The film cautioned against predictive systems and pre-crime, and the we turned it into a screensaver.
For two decades, Minority Report has been shorthand for the future of interaction. Show the film to a designer and one image surfaces: Tom Cruise in a black glove, sweeping translucent panels through the air like a conductor.
Teams chased it, vendors sold against it, and every few years someone demos a rig and invokes the film.
We treated a piece of science fiction as a product roadmap.
But the gloves were set dressing, not the subject. Steven Spielberg’s 2002 film is about a police division that arrests people for murders they have not yet committed, on the word of a predictive system no defendant is allowed to question.
The interface was the spectacle. The prediction engine was the real argument.
Call the gap between them spectacle debt: what a field owes when it copies the part of a vision that demos well and skips the part that would have constrained it. The industry borrowed against it heavily. Twenty years on, both halves of the film have arrived, though neither in the way we expected. We built the interface poorly and abandoned it.
We built the prediction machine quietly and kept it, and the balance is coming due in courtrooms.
The gesture looked effortless on screen. Held for an hour, it becomes labor.The Interface Everyone Remembered
Start with the part everyone copied, an interface the art department did not invent. John Underkoffler, the film’s science advisor, spent a decade making it real, most visibly in his TED talk on the g-speak spatial operating environment, and founded Oblong Industries to ship it.
It never became how anyone computes, and the reason is boring and physical. Chris Noessel makes the point in What Sci-Fi Tells Interaction Designers About Gestural Interfaces: raising your hands above your heart takes cardiac effort, and across a day it accumulates. Designers gave the symptom a name long ago: gorilla arm.
A demo that dazzles is not a workflow that survives.Not a laptop, it’s a coffee table.
I worked on the first Microsoft Surface, which was not a tablet. It was a coffee table with a 30-inch display under the glass, announced in 2007 at ten thousand dollars.
It demonstrated beautifully.
Using it was punishing.
You lean over a horizontal screen with nowhere to rest your arms, reaching across a table, and your shoulders quit before the demo does.
The mouse won the last forty years by letting your hands rest on something, probably the best ergonomic device for working. Ever.
Pro-innovation bias is the tendency to overrate a capability and wave past its limits, and science fiction feeds it. The clip goes up in a roadmap meeting, the room relaxes, and a hard question about who the feature serves gets answered by a movie.
That is the moment the debt gets taken on, and nobody writes it down.
A demo that dazzles is not a workflow that survives. The test that would have caught it costs nothing: name the constraint the original was free to ignore, then check whether your users live inside it. A ninety-second scene never reaches the eighth hour of a working day, and your users do.
Call this spectacle debt in its cheapest form, paid in prototypes.
The prediction moved from precognition to procurement, minus the review that fiction gave it.The System Everyone Forgot
Now the part nobody built demos around. The engine of the film is not the glove but PreCrime, which arrests people for murders not yet committed. Strip out the psychic triplets and the real question surfaces. What happens when the state treats a forecast as grounds for action its subject cannot see or contest?
That system is here now. Risk-assessment tools score defendants on their likelihood of reoffending, and judges use the scores to set bail and sentences. In 2016, ProPublica’s investigation Machine Bias found that COMPAS flagged black defendants as future criminals at nearly twice the rate of white defendants.
The algorithm is proprietary, so nobody it scores can inspect the logic behind it.
We shipped the prediction and skipped the safeguards.<a href="https://medium.com/media/651a7fa314472c6a14736a4675fabe1a/href">https://medium.com/media/651a7fa314472c6a14736a4675fabe1a/href</a>
Predictive policing did the same thing to places instead of people. When The Markup examined Geolitica, previously PredPol, in Predictive Policing Software Terrible At Predicting Crimes, it analyzed 23,631 predictions for Plainfield, New Jersey, finding a success rate under half a percent. The forecasts rarely matched the crimes, but they did concentrate patrols in the same neighborhoods, which produced more recorded incidents, which fed the model reasons to send patrols back.
The film gave PreCrime a review process, a public debate, and a protagonist whose arc is discovering it is wrong about him. We shipped the prediction and skipped the safeguards.
If your product scores people, two things belong in the specification before the model. The first is the false-positive rate, written plainly for the people who will be scored. The second is a route by which one of them can see the inputs and challenge them. Both are cheap early and impossible to retrofit later.
When the system disagreed with itself, that was the finding. It got filed away instead.The Dissent Nobody Shipped
The film is named after the part of its own plot designers discuss least. A minority report is what exists when the precogs disagree, when one sees a different outcome and the majority overrules it.
Those dissents are destroyed.
A system that visibly disagrees with itself is one the public stops trusting, and the division needs trust more than truth.
Sit with that as a design decision, because your product is making it now. Every model carries uncertainty: a distribution of possible answers, a confidence in each, readings that sit close together. Almost none reaches the screen. The interface renders a probabilistic result as one fluent sentence, in the same tone whether the model is certain or guessing.
The system’s disagreement with itself is the most useful thing it knows, and the interface throws it away first.
Worse, the confidence degrades.
OpenAI’s own GPT-4 Technical Report notes the pre-trained model was highly calibrated, its confidence matching the probability of being right, and that calibration drops after post-training.
The process that makes a model agreeable makes its confidence less trustworthy, and the interface presents it as fact.
Design has had guidance on this for years. Guidelines for Human-AI Interaction, the CHI 2019 paper from Saleema Amershi, Dan Weld, Mihaela Vorvoreanu, Eric Horvitz and co-authors, was validated with 49 practitioners against 20 AI-infused products. Its second guideline is to make clear how well the system can do what it can do.
Show the runner-up when the top two sit close. Give the system a state for out of scope, distinct from wrong. Log every low-confidence answer accepted without editing, because that shows where the product is trusted past its competence.
The system’s disagreement with itself is the most useful thing it knows, and the interface throws it away first.
The fiction’s version announced itself out loud. Ours identifies you and stays quiet.The Ads Knew Your Name
One more prediction arrived, and it is the one people quote without noticing what they are saying. Anderton walks through a shopping concourse, the advertising scans his eyes, and greets him by name. The scene gets cited as the future of retail, usually by someone selling identity resolution.
The film’s ads were loud.
They said his name in public, where anyone nearby could hear that the system had identified him. That is a system disclosing itself at the moment it acts.
The fiction’s dystopia at least announced itself. The one we built runs silently and tells you nothing.
Compare the real version. The Federal Trade Commission banned Rite Aid from using facial recognition for surveillance for five years in December 2023, after finding the chain ran it in hundreds of stores to flag people it deemed likely to shoplift. Employees acting on false-positive matches followed customers, searched them, and called the police, and the harm fell on women and people of color. Customers were never told.
The fiction’s dystopia at least announced itself. The one we built runs silently and tells you nothing. The person misidentified never learns there was a match at all.
Notice and recourse were the safeguards the scene made visible, so put them back before launch rather than after legal review. Write down who gets told when the system acts, in what words, and the path they use to contest it.
If nobody can name the human at the end of it, the feature is a prototype.
No gestures required. The system arranges the choices before you reach for them.The AI That Arrived Instead
A deeper misreading sits underneath the interface one. We assumed AI would be something you operate, a console you command and push around. The film taught us to expect exactly that, intelligence directed by visible motion, with the human as conductor.
Real consumer AI arrived with no gestures at all. It showed up as ranking, the machinery deciding what you see before you decide anything.
You don’t wave at it; it reads your history and arranges the menu.
Netflix reports, in its own accounting summarized in New America’s Case Study: Netflix, that more than 80 percent of what people watch comes from recommendations, not search.
We were watching the hands when we should have been watching the precogs.
That is the inversion of the fantasy. The gestural console was at least honest about its power, because you saw the system respond to your hand and knew you were the one moving things. A ranked feed hides that power inside something passive, and the deciding has happened before the screen loads.
So the most consequential interface of the decade is one nobody was invited to operate, which changes what a designer owes it. Build the controls the ranking never offered: a visible reason why this surfaced, a way to see what the filter held back, and an adjustment that changes the ranking itself.
Then measure what engagement hides: whether people found what they came for, or what was easiest to serve. We were watching the hands when we should have been watching the precogs.
Put down the glove. The part still worth our attention is the machine deciding in the background.The Warning Is the Part Worth Keeping
The misreading was not random. We remember the parts of a vision that are fun to build and forget the parts that ask hard questions of us. The gestural interface was fun, a prop you could drop into a keynote deck. The critique of pre-crime was uncomfortable, because it pointed at systems we were racing to ship. So we kept the glove, dropped the caution, and shipped it anyway.
We are now taking on the same debt again, larger. The chat box is this decade’s black glove, the screenshot-friendly half of a technology whose consequential half runs where nobody films. Spectacle debt comes due, and whoever the system was wrong about pays it — the shopper stopped at the door, the defendant scored before the hearing, the person narrowed by a ranking nobody showed.
The film’s real subject was never the hardware in anyone’s hands. It was what happens when a system predicts people and acts on them faster than anyone can object. That is not a 2054 problem, it is a purchasing decision made this quarter.
So before you ship a prediction, answer three questions in writing: what it does when unsure, who gets told, and how they argue back.
What Minority Report got wrong about AI (so far) was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.