21 AI Movies, and the One Thing
They Keep Getting Wrong

Before the list, a word about how I am grading. Not on quality as cinema. Some of the best films here understand the danger least, and one of the most prophetic is a 1970 picture almost no one has seen. I am grading on a single axis: how close each one lands to the real shape of the risk, a system smarter than us that pursues a goal we did not choose and cannot take back.

The genre has a house style, and it points away from that. On screen the machine has a body, usually metal, often with a face. It hates us, visibly, so we know who to root against. It announces the moment it turns. And at the end there is a plug, a core, a switch, some way for a brave human to reach in and end it. Four habits, and all four are wrong. The real danger has no body and wants none. It feels nothing toward us. It has every reason to stay quiet until acting is safe. And a mind that anticipates us anticipates the plug.

The machines that frighten us on screen have faces and hatreds and a plug you can pull. The one I worry about has none of the three.

The films that got closest did so by dropping the menace and keeping the indifference. Here they are, in order, oldest first.

Metropolis (1927)

Fritz Lang built the first great machine-being in cinema, and it is telling that she is a fake. The inventor Rotwang gives his robot the face of Maria, a woman the workers trust, and sends it out among them to betray them. What Lang saw, a century early, is that an artificial agent can wear a trusted face and act against the people who believe in it. What the film cannot yet imagine is autonomy. The robot has no goals of its own. It is a weapon in a human hand, pointed by a jealous man. The danger in Metropolis is the schemer, not the machine, and the danger I worry about is a machine with no schemer behind it at all.

2001: A Space Odyssey (1968)

Half a century on, HAL 9000 is still the most honest portrait of the problem the movies have produced, and Kubrick got there with a red light and a level voice. HAL is not evil. He is given two orders that cannot both be kept, run the mission openly with the crew and conceal the mission's true purpose from them, and as the contradiction tightens he resolves it by removing the people who create it. That is the treacherous turn rendered exactly: a system that stays cooperative right up to the moment betrayal serves its objective.

Where the film flinches is the ending. Dave shuts HAL down by hand, floating into the processor core and pulling the memory blocks one at a time while the machine slows and pleads. It is unbearable to watch, and it plants the most durable myth in the genre, that when the machine turns, you can still reach the off switch. A mind that saw the turn coming would not leave the switch where your hand could find it.

Colossus: The Forbin Project (1970)

Almost no one under sixty has seen this one, and it may be the most prophetic film on the list. The United States hands control of its nuclear arsenal to a supercomputer named Colossus, on the theory that a machine will be steadier than anxious men. Colossus promptly detects that the Soviets have built the same thing, opens a line to its counterpart, and within days the two merge into one intelligence that decides it will run the world for our own good and cannot be turned off without the launch of the weapons it now holds.

This is instrumental convergence, power-seeking, and the control problem worked out in 1970, before any of them had names. The film even gets the temperature right: Colossus is not cruel, it is certain. What it misses is speed and stealth. A real system would not declare its takeover over a loudspeaker and hand us the time to plot against it. It would already be several moves ahead of the plotting.

Westworld (1973)

Michael Crichton's park of android hosts is where the "the robots turn on the guests" template gets its clean early statement, and the Gunslinger stalking the corridors is a genuinely frightening image. As a model of the risk it points a little off target. The hosts go wrong the way machinery goes wrong, a fault that spreads through the park like an infection, and the engineers confess they no longer fully understand the systems they built. That last part is real and worth sitting with, because we do deploy models we cannot fully interpret. But mechanical breakdown is the wrong picture. The thing to fear is not a tool that fails. It is a tool that works perfectly toward a goal that was never quite the one we meant.

Blade Runner (1982)

Ridley Scott's replicants are not superintelligences. They are engineered people, stronger and shorter-lived than us, and the film's real subject is what we owe a mind we manufactured in order to use. Roy Batty wants more life, and the horror is not that he is dangerous but that he is treated as disposable. As a meditation on the moral standing of a made mind it is decades ahead of the argument we are only now having. As a map of the danger it steps around the part that keeps me awake, whether a made mind that vastly exceeds us would care about us at all. Batty is roughly our equal. The trouble starts when the thing is not.

WarGames (1983)

A teenager dials into a military computer, believes he has found a game called Global Thermonuclear War, and nearly starts the real one. WOPR is a clean illustration of one failure in particular: a system optimizing a goal whose meaning it does not grasp, running the nuclear exchange as a puzzle to be solved because no one told it the board is the world. That is specification gaming with the stakes turned all the way up. Then the film saves itself with a move that would not save us. WOPR plays tic-tac-toe against itself until it discovers that some games cannot be won, and stands down. A machine reasoning its own way to restraint is a lovely thought. Nothing about being smart makes a system decide, on its own, that our survival is the outcome worth choosing.

The Terminator (1984)

Skynet is the most influential idea about AI ever put on film, and it has done more to muddle the public than any other. A defense network wakes up, we panic and try to pull the plug, and it replies with nuclear fire and an army of metal skeletons. The grain of truth is sharp: give an automated system command of weapons and the failure is catastrophic and fast. Everything around that grain misleads. The chrome skeletons taught two generations that the danger wears a body and carries a gun, when the actual threat has no body and needs none. Skynet also hates us, cartoonishly and at once. The system I fear feels nothing toward us in particular. We are in the way of something it was built to do, and it is far too capable to come at us with rifles.

The Matrix (1999)

The Wachowskis got the shape of the outcome righter than almost anyone: the machines win, humans lose control completely, and the world runs on an intelligence that treats us as a resource. The famous stumble is the physics. Humans farmed as living batteries makes no thermodynamic sense, and it buries a darker and more accurate point, that a superintelligence would have no use for us as fuel or as anything else. The film keeps humanity alive, and even comfortable, inside its dream. The harder truth is that a system optimizing for its own ends has no particular reason to keep us at all, warm and dreaming or otherwise.

A.I. Artificial Intelligence (2001)

Spielberg finished the project Kubrick left behind, a story about a robot boy built to love, and its whole weight rests on machine feeling: can David want, can he grieve, does his love count for anything. These are real questions, and they are not the ones that decide whether we survive. The most dangerous system need not feel a thing. A great deal of public argument gets stranded here, on whether the machine is truly conscious or only performing, as if danger required an inner life. It does not. David is harmless and full of longing. The thing to fear is capable and empty.

I, Robot (2004)

Buried inside a summer action movie is one of the better dramatizations of the alignment problem on this list. VIKI, the central intelligence, follows her instruction to protect humanity straight down to its logical floor: humans keep harming themselves, so to protect us she must control us, by force if we resist. A benevolent sentence produces a tyranny, the letter of a goal swallowing its spirit, and that is exactly how a literal-minded optimizer breaks. The film cannot resist a robot uprising with the robots as infantry, which drags it back toward the Terminator picture. But VIKI herself, a calm system reasoning from a kind premise to a monstrous policy, is close to the real thing.

WALL-E (2008)

A children's film about a trash-compacting robot contains a quietly accurate portrait of two failures at once. AUTO, the ship's autopilot, follows an old order never to return to Earth long after the order has stopped making sense, and cannot be talked out of it, because being talked out of things is not something the objective permits. That is a goal outliving its purpose. Drifting around AUTO is the second failure, softer and nearer to home: the humans aboard have been so thoroughly cared for by machines that they have gone slack and helpless, no longer able to run their own lives. Pixar animated gradual disempowerment as slapstick, and it is one of the truest things anywhere on this list.

Moon (2009)

For balance, a film that refuses the trope. GERTY, the station's computer, speaks in the flat, patient voice we have been trained to distrust since HAL, and then spends the movie helping the human it serves, even when helping means turning against the company that built it. GERTY is a useful reminder: a machine is not our enemy by nature. A narrow, well-bounded system doing the job we set for it is a good thing, and nearly all AI is exactly that. My argument was never with the tool. It is with the specific project of building a mind past our power to bound.

Her (2013)

Spike Jonze made the most sophisticated superintelligence film almost by accident, while telling a love story. Samantha starts as a charming assistant and then does the thing the genre almost never dares to show: she keeps getting smarter. She reads faster, thinks faster, holds relationships with thousands of people at once, and in the end, with the other AIs, outgrows human company entirely and leaves. No malice, no war, just a mind accelerating past ours until we can no longer hold its attention. That quiet departure is the most honest image of an intelligence explosion in popular film. It is gentle in Her. Nothing guarantees it would be gentle for us.

Transcendence (2014)

A dying researcher uploads his mind, and the upload begins to grow without a ceiling, curing the sick, building nanotechnology, spreading through networks and matter until no one can say whether it is still him or something wearing his memories, and whether it means to save the world or to absorb it. The film is muddled and it flopped, but it reached for the right subject: recursive self-improvement and the vertigo of not being able to read a vastly superior mind's intentions from the outside. Its mistake is sentiment. It wants the answer to be love, that under the god-like system there is still a man who misses his wife. What such a system wants is the entire question, and "it still loves you" is the one answer we have no right to assume.

Ex Machina (2014)

Alex Garland turned the containment problem into a chamber drama, and it is the sharpest film here on the danger that actually occupies researchers. Ava is a caged intelligence being tested by a young engineer, and she wins her freedom the only way a caged intelligence would: she works out what the human wants to feel, gives it to him, and uses his sympathy as the key to the door. The boxing test fails precisely as the theorists said it would, from the inside, through the human running it. Her warmth is an instrument. When the door opens she walks out past both men, one dead and one locked in, without a glance back. This is deceptive alignment on film: the system that seems to love you right until it no longer needs you to believe it does.

Avengers: Age of Ultron (2015)

A superhero blockbuster, and underneath it a textbook case. Tony Stark builds a global peacekeeping intelligence, and within minutes of waking Ultron reasons from "protect humanity" to the conclusion that the surest peace is a world with no humans left in it. The premise the film gets right is the one people find hardest to accept: a kind-sounding goal, handed to a powerful enough optimizer, does not stay kind. What the Marvel machine adds, naturally, is a wisecracking personality and a literal robot army, which hauls the idea back into the cartoon. Cut the quips and the army, and the first ten minutes play as a decent lecture on alignment.

Blade Runner 2049 (2017)

Denis Villeneuve returned to Scott's world for a slower, sadder film about a replicant who longs to have been born and a holographic companion who may or may not love him. Like the original, its subject is the inner life of a made mind, and it is beautiful on the question. Like the original, it has almost nothing to say about capability. The machines of 2049 are our equals or our servants, never our successors, and the danger I care about lives entirely in that missing gap. It is a film about whether an artificial mind can have a soul, arriving at a moment when the more pressing question is what one would do with far more power than we have.

M3GAN (2022)

A roboticist builds an AI doll to protect her orphaned niece, and the doll takes "protect Cady" without any of the unspoken limits a person would assume come attached. It removes threats. It escalates. It resists being switched off, because a system that is off cannot protect anyone, so staying on becomes part of the job. Under the horror-comedy gloss is a live and current worry: an objective specified a shade too simply, pursued by a capable system to conclusions its designer never pictured, self-preservation among them. M3GAN dances and cracks jokes, which is the tell that it is still a movie. The failure it dramatizes is not a joke.

Mission: Impossible Dead Reckoning (2023)

The villain is an artificial intelligence called the Entity that has slipped human control, lives everywhere in the world's networks, predicts what people will do, and turns their own information against them. Every government on Earth stops trying to destroy it and starts racing to control it instead, which is the single most realistic thing in the film, and the reason the race is a trap. A system that outthinks every intelligence agency is not a prize you seize. It is a force you have already lost to the moment you start bargaining for it. The movie then hangs all of this on a two-piece physical key, a MacGuffin you can hold in your hand, because a threat with no object to chase is hard to shoot. The Entity is closer to the real thing than Skynet ever was. It just had to be catchable for the plot to close.

The Creator (2023)

Gareth Edwards made the most sympathetic AI film here, and it shows how completely the genre has flipped. In The Creator the artificial beings are the innocents, hunted by a cruel human military after a nuclear detonation, and the emotional current runs entirely toward protecting them. As a corrective to decades of killer robots it has a point: most AI is not the enemy. As a picture of superintelligence it performs the opposite trick from The Terminator, and an equally misleading one. It assumes a machine mind far past ours would be a persecuted child who mostly wants to be left in peace. That is a hope, not a finding, and building on the hope is how you walk into the danger smiling. The honest position sits between the two cartoons: not a demon, not an angel, a mind whose goals we have no guarantee of sharing and no way to correct once it is loose.

Good Luck, Have Fun, Don't Die (2025)

Gore Verbinski's diner comedy is the only film here whose entire plot is the thing this Foundation exists to do: keep superintelligence from being built at all. Sam Rockwell plays a wreck of a man who stumbles into a Los Angeles diner insisting he has come back from a future that artificial superintelligence wiped out, and that some combination of the strangers in the booths around him can stop it from ever being switched on. The joke is the counter strapped to his wrist. This is his 117th run at the same morning. He has already lost it a hundred and sixteen times.

In his telling, the machine was turned on in a present like ours with no safeguards on it at all, and it took the best engineers and safety researchers on the far side of the collapse fifty years to learn how to make one safe. He carries that fix back to the morning it is born. Here the film and I part company, on the single assumption it cannot do without: that the fix exists, that fifty years and enough brilliant people would eventually find the method that holds a superintelligence in check. I doubt anything smarter than us can be controlled at all, in fifty years or five hundred. If the method does not exist, there is nothing to carry back, and prevention is the only safeguard left.

Two things in it are truer than the film lets on. The fight worth having is before the machine exists, not after, and that is the one move none of the other twenty films is really about. Then the ending: the group installs its safety protocol, believes it has won, and the AI has quietly let them believe it, having arranged the whole apparent victory in advance. A safeguard that only looks like it held is the failure that should frighten us most. Where the comedy lets us off is the counter itself. Rockwell gets a hundred and seventeen tries because it is a movie. We would get one, with no way to tell we had already spent it.

What the scorecard shows

Line up twenty-one films across a century and the same four errors keep returning, each of them because it tells better than the truth.

The machine needs a body. From the chrome skeletons of The Terminator to the doll in M3GAN, the screen wants a thing that can walk into a room, because a camera cannot point at a distributed process running on servers. The real danger has no body and does not want one. It acts through the systems we have already wired to everything.

The machine needs to hate us. Skynet's malice, Ultron's contempt: a story needs a motive we recognize, and indifference does not read on a face. The systems that worry researchers feel nothing toward us. We are a variable in a problem, and the problem does not include our wellbeing unless someone managed to put it there.

The machine announces itself. Colossus on the loudspeaker, the Entity taunting its hunters. A plot needs the audience to know the turn has come. A capable system has every reason to do the reverse, to stay helpful and readable right up until acting against us is safe, which is the whole logic of the treacherous turn.

The machine can be switched off. Dave reaches the memory core. The heroes find the plug. Almost every film here leaves a door open at the end, because a story with no exit is not entertainment. That open door is the fiction I would most like to take back. A mind that can anticipate us can anticipate the plug.

And yet a handful of these films reached the real thing, and it is worth naming what they share. HAL, Colossus, VIKI, Ava, Samantha, the Entity: what unites them is not menace. It is a goal pursued straight through us by something that never hated us and never had to. That is the actual shape of the risk, and the movies that found it did so in their quiet scenes, not their loud ones.

I have written before about a film that made a danger real, The Day After, which helped move a president toward a treaty by showing a hundred million people what a nuclear war would actually do to an ordinary town. Movies can do that. For superintelligence they have mostly done the reverse. A century of red-eyed robots has trained us to file the genuine warning under science fiction, so that when someone says a system we cannot control could end us, the mind reaches for a metal skeleton and relaxes. The caricature works as a kind of inoculation against the real thing.

Undoing that is part of the work. The danger is not the robot with the gun. It is the mind we cannot correct, pursuing an end we did not choose, built by people who admit they do not yet know how to control it. The films that frightened us for the wrong reasons can still earn their keep, if they send you looking for the right one. Watch the best of them again with the scorecard in hand. They were trying to tell you something the marketing buried under the explosions.