It started with an absurdity. In Bengaluru, a motorcyclist wearing a helmet was fined ₹500 for carrying a passenger without one—the camera mistook the guitar in a case on his back for a passenger. The automated system issued a ticket with a QR code and a photograph that makes one thing perfectly clear: there is one person on the motorcycle; there is no second figure at all. The city police advised him to file a complaint. On paper, he remains the offender, and the guitar is physical evidence.
There is nothing to unpack here about the camera’s optics. The process is what matters: the machine brought the charge; the person has to contest it. Nobody asked whether the system worked, or why, after the wave of false fines in 2022, the city had promised human review of every alert.
Then I remembered that it had not always worked this way. British law once required the reverse: until a machine proved itself reliable, it kept quiet in court. A rare rule, quietly repealed and still contested. That is the one worth looking at.
Automated cameras in Bengaluru were issuing wrongful fines long before the guitar. In late 2022 and throughout 2023, a wave of false tickets swept Karnataka: the system could not tell the difference between a seatbelt and clothing of a similar colour. Drivers in dark shirts and cars with light-coloured belts were especially at risk. Mysuru police said the technology was being refined, while Bengaluru’s additional commissioner insisted that, from December 2023 onward, the issue had largely been solved and every alert was now manually confirmed by a human being.
The guitar test showed that the “solved” problem survives to 2026 in exactly the same form. But the legal mechanics matter more here than the detector’s quality: the fine is issued by a machine, the appeal falls to the person, and he alone bears the burden of arguing with someone else’s algorithm. The same imbalance that brought everything down in Britain a quarter-century ago—only framed not by regulation, but by default.
Where did this default order come from, and who abolished it? That takes us back forty years.
Britain’s Police and Criminal Evidence Act 1984 contains section 69—a short provision in two subsections. It says that a document produced by a computer is inadmissible as evidence unless it is shown that there are no reasonable grounds for believing it inaccurate because of improper use of the machine, and that at all material times the computer was operating properly—or, if it was not, that the malfunction did not affect the output or the accuracy of the result. The second subsection allowed court procedure rules to require further information about such output.
The law stamped the machine as a suspect. Until its reliability was confirmed, its words meant nothing in court. Almost nowhere else in the world had anything like this existed: the Law Commission later noted that Scotland, some Australian states, New Zealand, the United States, and Canada had no separate schemes for computer evidence at all.
The rule did not survive long in its pure form. Just four years later, the Court of Appeal (Criminal Division) began eating away at it.
In December 1988, Lord Justice Steyn delivered judgment in two recent cases where prosecutor, defence, and judge had all misunderstood the section in the same way. It begins with a paragraph that now reads like a warning sent to itself: the law of evidence must adapt to the realities of modern business; computers played, and would continue to play, an ever-increasing role; and if computer output could not relatively easily be admitted, a substantial number of crimes—particularly crimes involving dishonesty—would go unpunished. Three lines later, Steyn added what would become an indictment thirty years on: computers are not infallible; they do occasionally malfunction; software systems often contain bugs; stored information may be subject to unauthorised alteration; and the phenomenon of viruses is also established. Realistically, then, a computer should be treated as an imperfect device.
The judge obliged to demand proof of reliability was the first to say that computers were inherently unreliable. After that, section 69 began rolling downhill faster.
In 1992, the House of Lords heard the appeal of Hilda Shephard, accused of stealing £78.36 worth of goods from a Marks & Spencer in St Albans—the very goods for which she had no receipt. The crucial evidence was the till roll: the store detective had taken all the rolls for the relevant day, and there was no record there of the items found in her car. The Lords held that complying with section 69 did not require a computer expert at all: someone familiar with the operation of the tills was enough, even if they made no claim to technical understanding of the machine.
This did not abolish the requirement to prove reliability, but it turned it into a formality. It was enough to bring in an employee who said that, over years of work, they had not noticed any problems.
Then came the blow that stripped the section of its meaning. In 1997, a driver challenged the reading of a Lion Intoximeter breathalyser because its internal clock was running slow. The lower court held the data inadmissible, and several years passed before the House of Lords ruled otherwise: only faults affecting how the machine processed, stored, or retrieved information mattered; a timer malfunction was irrelevant. What had to be proved was not reliability in general, but only the kind of unreliability that distorted the result. That is a much narrower target.
By the time of that decision, the section was already queued for repeal.
In June 1997, the Law Commission for England and Wales published a report on hearsay and related topics in criminal proceedings. The section on section 69 runs to several pages and is written with remarkable calm. The Commission acknowledges that the rule does not address the main causes of computer error—most of which, as Professor Tapper observed, are either immediately apparent or arise from wrongly entered data—that as technology becomes more complex it is ever less realistic to check and certify every detail of a machine’s operation, and that it is odd to impose this requirement on machine output but not on the expert using it.
The one serious counterargument is stated twice, and rejected both times. Its meaning is this: the presumption may place an excessive burden on the defence, especially where machine evidence is central. The Commission acknowledges that the question is what evidence the defence should have to produce at all, and how realistic it is to expect this from someone who knows nothing about the machine’s workings. Then it immediately writes that it is satisfied that such an arrangement will operate fairly.
Years later, researchers returned to that report with a pencil. In a 2023 article, James Christie examined the sources cited by the Commission and reached an unvarnished conclusion: the Commission misunderstood or misrepresented the views of the principal authors invoked to support repeal; it failed to address the strongest arguments against repeal without replacement; and it ignored the advice of experts, all of whom insisted that courts should look at the reliability of machine evidence. The same issue of the journal also published the responses received by the Commission, including the positions of those who wanted repeal.
The dispute, tellingly, remains unresolved to this day.
The repeal was slipped into the most unexpected place—the Youth Justice and Criminal Evidence Act 1999. Section 60 consists of a single sentence: section 69 of the 1984 Act ceases to have effect. The legislation received Royal Assent in 1999 and came into force on 14 April 2000.
Pinning down the timing was a separate exercise, because it looks too neat to be accidental. In 1999, Britain’s Post Office rolled out Horizon, the system developed by Fujitsu. The rule requiring the reliability of a computer to be proved, and the system that would send hundreds of people to court, began operating at roughly the same time—and the latter outlived the former by twenty-five years. The people later exonerated had nothing to set against Horizon’s data: Fujitsu first disclosed its error logs in court only in 2018, years after the last criminal case, because when the defence asked who would pay to retrieve them, the Post Office suggested doing the maths—it came to roughly £20,000. Tracey Felstead, then nineteen, could not manage that.
The numbers came out like this: the Post Office prosecuted 700 people between 1999 and 2015; another 283 cases were brought by other bodies, including the Crown Prosecution Service. In all, more than 900 people were harmed by faulty data. Twenty years passed before, in April 2021, the Court of Appeal quashed 39 convictions, calling the Post Office’s failures of investigation and disclosure so egregious as to make the prosecution of any of the Horizon cases an affront to the conscience of the court.
By autumn 2025, something else was more interesting—the arithmetic of correction. The Ministry of Justice sent letters concerning 944 records that the Post Office and prosecutors had submitted in February 2024 as possible victims of the faulty system. It turned out that 46 records contained so little information that the person could not be traced, while another 53 people identified themselves. In all, as of 10 January 2025, 951 cases had been accepted for consideration, 98% had been reviewed, and it had been established that the legislation had quashed 2,108 convictions involving 582 people. The gap between the first and final number is simple: one person can have several convictions, and the “guilty” were sought through records kept by that very system.
The Australian version shows what happens when the presumption is built directly into the algorithm, bypassing court altogether. The 2015–2019 scheme came to be called Robodebt by the people it targeted, and it worked like this: the system took income data from the tax office, spread it evenly across fortnightly periods—“income averaging”—and compared it with what the person had once reported themselves. If it found a discrepancy, the person received a letter; if they did not respond in time, the system treated the tax-office data as correct, assessed a debt, and added a penalty for delay.
Silence stood in for admission—that was the scheme’s entire legal content. More than 1.76 billion dollars in debts were raised against roughly 433,000 people, of whom around 751 million was actually recovered from 381,000. In June 2021, the Federal Court approved a 112 million settlement—with interest, an undertaking not to recover invalid debts, and an agreement to acknowledge that administrative decisions had been made without a proper basis. In his ruling, Justice Murphy described what had happened as a shameful chapter in public administration and a massive failure of government, separately noting that the responsible officials and ministers were unlikely to have known about the unreliability of averaging—a proposition fiercely contested for a long time afterward.
It is worth fixing the discrepancy here, because it has not gone away. News reports cite 1.8 billion as the total settlement figure, but that is invalid debt written off; the cash compensation actually paid under the settlement was 112 million. The description of the scheme is the same across all sources, while the amount depends on what one counts as a payment. That is how cases on this scale live: while the court counts one thing, the press office counts another.
And two years later, in the summer of 2023, the royal commission issued a report of nearly a thousand pages and 57 recommendations, with a separate sealed chapter containing the names of people referred for possible criminal and civil proceedings. The same logic—silence as agreement with the data—had travelled from budget saving to the personal responsibility of specific officials.
The direct consequence of the Horizon case was not what people expected. In January 2025, the Ministry of Justice launched a twelve-week call for evidence on the presumption itself, declaring that the rule had demonstrated its harmfulness in one particular scandal. In February 2026, Baroness Shami Chakrabarti, Lord Arbuthnot, and Lord Beamish introduced an amendment in the House of Lords with a straightforward remedy: restore section 69; repeal section 60 of the 1999 Act. The legal organisation JUSTICE, in a nearby briefing, noted that the current presumption is inaccurate, unsafe, and not built for the future. One separate argument concerned generative systems: their output is increasingly produced not by deterministic calculation but probabilistically—and is all the harder to explain for it.
The amendment was not adopted. At the report stage on 10 March, it was withdrawn after the minister promised to ask the Criminal Procedure Rule Committee to develop rules adding safeguards and disclosure where machine evidence is determinative. In the final Victims and Courts Act 2026, which received Royal Assent in April, the word “computer” does not appear even once—I checked with a full-text search of the legislation.
Twenty-five years passed, a scandal involving more than nine hundred people happened, and a four-hundred-page royal commission report was produced—and the solution came down to an instruction to write regulations. What comes of it will not be known until the rules appear. For now, the system remains one in which a mistaken algorithm issues the ticket itself, and a person with a guitar goes to argue with it.
I saved the driest detail for last, because it turns the meaning of the whole story upside down. Section 69 had a second subsection—the very one allowing courts to demand additional information about how the machine worked. In all sixteen years of its existence, not a single such court rule was adopted. The instrument meant to protect people was never once put into operation—and it was repealed partly because it “did not work.”