine, prices visible at the shelf, and a hard limit on what may be inferred without being declared. Britain’s slide toward a cashless economy and the death of the free cash machine make a fair test of whether any of that gets built, since a country without notes prices the unbanked and the unscored out of the room. Sweden went furthest fastest and is now legislating to slow itself.
Cash payments as a share of all payments, UK, %
40 ┤╮
30 ┤ ╰╮
20 ┤ ╰╮
10 ┤ ╰──╮
5 ┤ ╰──╮
0 ┼───────────╰──────────
2014 2017 2019 2022 2025
Whether anyone tells you what the watching concluded, and whether a person holding three children, one job and a spare hour on a Thursday can argue with it, will matter more by then than whether you were watched at all.
Most of them will spend the hour on something else.
Tweet
Continue reading...
[ H/T The Burning Platform ]
Four hundred times a shift, a cashier hears the same flat slap of plastic landing on the scanner glass. Nobody looks up. A woman in her forties digs the card out of a purse fat with others, lays it down, and the till begins writing.
Bread, milk, frozen peas, two tins of tomatoes, a reduced chicken, forty-two pence off a chocolate bar. Then the payment, on a card ending in digits this shop has seen every second Thursday for six years. Attach an address and the basket has a household. Multiply by four hundred thousand similar baskets and the shop holds a working model of how a city eats in the week before payday.
Ninety seconds of ordinary shopping, and one woman has been converted from a customer to a profile with a value attached. Companies she will never hear of buy that profile before the peas are in the cupboard. Most of them want to sell her something. A few want to decide whether she is worth lending to, insuring, employing, or treating.
Speed is the part worth holding on to. Fifteen years ago money was spent and the state caught up afterwards, on paper, slowly, by post. Watching was an event, announced by knocking. Bookkeeping now runs the whole time, at the till, at the meter, on the gantry over the motorway, in the app, in the doorbell camera two doors down that happens to face a drive. The resulting judgment reaches the landlord’s screen before a viewing and the insurer’s pricing engine before a quote.
Six threads run through this piece, and they knot together: what a till records and who buys it; the tens of millions of adults whose credit records are too thin to be read; the point where an income model stops watching and starts accusing; the geography that fixes a life before its occupants arrive; the small surcharges on people who cannot shop around; and the shelf that will reprice itself while you stand at it.
What the till records, and what the brokers pay for it
Loyalty schemes were sold as a favour and built as a research asset, which is worth remembering when the discount gets waved in your face. Tesco introduced the Clubcard in 1995 and spent twenty years working out what the data could do before attaching the discount to the card itself through Clubcard Prices, where the shelf label carries two figures and one belongs only to the identified. Shopping anonymously now costs more for the identical loaf, so the card has stopped being a perk and become the entry fee for a fair price.
Analytics arms like Tesco’s dunnhumby, since largely sold off, made their reputation by turning baskets into households and households into segments. Kroger runs the same operation in America under different branding. Walmart began fitting electronic shelf labels at scale in 2024, reaching thousands of stores — a stationery upgrade on the surface, groundwork for the last section of this article underneath, because a price that lives on a screen needs nobody to walk the aisle with a labelling gun.
Pack sizes betray how many mouths are being fed. The own-brand ratio betrays how tight the month has gone. A slide from fresh to frozen, coffee dropping out, value lines appearing in week four and everything returning in week two, draws a pay cycle and the margin left at the end of it. Nobody has to be told the salary; the trolley says it.
Heavier plumbing sits under the shops. Payment networks see a transaction even when the loyalty card stays in the purse, and each one carries a merchant category code that sketches a life with uncomfortable clarity — payday lender, pawnbroker, betting shop, storage unit, launderette, cheque casher, the petrol station at eleven at night. Open banking, compulsory in Britain and patchy across Europe, hands over the statement itself: salary credits, direct debits, overdraft use, the exact day the balance hits zero. Pay-in-four credit occupies the newest and least policed seam of the same rock, and plenty of providers now report to credit agencies, so forty pounds of trainers split across a fortnight can dent a record a mortgage underwriter reads four years later.
Two decades of regulatory chasing have mostly ended in the regulator losing ground. A Federal Trade Commission study in 2014 found firms holding dossiers on hundreds of millions of consumers, largely unvisited by the people inside them, and recommended legislation that has since arrived in fragments. Chris Jay Hoofnagle, then at Berkeley, put the distributive problem plainly: lower-income consumers take the sharp end of data-driven discrimination, because the products pitched at them cost more and carry worse terms, and because the data gathered about them is collected for somebody else’s benefit. Rich households generate rich records too. What differs is the use — a fat record earns better offers, a thin one justifies worse.
Target’s pregnancy model, described by Charles Duhigg in The New York Times Magazine in 2012, remains the anecdote everyone repeats, and it holds up because a statistician spotted that certain basket combinations clustered around early pregnancy and coupons began arriving at a teenager’s family home before her father had been told. Aim the same technique at income and the household gets approached in its worst week: the week a second wage vanishes, the fortnight after a split, the month the car fails inspection and the commute starts to wobble.
Accusation is where the models turn dangerous. Australia’s Robodebt, from 2015, averaged annual income across fortnights, compared the average with what claimants had reported and billed them for the difference. An accounting shortcut wore an auditor’s uniform. Around 433,000 people were pursued and some A$1.76 billion in debts raised before a royal commission in 2023 found the scheme unlawful at its base; two suicides were directly linked in evidence.
Debts raised under Robodebt, cumulative, A$ million
1,760 ┤ ╭────
1,320 ┤ ╭───────────╯
880 ┤ ╭─────────────╯
440 ┤ ╭───────╯
0 ┼──────╯
2015 2016 2017 2018 2019 → collapse
Michigan’s MiDAS flagged roughly 40,000 unemployment claimants in 2013 on a software mismatch and the state refunded about twenty million dollars. Between 2013 and 2019 the Dutch childcare benefits affair caught an estimated 26,000 families, dual nationals out of all proportion, on risk flags and impossible paperwork, and brought down the Rutte government in January 2021. Three countries, three political traditions, one shape: a score built from income data, the burden of proof reversed, and an appeal too slow to reach anyone in time.
Mathematician Cathy O’Neil gave the pattern a name in Weapons of Math Destruction in 2016, and one line from it belongs screwed to the wall above every procurement desk in every ministry: “Big Data processes codify the past. They do not invent the future.” A model trained on who has historically paid, defaulted and been flagged will keep producing the silhouette of the past and calling it prophecy. That past was sorted by race and class and street long before spreadsheets existed, so the codified version arrives pre-loaded with the old unfairness and wearing a lab coat.
Twenty-six million American adults held no credit record at all in 2015, by the Consumer Financial Protection Bureau’s estimate, with another 19 million carrying files too thin to score. Pay rent in cash for fifteen years, hold a job, raise children, and the apparatus still treats you as a rumour. Experian’s Boost lets people volunteer utility and phone payments into a file, which is generous in the way a toll road is generous. Meanwhile the same trade that cannot see a tenant with no history sees a landlord’s holding company perfectly.
Invisibility gets answered with more instrumentation, never less: rent reporting, phone contracts, streaming subscriptions, gym memberships, insurance direct debits, prepaid cards dressed as accounts. Everyone acquires a record in the end. The question designers skip concerns the people whose record turns out accurate and still ruinous, since accuracy without context is only a better-aimed insult.
Brussels has begun to legislate. The EU’s AI Act, in force since 2024, classes creditworthiness assessment as high-risk and switches the obligations on in August 2026, obliging lenders to document their models, keep a human in the loop and explain the score to its subject. Britain’s Data (Use and Access) Act, with royal assent in June 2025, rewrote the consent architecture underneath much of this traffic. Whether either reaches the Leeds supermarket depends on enforcement budgets nobody has voted through.
A street gets sentenced before anyone living on it does
Four miles across almost any American city, insuring the same car against the same hazards can cost hundreds of dollars more. Same driver, same vehicle, same mileage. The trade calls it actuarial honesty. Anyone who has been poor recognises a verdict reached before the evidence. Territory pricing stands in for whatever the industry would rather describe obliquely — race, density, housing age, how fast the police come — all of it crushed into a postcode and presented as a fact about a vehicle.
Home Owners’ Loan Corporation surveyors drew colour-coded maps of American cities in the 1930s and shaded the districts they judged risky in red, and that shading governed lending for forty years. The practice died. Its geography did not. Wealth settled where the maps were green, disinvestment where they were red, the actuarial tables inherited the result, and the result now gets read back to us as though the streets had done it to themselves.
Compounding works in both directions. A cheap postcode means a low valuation, a smaller loan, a slightly worse house on a slightly worse street, a worse school, a slower sale, and the whole arrangement handed to the next family. Two cars and a drive in a leafy district insure cheaply and the saving goes into a deposit. Add the small fees sprinkled along the route — higher utility deposit, loading for paying monthly instead of annually, short-term cover bought by the week because the annual lump sum sits out of reach — and the wrong street stops being a bill and becomes a levy.
Wildfire made the sentence explicit. Insurers pulled out of Californian fire country through 2023 and 2024, State Farm halting new business and later moving against roughly 72,000 existing policies, while the FAIR Plan, the state’s insurer of last resort, swelled past 450,000 policies and became load-bearing by accident. Florida’s Citizens Property Insurance arrived at the same position by the same route, growing into one of the state’s largest carriers because the private market left first. Retreat by insurers throws the exposure onto the residual pool, and the pool bills everyone still standing.
FAIR Plan policies in force, California, thousands
450 ┤ ╭─────
360 ┤ ╭───────╯
270 ┤ ╭───────╯
180 ┤ ╭───────╯
90 ┤ ╭───────╯
0 ┼──╯
2019 2021 2022 2023 2024 2026
Renewal letters arrive on Saturdays and carry the news years before the fire does.
Property portals went further into the same terrain and got burned. Zillow’s Zestimate put a number on nearly every roof in America, and for a season the company believed its own number hard enough to start buying houses against it, until the Offers division collapsed in one quarter in November 2021 with an $881 million writedown and a quarter of the staff gone. Business pages filed it as an expensive misjudgment. The interesting failure was epistemological: the model guessed well what a house was worth and badly what it would be worth in eight weeks, and the gap between those two skills is where the money sat. Millions still read a Zestimate as a valuation rather than a guess with a decimal point.
Catchment areas in Britain and attendance zones in America are priced the way everything else is priced. Estate agents print the Ofsted grade or the district rating in the same typeface as the bedroom count, and traffic on a particular crescent thickens at ten to nine. Choosing a house means choosing a peer group for a six-year-old and, without anyone saying it aloud, a set of future earnings for that six-year-old.
Emergencies are where geography bites hardest. Response times, fire station closures, distance to a trauma centre, the number of crews stacked outside a hospital waiting to hand over a patient — all distributed unevenly, all tracking deprivation with dreary loyalty. British figures have shown the slowest arrivals in the poorest and most rural districts for years, and pressure widens the gap, because the nearest available crew is a fiction once every crew is queuing. Cardiac arrest in a comfortable suburb and cardiac arrest on a hard estate are one clinical event with two different clocks. Ten minutes is survivable. Twenty-two often is not.
Broadband follows, and with it the chance to work from home, claim online, attend an appointment by video. Branches close, the post office counter becomes the last place to pay a bill in cash, then the counter shuts. Somebody reviews the bus route. A neighbourhood can be de-funded by a run of individually sensible decisions taken in offices that overlook somewhere else, and nobody in the chain needs to be hostile for the result to be.
Amazon’s HQ2 contest through 2017 and 2018 turned the whole thing into theatre — 238 bids, abatements offered up, waterfront campuses rendered for a prize of 25,000 jobs split between two winners. What the spectacle hid is that every city runs a low-grade version of the contest permanently, and the winners already had the airport and the university. Money moves toward money, one lease at a time, and the losers are told they were brave.
Cameras have become street furniture. Automatic number plate recognition is dense enough in some British counties to follow a car across a whole commute, and in America the private network Flock Safety has planted readers in thousands of communities, paid for by residents’ associations and police foundations, which puts coverage precisely where people can afford to buy it. Plate readers are sold as a crime tool and function as a traffic census. An estate that cannot fund cameras gets documented by the cameras of the estate next door.
Rent has been priced by yield-management software for years, and in some metros a handful of systems set whole districts at once, which raises the awkward question of whether competing landlords are competing or consulting one oracle. Tenants bid against an expectation trained on what the last tenant paid, and the ratchet only turns one way.
Move to a cheaper street and insurance climbs, the commute lengthens, childcare gets harder to reach, broadband drops out mid-interview. Stay and the rent rises. Nobody built the loop end to end, which is why it is so hard to argue with — no author to complain to, only reasonable decisions made by reasonable people in separate rooms.
The shelf that reprices itself, the wrist that underwrites the policy
Push the curve forward and the sums turn nasty within a paragraph.
Electronic shelf labels are the hinge. Time-of-day pricing comes first, ordinary in electricity and airline seats and creeping into groceries, where the same chicken costs more at six in the evening than at nine in the morning because the six o’clock shoppers could not shop at nine. Store-level pricing follows, with the wealthy branch and the poor branch running different price files for identical goods. Last comes personalisation keyed to the loyalty card, where the label is a display and the real figure is computed from your record — the discount depends on how much the shop thinks you can be made to pay before you walk.
Shopping unidentified already carries a surcharge, visible in the gap between card price and till price. Paying monthly instead of annually carries a loading. Short-term insurance, bought by people who cannot find a year’s premium in one movement, costs several times the annual equivalent. Overdraft fees, late fees, cash machine charges, prepaid top-up fees, interest on the rent-to-own washing machine at an effective rate that would embarrass a loan shark. Virginia Eubanks wrote in Automating Inequality that “The poor are the testing ground for new technology,” and they are the testing ground for new pricing too, absorbing the failure modes of both without being asked.
Mobility will be priced by the hour and by the metre of behaviour. Usage-based cover started as a careful-driver discount when Progressive launched Snapshot in 2008 and Allstate followed with Drivewise in 2011, and has matured into underwriting that scores braking, cornering, phone handling, time of day and, in some versions, location. The category sounds fair and distributes unfairly: a care worker doing 30,000 miles a year earns a worse rate than a consultant doing 4,000. Within the decade the pricing unit will be shorter than a year, possibly shorter than a week, with a premium that resets and a phone that serves as the meter, so a bad month raises the cost of getting to work inside that month.
Wearables will do the same to life cover. John Hancock stopped selling traditional life policies in 2018 and moved its book onto a Vitality model where activity tracking earns discounts, which works as a surcharge on the sedentary — a device counting steps, resting heart rate and sleep hands an insurer a rolling clinical impression it never had to commission. Underwriting shifts from what happened to you toward what your wrist reports, and that report is continuous, so the price can be continuous too.
Health records have already been consolidated nationally. England’s NHS signed a seven-year, £330 million contract with Palantir in November 2023 for a Federated Data Platform built to let separate trusts see the same operational picture — sensible on a ward, permanently contentious everywhere else, since the platform spans records for tens of millions of people and the terms for secondary use are still being negotiated in public. In America the Change Healthcare breach of 2024 exposed data on the order of 190 million people, showing that consolidation concentrates risk as well as records, one contractor’s bad night becoming a continent’s bad year. Equifax had taught the same lesson about credit bureaus in 2017 at 147 million records, and the lesson failed to stick.
Continuous underwriting is the likely endpoint. Mortgage approvals refreshed against live income feeds. Rents quoted off a tenant’s present cash position rather than last year’s payslips. Premiums adjusting after a bad fortnight. Little of it needs new bureaucracy: the plumbing exists, the consent hides in terms nobody reads, and the commercial incentive runs one way. Better risk pricing gets passed along as segmentation, accurate segmentation keeps good risks cheap and prices bad ones out of cover, and the efficient outcome is a population split into insured and uninsured halves.
Queues for treatment are rationed by clinical priority and, increasingly, by whether an appointment stands a chance of being attended — and attendance depends on transport, childcare and shift patterns, which depend on income. Triaging by likelihood of turning up favours the comfortable without meaning to. Remote-first consulting hardens it further, since a video call needs a room, a connection and the privacy to describe a symptom without a landlord walking through the flat. At the front of the queue the practical question becomes who is easiest to process, and the answer matches the postcode at the top of this article.
Climate accelerates the sorting. Insurance retreat from fire and flood country will spread past California, and where cover vanishes the lenders pull back next, because nobody lends against an uninsurable asset for long. Lending gone, the market for the house dies. Districts enter managed decline through a stack of renewals and refusals rather than a single decision. Households with equity leave early and take it. Households without stay in houses nobody will advance money against.
Counter-currents exist and deserve their space. Pricing behaviour rather than the demographic someone was born into helps the young, the self-employed and the migrant with a short history. Open banking lets thin-file borrowers evidence income payslips never captured, and rent reporting drags tenants toward visibility. Aadhaar in India and M-Pesa in Kenya pulled hundreds of millions of adults into formal finance within a generation, with real gains. Estonia’s digital state does in an afternoon what other countries take a fortnight and three stamped forms.
Every one of those is also a register, and a register that includes you retains the power to exclude. The identity rails that let a farmer in Maharashtra draw a payment also let a clerk in a district office rule him ineligible. Shoshana Zuboff’s phrase in The Age of Surveillance Capitalism — human experience claimed as free raw material for translation into behavioural data — describes the direction of travel well enough, though the commodity refined here is solvency rather than attention. We send the ore of our own household budgets out to systems we cannot inspect and get a price back.
Prediction is where it turns strange rather than merely dear. Models already forecast default, divorce, job change, pregnancy, illness and moving house from transaction patterns, so a lender that expects you to move in eleven months prices differently from one that thinks you are settled, and a retailer that knows about a separation sends single portions and cheaper wine. By 2030 the system’s opinion of your next three years will be formed and sold before you have formed it, and it will stick, because a score that follows you is a reputation you never wrote and cannot appeal to a person.
The raw material was never meant as evidence. Gym check-ins. Subscription churn. Phone unlocks at three in the morning. Steps. Whether the heating came on. Whether the car moved. Each is a weak signal; the composite is not weak, because the composite gets read against millions of comparable households and the deviation is what draws notice. A household quietly coming apart leaves a signature — cheaper cuts, more frozen and fewer fresh, more cash withdrawals, cancelled subscriptions, a cluster of pharmacy visits — and it reaches the market weeks before anyone in the house says the word struggling out loud. Being seen early is a mercy. Somebody acting on it rather than selling against it would be a bigger one.
Employers, landlords and, in places, benefit offices read the same telemetry. Warehouse productivity runs on scan rates and idle time, remote work normalised keystroke counters and screenshot agents, delivery routing scores the stop rather than the shift. Eric Schmidt’s boast from the middle of the last decade, that the platforms know where you are and where you have been, was a remark about a phone in a pocket and has since become a fair description of a payroll.
Come 2030 the picture looks undramatic. Nothing explodes, the traffic moves, the shops open. What has changed is that the price of everything necessary depends on what the machinery has concluded about you, drawn from loyalty records, meter readings, telematics, statement feeds and the phone in your hand. Two neighbours in identical houses doing identical jobs get charged differently for the same basket, the same cover, the same credit, and neither learns the other’s number.
Asymmetry matters more than the cameras. A system that prices you well knows more about your circumstances than you do, and will use the advantage by default, because nobody has written a rule obliging it to do otherwise. The statements are in the drawer. Somewhere in a building you will never visit, a model has an opinion about your next three years and has held it a while.
Friction at the right joints is the counterweight: scores that travel with the person rather than the address, a right to read the record, an appeal that reaches a human with power to reverse the machine, prices visible at the shelf, and a hard limit on what may be inferred without being declared. Britain’s slide toward a cashless economy and the death of the free cash machine make a fair test of whether any of that gets built, since a country without notes prices the unbanked and the unscored out of the room. Sweden went furthest fastest and is now legislating to slow itself.
Cash payments as a share of all payments, UK, %
40 ┤╮
30 ┤ ╰╮
20 ┤ ╰╮
10 ┤ ╰──╮
5 ┤ ╰──╮
0 ┼───────────╰──────────
2014 2017 2019 2022 2025
Whether anyone tells you what the watching concluded, and whether a person holding three children, one job and a spare hour on a Thursday can argue with it, will matter more by then than whether you were watched at all.
Most of them will spend the hour on something else.
Tweet
Continue reading...
[ H/T The Burning Platform ]
