Vol. II, No. 29
Covering 15 January - 28 January 2024
Monday, January 29, 2024
Late edition · A Relentless publication
All the fortnight that mattered, in technology and in the world, read next to what we were building at the time.
BRITAIN

The computer was wrong for twenty years, and it took a television drama to make anyone believe the people it destroyed

More than 900 sub-postmasters were prosecuted for stealing money that was never missing. The thief was a software bug. It is Britain's largest miscarriage of justice, and it happened because an institution trusted a system over its own people.

Britain spent this fortnight reckoning with a scandal it had known about for years and only now, shamed by a television drama, decided to treat as the catastrophe it is. Between 1999 and 2015, the Post Office prosecuted more than 900 of its own sub-postmasters, the people who run local post office branches, for theft and false accounting, on the evidence of a computer system called Horizon, built by the Japanese firm Fujitsu, that showed money missing from their branch accounts. The money was not missing. The system was riddled with bugs that invented shortfalls out of nothing. But the Post Office believed the computer, prosecuted the humans, and ruined them: people were imprisoned, bankrupted, hounded from their communities as thieves, and at least four are believed to have taken their own lives. Many died before their names were cleared.

What forced the reckoning was not the courts, which had begun to overturn convictions, nor the public inquiry grinding on in the background. It was a four-part ITV drama, broadcast at the start of January, that told the story through the sub-postmasters themselves and made a nation that had shrugged at two decades of news reports suddenly, furiously care. Within days the government announced it would legislate to exonerate the wrongly convicted en masse, an extraordinary step, and to compensate them. The campaigner who had fought for twenty years to be believed watched the country discover, in a week of prime-time television, what he had been shouting into an indifferent void since before some of the guilty verdicts were handed down.

The horror of it is not only the bug. Software has bugs; that is ordinary. The horror is the institution that decided, again and again, over sixteen years, across hundreds of cases, that the computer was more credible than the person standing in front of it, that a system it did not understand and could not audit was a more reliable witness than a human being with an unblemished record and a desperate, consistent story. The sub-postmasters said the numbers were wrong. The Post Office said the computer does not lie. The computer was lying the whole time, and everyone chose to believe it, because it is easier to believe a machine than to admit you have built your operation on one you cannot see inside.

AMERICA

Trump wins Iowa and New Hampshire, and the race is effectively over before February

Donald Trump won the Iowa caucuses on January 15 with a record margin and the New Hampshire primary on January 23, dispatching his last serious rival, Nikki Haley, in all but name and making the Republican nomination a formality months before the convention. Haley vowed to fight on to her home state of South Carolina, but the arithmetic and the party had already decided. The 2024 election, barring the extraordinary, will be the rematch almost no one says they want: Trump against Biden, a contest between the two oldest candidates in the nation's history, its outcome hanging over every other story in this paper for the year to come.

IN BRIEF

The Hague rules on Gaza; the economy defies gravity; a chip in a skull

The International Court of Justice ruled on January 26, in a case brought by South Africa accusing Israel of genocide in Gaza, that the claim was plausible enough to proceed and ordered Israel to take steps to prevent genocidal acts and allow aid, though it stopped short of ordering a ceasefire. The American economy grew at a startling 3.3 percent annual rate in the last quarter of 2023, defying every recession forecast and strengthening the case that inflation had been tamed without a downturn. And Elon Musk's company Neuralink implanted its first computer chip in a human brain, it announced on January 29, an early step in a long and contested project to let people control machines with thought alone.

The Column

The computer said so, and no one was left to disagree

Britain destroyed hundreds of innocent people because a piece of software said money was missing, and an institution chose to believe the software over the people. The horror is not that the computer was wrong. It is that the institution arranged itself so that no human had the authority, or the will, to overrule it.

For twenty years the British Post Office prosecuted, bankrupted, imprisoned, and in some cases drove to suicide hundreds of its own subpostmasters, ordinary people running village post offices, because a computer system called Horizon reported shortfalls in their accounts, money that was never actually missing but that the software, riddled with faults, invented, and it took a television drama, this fortnight, to force a nation to believe the victims it had spent two decades disbelieving. The scandal is being told, rightly, as a story of institutional cruelty and cover-up. But underneath the cruelty is a structural failure that I encounter, in smaller and less murderous forms, in nearly every organisation that runs on software, and I want to name it precisely, because it is spreading and it is about to get much worse.

Lisanne Bainbridge wrote a paper in 1983 called "Ironies of Automation" that everyone who builds these systems should be made to read, in which she identified a perverse dynamic: as we automate a process and it becomes more reliable, the humans left to supervise it lose the practice, the context, and eventually the standing to challenge it, so that precisely as the system's rare failures become more consequential, the humans become less capable of catching them, and more disposed to defer to a machine that is usually right. This is automation bias, the well-documented human tendency to trust the automated output over our own judgment and over other humans, and it grows stronger the more reliable the system usually is, because every day the computer is right trains the operator to assume it is right, until the day it is catastrophically wrong and there is no longer anyone in the loop with the confidence, the knowledge, or the authority to say so. Horizon is Bainbridge's irony at industrial scale and across two decades: a system trusted so completely that the word of the machine outweighed the word, and the character, and the desperate protestations, of hundreds of people who had kept honest books their whole lives.

Let me put the counterargument fairly, because it is not stupid and it is in fact usually correct, which is exactly what makes the trap so deep. Computers are, at arithmetic and record-keeping, genuinely more reliable than humans; the whole reason we automate accounting is that human bookkeeping is error-prone and slow and inconsistent, and an institution that overruled its accounting system every time a person objected would be paralysed, defrauded, and correct far less often than one that trusted the system. In the overwhelming majority of cases, believing the computer over the individual's protest is the right call, statistically, and an organisation cannot function if every automated result is up for negotiation with whoever it inconveniences. This is true, and it is the reason the Post Office's posture was not obviously insane from the inside, and that is the whole horror of it: the failure was not a crude refusal to think, it was the over-application of a heuristic that is usually right, pushed past the point where anyone retained the capacity to notice it had become catastrophically wrong. The danger is never that the machine is trusted; it is that it is trusted structurally, with no preserved mechanism, no empowered human, no institutional will, to catch the case where the usually-right system is this time destroying the innocent.

Because that is where the real fault lies, and it is not in the software, which was merely, banally, buggy, as all software is. The fault is that the institution arranged itself so that the machine's output was effectively unchallengeable: it stripped the local humans of the authority to dispute it, gave them no window into how it worked, structured the incentives so that admitting the system could be wrong threatened the whole edifice and the careers atop it, and thereby removed, one by one, every point at which a human being might have interposed judgment between the machine's error and a person's ruin. The computer said money was missing, and the institution had, over years, dismantled every position from which anyone could effectively say: the computer is wrong, and this person is telling the truth, and I have the standing and the duty to act on that. There was no one left to disagree, not because the people were absent but because the structure had disempowered them, and a machine that cannot be contradicted is not a tool. It is an oracle, and an institution that has built an oracle has abolished responsibility, because responsibility requires a human who can be held to account, and no one can be held to account for deferring to the computer.

I write this now because we are about to do it everywhere, at a scale that will make Horizon look small. We are wiring automated judgment into hiring, lending, policing, medicine, welfare, the allocation of life chances, and we are doing it with the same reassurance that the system is usually right, which is usually true, and with the same quiet erosion of the human positions from which the system could be challenged, and with the same institutional preference for the machine's clean answer over the messy, inconvenient, expensive protest of the individual it has wronged. The lesson of the Post Office is not that we should distrust computers, which is neither possible nor wise. It is that a system, however reliable, must never be allowed to become unchallengeable, that there must always be a human with the knowledge to understand it, the authority to overrule it, and the incentive to listen to the person it is about to destroy, and that building that human into the loop, and protecting them, is not a brake on efficiency but the whole difference between a tool and an oracle. Britain built an oracle in its post offices and fed hundreds of innocent people to it over twenty years, and the machine was only wrong. It was the people who arranged for no one to be left to disagree who were guilty, and we are, right now, arranging the same thing, on a hundred times the scale, and calling it progress.

The Ledger
AI
The Post Office scandal is not an AI story, but it is the most important warning the AI industry could receive: it is the case study of what happens when an institution trusts an opaque system over the humans it accuses. We are about to build far more opaque systems and give them far more authority.
Data centers & power
Quiet this fortnight; the build-out continues beneath the geopolitics.
Rates
Q4 growth of 3.3 percent all but confirms the soft landing. The Fed meets next fortnight; markets expect a hold and hints of the first cut later in the year.
Real estate
US 30-year mortgage 6.69 percent (January 25). Stable, still high, still frozen, still waiting on cuts.
India tech
A quieter fortnight domestically, with the general election approaching. The Horizon scandal is worth Indian attention for a specific reason: much of the world's software, including systems exactly this consequential, is built and maintained by Indian firms, and the lesson about auditability is one the whole industry should carry.
What we called wrong
Nothing to retract. The Horizon column above extends, into its darkest form, the thread this paper opened in its first issue: that a system trusted beyond its warrant, and never questioned, is the most dangerous thing an organisation can own.
The Back Page

The new man in Melbourne

On January 28, on the blue courts of Melbourne, a 22-year-old Italian named Jannik Sinner came back from two sets down to beat Daniil Medvedev and win the Australian Open, his first Grand Slam title. It matters beyond the trophy, because of who was not there. For twenty years the men's game belonged to Djokovic, Federer and Nadal, and this paper has twice this year watched the young ones knock politely at the door, Alcaraz at Wimbledon, and get told to wait. This time the door opened. Sinner beat Djokovic in the semi-final, ending the champion's decade-long ownership of that tournament, and then showed the temperament of someone who intends to stay, calm and merciless from two sets down in his first final. Red-haired, quiet, from a German-speaking corner of the Italian Alps where he grew up skiing before he chose tennis, he is the least flashy of the new generation and, this fortnight, the first of them to actually take the prize. The changing of the guard has been announced for three years. In Melbourne, on a Sunday night, it finally happened.