Minority Report Was Wrong About the Crime
Spielberg's film thought prediction would be used to catch us. It turned out to be far more profitable to finish us.
In the late 1920s, a sociologist at the University of Chicago named Ernest Burgess was handed a question that sounds, now, like the pitch for a science-fiction film. The state of Illinois wanted to know which of its prisoners could safely be let out early. Parole boards had always decided this the old way: a man stood in front of them, they looked at him, and they formed a view.
Burgess did something stranger. He went to the paper. He worked through the records of about three thousand men who had already been paroled from Illinois prisons, and he picked out twenty-one facts about each of them: the kind of offence, the man's work history, whether he had a previous record, the sort of neighbourhood he came from, and so on. Each fact got a score of zero or one. Then he added them up.
The totals were startling. Of the men who scored highest, from fourteen to twenty-one, 98 per cent had got through parole without a violation. Of the men who scored four or below, only 24 per cent had. Twenty-one ticks in twenty-one boxes, and a column of numbers had done what the parole board's long look could not. Burgess published his findings in 1928, in a report for the Illinois parole board. From 1933 Illinois began using prediction tables built on his method, and it kept using them for decades.
Nobody in this story could see the future. There were no psychic mutants in a flotation tank, no wooden balls rolling down a chute with a name carved into them. There was a man with a stack of files and a pencil. But in every way that matters, Ernest Burgess built the first precog.
It is a coincidence, as far as I can tell, that seventy-four years later the head of the Precrime division in Steven Spielberg's Minority Report would be called Lamar Burgess. It is a useful one, though. Because the film, which almost everyone remembers as a warning about being arrested for crimes we haven't committed, was really standing on top of a much older and much duller machine. And it got that machine's future almost entirely wrong.
You know the set-up. Philip K. Dick wrote the story in the mid-1950s; it appeared in the January 1956 issue of Fantastic Universe. Spielberg's film came out in June 2002 and moved it to Washington, DC, in 2054. Three precognitive "precogs" lie in a pool and dream murders before they happen. Tom Cruise's John Anderton reads the visions, waves his hands at a glass screen, and arrives just in time to slap a halo on the would-be killer. Precrime has run for six years and premeditated murder has all but vanished. Then the precogs see Anderton himself killing a man he has never met, and he runs.
The question the film asks, the one that launched a thousand undergraduate essays, is about guilt. Can you punish someone for an act that hasn't happened? If the future is known, is it fixed? If it's fixed, was the man ever free?
These are good questions. They are just not the ones the twenty-first century ended up asking. Prediction did arrive, more or less on schedule. It simply didn't show up at the scene of the crime. It went to work in shops, banks and insurance offices, where the money was.
To see why, it helps to go back further than Burgess, to the people who invented the habit of turning a crowd into a forecast.
In 1693 the astronomer Edmond Halley, he of the comet, published a paper in the Philosophical Transactions of the Royal Society. He had been sent birth and death records from the city of Breslau, and from them he built a table showing how many people born in the same year would still be alive at each age after that. He then used the table to work out what a life annuity ought to cost. It was one of the first times anyone had taken the chaos of individual deaths and found in it a price.
Seventy years later, in London, the Society for Equitable Assurances on Lives and Survivorships opened its doors. Its premiums were based on a method worked out by the mathematician James Dodson, who had died a few years before it started trading. Dodson's idea was simple, and it changed the world: charge each customer a level premium according to the age at which they joined. The young paid less. The old paid more. The Equitable is also where the job title "actuary" was first used.
Notice what these people did not do with their predictions. Halley did not try to stop anyone dying. Dodson did not lecture the elderly about their mortality. An actuary has no interest in preventing the future. An actuary wants to price it.
This is the thread that Minority Report dropped. In the film, prediction is a tool of the state, used to interrupt. In history, prediction was mostly a tool of commerce, used to sort. And the clearest sign of where things were heading came in the same year Dick's story was published. In 1956, an engineer called Bill Fair and a mathematician called Earl Isaac put in $400 each and founded a company to help lenders score their customers. You know it now as FICO. The precogs' first real job was in credit.
To be fair to Spielberg, the state did try his version. For about a decade, in the period when "big data" was still said with a straight face, police departments in America bought into the dream of Precrime with real money.
The best-known tool was PredPol, which grew out of a collaboration between the Los Angeles Police Department and the UCLA anthropologist Jeff Brantingham. It drew small boxes on a map, a few hundred feet across, and told officers where crime was likely to happen on their shift. Chicago went further and predicted people. Its Strategic Subject List gave individuals a score for how likely they were to shoot someone or be shot. In Pasco County, Florida, the sheriff's office built a programme that, according to the Tampa Bay Times, kept watch on about a thousand residents it had marked as likely future offenders. Deputies came to their homes again and again, and parents and children found themselves treated as suspects.
Here is what happened to all of it.
In March 2019, an internal audit of the LAPD's programmes found too little evidence to say PredPol had cut crime. In April 2020 the department dropped it, citing the cost during the pandemic, though the Stop LAPD Spying Coalition, which had campaigned against it for years, believed its organising had more to do with the decision. That June, Santa Cruz, one of the first cities to try predictive policing, became the first city in the United States to ban it. Chicago quietly decommissioned its Strategic Subject List in November 2019, after about ten years, and the city's inspector general published a critical advisory on its risk models two months later. The Tampa Bay Times series on Pasco, "Targeted", won the 2021 Pulitzer Prize for Local Reporting. Residents sued, with the help of the Institute for Justice, and in December 2024 the sheriff's office settled. It admitted the programme had violated their constitutional rights, agreed to pay $105,000, and promised never to bring it back.
The company behind PredPol renamed itself Geolitica. In October 2023, the news site The Markup looked at 23,631 predictions it had made for the police in Plainfield, New Jersey, in 2018. Fewer than half of one per cent had matched a reported crime. By the end of that year the company was closing down, and a rival firm had taken on its engineers, patents and customers.
So the literal prophecy of Minority Report was tried, and it failed. That is worth saying plainly, because people still use the film's title as shorthand for a future that is supposedly coming. It came. It mostly didn't work. In several places it was thrown out.
But buried in that failure is the most important idea in this whole story, and it explains where prediction went instead.
In 2016, two statisticians, Kristian Lum and William Isaac, ran an experiment. They took a publicly available version of PredPol's algorithm and fed it Oakland's 2010 drug-arrest records to see where it would send police in 2011. Then they compared those predictions with public-health survey data on where drug use actually happened. Drug use was spread fairly evenly across the city. The algorithm's predictions were not. They clustered in a few poorer, mostly non-white neighbourhoods, the places where police had already been making the arrests.
Then Lum and Isaac did the clever part. They assumed that sending more officers to the predicted places would turn up somewhat more recorded crime there, and fed the new arrests back into the model. The model became more certain. It sent police back to the same streets, which produced more arrests on those streets, which made the model more certain still.
The prediction was not describing the world. It was making it.
There is a name for this, and it is older than any computer. In 1948 the sociologist Robert K. Merton wrote an essay called "The Self-Fulfilling Prophecy". His best-known example was a made-up bank, the Last National Bank, run by a man called Cartwright Millingville. The bank is perfectly sound until a rumour goes round that it is about to fail. The customers rush to take their money out, and so the bank fails. The false prediction became true because people acted on it.
Merton saw this as a disease of social life. The twenty-first century saw it as a business model.
Think about what a prediction is worth to someone selling you something. A weather forecast can't change the weather. But a forecast of what you will do is different, because you are listening. When a system predicts that you will watch a certain show next and then puts it, enormous and already playing, at the top of your screen, it is no longer forecasting your choice. It is helping to make it. Netflix's own engineers wrote in 2015 that about two of every three hours streamed on the service were found through the homepage, the place where their recommendations live.
That is not sinister in itself. Most recommendations are helpful, and some are wonderful. But it changes the economics of prediction in a way Philip K. Dick never had to think about. If you're paid when the prediction comes true, the cheapest way to be right is not to see the future more clearly. It is to push.
The film actually caught the edge of this. In its most famous scenes, Anderton walks through a shopping centre while billboards scan his eyes and call to him by name. Later, after a black-market eye transplant, the Gap's hologram shop assistant greets him warmly as Mr Yakamoto, the previous owner of his eyes, and asks how his last purchase worked out. The joke is that the system is completely confident and completely wrong. In 2002 that felt like satire. Now it reads like a slightly dated product demo.
And then there is the screen. Spielberg asked for an interface that would feel like conducting an orchestra, and the designer John Underkoffler, who had worked on gestural computing at MIT, gave him one. Anderton stands in front of glass with his arms raised, sweeping and pinching and flinging data aside. The technology world loved it so much that Underkoffler went on to build a real version, g-speak, through his company Oblong.
But look at what the image says. A man stands in front of a wall of predictions and moves them with his hands. He is in charge. The data waits for him. The future we actually built turned the picture around. The predictions arrange themselves around us, and we are the ones being swept and sorted.
So here is the near-future question, and it isn't the one on the film's poster. It isn't "Will I be arrested for something I haven't done?" It's this: what happens to free will when the most profitable thing a company can do is make your predicted self come true?
You can already see the shape of the answer in money. A credit score predicts that you are a risk. A higher interest rate follows. The higher rate makes you more of a risk. Burgess's twenty-one boxes never asked a man what he intended to do with his life. Their descendants don't need to ask, because they can set the terms that shape what he does next.
For years, the clearest example in Britain was the insurance renewal. Insurers learned, through exactly the kind of modelling Halley and Dodson had pioneered, which customers were unlikely to shop around. Those customers were charged more each year for the same cover, a practice known as "price walking". The model predicted your inertia and then billed you for it. The more loyal you were, the more you paid.
And then something small happened. The Financial Conduct Authority, Britain's financial regulator, wrote a rule. From 1 January 2022, home and motor insurers could no longer charge a renewing customer more than they would charge an equivalent new customer for the same policy. The prediction was still allowed. The modelling was still allowed. What was banned was the step that turned a forecast of your habits into a penalty for having them.
It was a dull rule, written in dull language and argued over by people in dull meetings. It was also exactly the kind of change that tips a fork.
It is worth remembering what the "minority report" actually is in Dick's story and Spielberg's film. Sometimes the three precogs don't agree. One of them sees a different future, one in which the killing doesn't happen. That dissenting vision is the minority report. And the whole scandal of Precrime is that it threw those reports away. The system could only work if everyone believed the future had one version.
That is the real choice in front of us, and it will be made in the next ten years, mostly by people we have never heard of. AI systems will soon know our habits better than our friends do, and the agents that act for us will be built by companies that profit from what we do next. The prediction will get better. That part is not in doubt.
What is in doubt is whether we keep the minority report: the other branch, the version of you the model didn't expect.
On one branch, prediction is a mirror. It tells you what you are likely to do and leaves the doing to you. It shows its working. It gives you the new-customer price. It lets you see the feed in a plain order, without ranking, and makes "show me something I wouldn't normally choose" a button rather than a bug. On the other branch, prediction is a mould, and the most valuable customer is the one who has stopped surprising anyone.
The difference between those branches is not some great philosophical breakthrough. It is a handful of rules about what a forecast is allowed to do once it's made. Separate the prediction from the push. Let the model guess, but don't let it get paid for being right about you.
Ernest Burgess's tables told a parole board which men were likely to fail. They could never tell the board which men would decide not to. That gap, between the likely and the chosen, is the most valuable thing any of us owns. It's small. It's easy to price. And it's worth fighting for.