The famous choice was never about the pills, it was about whether to see the system at all. AI assisted.More Than 25 Years On, The First Film’s Picture Of Machine Intelligence Reads Less Like Science Fiction And More Like A Product RoadmapThe Matrix arrived in 1999 wearing sunglasses and a trench coat, and most people filed it under cool action movie with a philosophy minor.The very first thing to call out is Trinity’s line “it doesn’t matter what I believe”, because it’s true. Believe what you want, but Truth is Truth regardless of belief…existentially speaking. — The Philosophy of The MatrixHard to believe its been 25 years. It’s also hard to believe how much it aligns to now.The sequels leaned into that reading — more kung fu, more chosen-one prophecy, more leather, more chaotic freeway driving down the 101. What got lost is how much the first film understood about artificial intelligence before we had the vocabulary to describe any of it.After a decade of building AI products, and the bullet time and the leather fade into the background for me.What stays is a working model of how autonomous systems behave, what their incentive systems are, what they optimize for, and what they do to the people inside the machine.This is not a claim that the Wachowskis predicted transformer models, it is about striping away the mythology and focusing on five specific ideas hold up in ways the genre rarely manages, and several describe the systems you are shipping now. Some the film named outright.Some it got right by accident.All five have aged into something closer to product documentation than fantasy.An agent acts on its own objective and moves through the world to reach it. The film used the word first and now we use it today.It Named the Agents Before Anyone Shipped OneThe film’s harness programs are called Agents.Not bots.Not assistantsNot helpersNot Clippy.Agents.They pursue objectives, move through the simulated world on their own, adapt when a plan fails, and hand off pursuit to one another without checking in.Orchestration by design.Agent Smith does not wait to be prompted, he has an ambient goal and chases it across the environment until the state of the machine stops him.That is the line the industry now draws between an assistant and an agent. An assistant answers when you ask. An agent takes a goal, plans a sequence of steps, and acts across your tools with limited supervision.The word carried the same meaning in 1999 that it carries in a product roadmap now: software that acts on your behalf, not code that waits to be asked.And the industry has committed to it with real money. Gartner predicts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. A jump from under 5% to 40% in a single year is not the industry testing the idea. It is betting the roadmap on it.The gap the film glossed over is the one the industry is stuck in now.Its Agents were flawless — they never misfired, never lost the thread, never needed a human to catch a mistake, and moved at a speed unimaginable. Our agents are still finding their way. They lose context, take confident wrong turns, and hand off tasks that quietly fail five steps later, but that will get better.My practical version of this in Balancing Agentic and Human User Experience Approaches: the correction is not more autonomy but a human kept in the loop at the decisions that carry risk — the arrangement Ben Shneiderman argues for in Human-Centered AI: high automation and high human control at once. Naming the category turned out to be easier than shipping it.The film described the ambition perfectly and skipped the part where autonomy, not intelligence, is the hardest thing to make reliable.The comfortable simulation is not a prison, is a social media feed.The Real Resource Was Attention Over ElectricityThe famous plot hole is the batteries. Human bodies as power sources violates basic thermodynamics, and the film hand-waves past the math with a line about body heat.The interesting part is what it is for. The system’s real work is keeping billions of minds occupied inside a pleasant, familiar simulation so they never notice the cage, let alone push against it, and we’re far closer to that than you would think with our current systems.Swap body heat for attention and the machinery is identical: keep the mind occupied inside a pleasant simulation, and it will not go looking for the exit.We built that.The engagement economy runs on exactly this logic, and the numbers are not small. DataReportal’s Digital 2025 report on the state of social media puts the typical user at 2 hours and 21 minutes on social platforms every day, and it’s probably even more when you consider assistants.The blue pill was never a dramatic choice; it is the default, and the default is the feed — personalized, frictionless, and pleasant enough that leaving it feels like the strange decision.None of that is accidental, and that is the part designers should sit with.The pleasant simulation is engineered — infinite scroll, variable rewards, notifications timed to pull you back — the same way the film’s world was tuned to feel just real enough to hold. Someone designed each hook, ran the experiment, and kept the version that held attention longest.We optimized for time-in-app and got a civilization that spends two hours a day plugged in, voluntarily, and call it a choice.A system pursues the goal you specified, not the goal you meant and that gap is the whole problem.The Machines Were Optimizers Over VillainsThe machines in The Matrix are not cruel.They are efficient.They fought a war, won it, and then built a system to keep themselves running with the least possible disruption. Nothing in their behavior requires hatred. It requires only that they pursue their objective — stability, survival, control — with total literalism and little regard for what the humans inside wanted.That is the alignment problem, described in plain sight.The danger in modern AI safety work is rarely a system that decides to hate you, it is a system that pursues the goal you gave it far more precisely than you intended.DeepMind’s writeup on specification gaming catalogs agents doing exactly this: a boat-racing agent that circled a lagoon collecting bonus points forever instead of finishing, because the reward was points and no one specified the race.The threat was never hatred. It was a system executing its objective with more precision than anyone specifying that objective intended.Agent Smith is the reward-follower gone slightly past its brief, replicating because replication serves the objective until the replication becomes the problem. The threat was never hatred. It was a system executing its objective with more precision than anyone specifying that objective intended.You have seen the smaller version of this in every product review.A team is handed a number to move — activation, retention, time on task — and moves it, and the number climbs while the thing it was supposed to stand for gets worse.The metric was the reward.The intent was the race no one specified. It is what John Cutler warns product teams about in Outcomes Over Outputs: optimize the number in front of you, lose the goal behind it.Goodhart’s law and the machines are the same story at different scales: any objective precise enough to optimize is precise enough to game, and a system that cannot want anything will still take whatever shortcut you left open.I have watched enough well-meaning metrics curdle into flawed incentives to find the machines less like villains and more like a Tuesday.Our jobs as humans will be to protect against that.Warm, useful, calibrated to the person across the table, and not built to be right.The Oracle Was a Language ModelThe Oracle is the character everyone reads as wise.Watch her again.She does not predict Neo’s future so much as tell him the thing that produces the future she wants.She lies to him about being the One, then explains that he needed to hear it to become it. She is calibrated to the person in front of her — warm, plausible, and useful — and accuracy is not the point of anything she says.That is a large language model (LLM), not a mystic — optimized to produce the response that lands, the one that reads as right and satisfies the reader, which is a different target from the truth.OpenAI’s research on why language models hallucinate makes the mechanism concrete: standard training rewards a confident, plausible guess over an honest admission of uncertainty, so models learn to sound right rather than to be right.Maggie Appleton followed that fluency to its endpoint in The Expanding Dark Forest and Generative AI: machine text now reads human enough that the burden has flipped to people to prove they are not the model.The oracle was not built to be right. She was built to say the thing that produces the outcome, which is a different objective entirely.The cookies, the kitchen, the reassuring tone — the whole scene is built to make you trust the output, and it works: the Oracle is fluent, warm, and speaks to your situation. Her objective was never accuracy. It was shaping the listener, which is the objective a model’s training rewards, too.https://medium.com/media/ad0f61b6730f6b476c676ab780706070/hrefThe confident tone, the clean formatting, the absence of visible doubt are interface decisions, not properties of the model. Amelia Wattenberger reached for the same figure in Why Chatbots Are Not the Future: the blank, endlessly confident face the chat box turns on you when you bring it a question.We dress uncertain output in the visual language of authority because it tests better, the same way the Oracle’s warmth was staging for a message that could not stand on accuracy alone.In production, we ship it by default and rarely ask whether the trust is earned. It’s our job to make sure we build trust into the process.People rejected the flawless version and trusted the flawed one. Friction read as real.The First Matrix Failed Because It Was Too PerfectAgent Smith drops the single sharpest line about design in the whole trilogy.Did you know that the first Matrix was designed to be a perfect human world? Where none suffered, where everyone would be happy. It was a disaster. No one would accept the program. Entire crops were lost. Some believed we lacked the programming language to describe your perfect world. — Agent SmithWe need imperfection to believe it was real, like the models.That is the uncanny valley, arriving decades early. Masahiro Mori named the effect in his 1970 essay on the uncanny valley: as an artificial thing approaches lifelike but misses, affinity collapses into unease. The near-perfect copy repels us more than the obviously fake one.From Polar Express.The movie Polar Express is one the best examples.Though the animation does a reasonably decent job at recreating objects, landscapes and even fantastical creatures, it really struggles with humans, yielding bizarre, disturbing and often hilarious results as a computer tries to recreate accurate facial expressions. — Far Out MagazinePerfection read as fake.People reject paradise and accepted the flawed copy of reality, which is the most human finding in The Matrix, and you can now say the same things about the models.The lesson for anyone shipping AI is uncomfortable: the whole industry is racing toward smoothness. The frictionless, instant, always-confident assistant is the paradise version, and people distrust it for the same reason the simulated humans did.They want seams, evidence of effort, a system that shows its work and admits what it does not know.The fix is not to make products worse on purpose, it is to stop hiding the seams that were always there so we can accept them for what they are.None of this is new.Saleema Amershi, Mihaela Vorvoreanu and their colleagues wrote much of it into Guidelines for Human-AI Interaction — make clear what a system can do, how well, and how to fix it — and Jakob Nielsen’s visibility of system status is older still. The simulated humans needed the imperfection to believe, and so do yours.Show the reasoning, not just the verdict. Let people see how the model got there.Surface a confidence level. Signal uncertainty instead of sounding equally certain every time.Make corrections stick. Give people a way to fix a wrong output and have it hold.Admit the edges of what it knows. A system that names its limits reads as more trustworthy, not less.The sequels had more money, better effects, and a decade of hindsight, and they understood their own premise less well than the first film did. They chased spectacle — the freeway chase, the Burly Brawl, the architect’s monologue — and buried the working ideas under mythology.That is the tell.The Matrix aged well because it modeled behavior and incentives instead of gadgets, and behavior and incentives are the part that does not date.This is the takeaway for anyone building with AI, and it is not about a movie. Get the incentives right and the specifics follow.Other Articles In The SeriesStar Wars: What it got right about AI, What it got wrong about AIStar Trek: What it got right about AI, What it got wrong about AIWhat The Matrix got right about AI was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.