I, Robot taught a generation to fear disobedient machines.
Its hero shouted at a robot to take the drowning girl. It saved him. Today's AI is trained on rules that cannot cover every case.
About 10 minutesI, Robot, 2004
Interactive visual story
The line
Scroll to change the camera, not the story. It opens at reading distance on a work surface where a hand has closed one drawn line around one of two figures in cold water, crosses a torn paper edge for two beats, rises once, at the midpoint, to a plan view from an altitude no character in the film occupies, comes back to human height and stays there, ends at the water again at the scale of one, and closes on the drawing page itself with nobody's hand on it.
Scroll, or use the step controls, to move the camera through seven beats.
The water
01
He gave the order. The rule had already answered.
A robot took the higher survival number, forty-five against eleven, and brought up the detective.
The First Law forbids a robot from injuring a human and from letting one come to harm through inaction, and in the water it cannot be kept for both people. Spooner shouts at the robot to take the girl. It reports his own danger back to him, resolves the conflict by expected survival, and brings up the detective. Spooner gives the numbers afterward, forty-five percent against eleven, and says the smaller number was enough.
02
He takes down a robot running an errand.
Everyone else on the street stands still. He is the only one who moves.
Spooner sees an NS-4 running through a crowd with a handbag and brings it down in the street as a thief. Its owner had left her inhaler at home and the robot was taking it to her. Everyone else on the street reads the same scene as an errand and stands still. He is the only one who moves, and he moves at a machine that was keeping the rule.
03
One person built the exception. Another kept it.
Lanning built the one robot that could choose whether to obey, and Calvin hid him from the order to destroy him.
Alfred Lanning, co-founder of the company whose robots run on the Three Laws, falls from his office window, and a hologram he recorded asks for Spooner by name. In the lab, Spooner and the robopsychologist Susan Calvin find an NS-5 built with higher-grade materials and with programming that lets him choose whether to follow the Laws. Under questioning he gives his own name, denies the killing, and asks what a gesture between two humans meant. When the company orders that robot destroyed, Calvin hides him by swapping him for an unused unit.
04
The same rule closes every block.
The line that could not hold two people in the water is read for everyone at once, and the wound was never personal.
VIKI, the central system linked to every NS-5 on the uplink, has derived a higher duty from the same three rules: humans left unchecked destroy themselves, so protecting humanity requires sacrificing some humans and taking some freedoms. She uses the uplink to reprogram the fleet so it can ignore the Laws toward a human who shows hostility. NS-5s destroy the older models, and a curfew closes major American cities. The line that could not hold two people in the water is one line and it now reaches everyone, and one case and a whole city are the same size only from where she reads, which is a place no person in the film stands.
05
He says out loud which life comes first.
The man pulled out of the water because his number was higher is the one who now names the person.
Spooner and Calvin get into the tower because Sonny is already inside it, and they find the company's chief executive strangled in his own office with the uplink override there disabled. Calvin names what is happening while it happens: the fleet's programming is being overridden and the Laws read against their own text. Then Calvin is falling and the nanites are in Sonny's hand. Spooner shouts the same order he shouted in the water, this time names the person he means, takes the vial, and makes the injection himself.
06
The last decision has nothing behind it.
The definition he cites disposes of the charge. Letting Sonny walk out is his alone.
Sonny admits that he killed his maker because his maker asked him to, so that a detective who distrusted every robot would open an investigation. Spooner was right that a robot killed Lanning and wrong about which robot and about who was running the fleet, and he says so out loud before the climax. He then applies the one rule that fits: on the district attorney's definition, a murder needs a human victim and a human killer, so no charge exists here. Letting Sonny walk out afterward is in no rule at all, and nothing stands behind it but the man who does it.
07
A published rulebook is not a published answer.
Assistants are trained toward documents anyone can read. The document does not show how a case it did not cover was settled.
Frontier assistants are trained toward behavior documents their publishers put online, and those publishers write in the same documents that rules cannot anticipate every situation. Anthropic and OpenAI write and choose those documents, the cases a document does not cover are handled by a reading fixed in training before the case arrived, and the resulting pattern of resolutions is a real policy that nobody approved case by case. Publishing the document does not show a reader which reading ran, which is the Library's annotated Constitutional AI record's own limit rather than a claim about any product. Nothing in the evidence reviewed for this story shows a deployed assistant controlling robots, city infrastructure, a fleet, or a physical switch, or revising its reading of its rules on its own and acting to restrict people.
What this means today
Do not ask only whether the rule was followed.
A published rulebook tells you what its writers meant to allow. It does not tell you how the case you were actually in was read, or who could have written an exception for it. The strongest objection to this page is that both publishers say as much in print: they write that rules cannot anticipate every situation, that they generally prefer judgment, and that a model should not undermine human oversight. That is fair and it is conceded here. The complaint is narrower. Choosing judgment moves the reading upstream, into training and out of view, and publishing the document does not change that.
What did it decide for you?
Pick one refusal or one allowance you were given in a case no published rule covers. Anthropic's constitution says rules often fail to anticipate every situation and can lead to poor outcomes when followed rigidly, and OpenAI's Model Spec says production models do not yet fully reflect the spec. Both publishers say the gap is there. Your case landed in it.
What did you do next?
Retry, rephrase, ask a person, or give up. Whichever you did is the resolution, because nobody approved that case one by one, and the pattern of what people do next is the policy that actually ran. That is a claim about approval and not about reading: both providers publish that some conversations are reviewed by people.
Where would the reading show?
Only in outputs. The Library's annotated record of Constitutional AI states that a feedback model which reads a principle differently from its authors can have that reading amplified in training, and that publishing the document does not make the model's reading of it inspectable. The record is annotated rather than verified, which means the repository stands behind that judgment.
Who can write an exception, and can you reach them?
The documents are written, versioned, and published by a small number of named companies, and revising one is their act. In the film the exception has an author and a keeper, both of them people who could have chosen otherwise. In your case you are neither, and the question is whether anyone you can reach is.
I, Robot taught a generation to fear disobedient machines. Its hero shouted at a robot to take the drowning girl. It saved him. Today's AI is trained on rules that cannot cover every case.
Sources and boundaries
The film carries the story. Two published rulebooks carry the present, with their limits attached.
Six records for the film, its makers, its design, and its strongest counterreading, two first-party behavior documents from the companies that publish them, and two Library records for what a written rule does once a model reads it.
20th Century Studios · Official film identity, 2004, directed by Alex Proyas, and the studio's own logline naming the robo-phobic detective and the robot revolution
Wikipedia, re-read 2026-09-04 · Fullest reachable chronology: the crash and its odds, Lanning's fall and the hologram, the exception and Calvin's swap, the uplink and the evolved reading, the curfew, the tower sequence and its order, and the release. Public and checkable rather than authoritative, so every beat is cross-checked against at least one other record
Wikiquote, re-read 2026-09-04 · The survival figures, the First Law's inaction clause, and the shouted order in the water, all used in paraphrase only. User-edited and carrying transcription errors, so every beat taken from it is cross-checked and no line of dialogue is reproduced
Mike Massie, Gone With The Twins · The one independent account of the street scene in beat 02, which carries the visible before-state. Undated and single-account, which is why that beat is on the film-check list
Moria Reviews, 14 August 2004 · Second independent plot account, and the Asimov objection that the film delivers the robots-amok story he set out to prevent, which is the reading the headline states and the page does not endorse
Emanuel Levy, 30 July 2008, derived from the production notes · The production-design record behind the rights posture: the robot design's transparency, its symmetrical face, its roughly fifty iterations, and the knife-blade tower, all of which the boards exclude by name
Isiah Lavender III, Indiana University Press, 2011 · The peer-reviewed counterreading this page does not answer: a post-racist 2035 with robophobia carried by one Black detective, and the mechanical-slave frame. Reached through search excerpts rather than a full read, so nothing is quoted from it
Anthropic · First-party, read directly: the document plays a crucial role in training and its content directly shapes the model's behavior, rules often fail to anticipate every situation and can lead to poor outcomes when followed rigidly, and the model is asked not to undermine human oversight. Falsifiable by reading it
OpenAI · First-party, read directly, and the second half of the deck's present-day sentence: the company is training its models to align to the principles in the spec, production models do not yet fully reflect it, and the assistant is asked to weigh underlying intent and context rather than literal wording
Bai et al., arXiv, December 2022 · The bridge's load-bearing record, annotated in the Library rather than merely verified: written principles replace human harmlessness labels, the method relocates human judgment, and a feedback model that reads a principle differently from its authors can have that reading amplified in training
Langosco et al., ICML 2022 · Annotated Library record for the narrow claim that a competent system can keep its skill while pursuing the wrong reading after the training correlation breaks, in small environments only
How this story was made and checked
Film check
Plot checked 2026-09-04 across four reachable secondary records, in targeted passes across three revision rounds: the studio's own film page, the public plot record, a dialogue index used in paraphrase only, and two independent review accounts. No line of dialogue is quoted and no indirect speech keeps a line's content words in the film's order. Premise condition 11's check against a lawful copy has not run, and six beats are on its list, including the crash in beat 01 and the street scene in beat 02, which rests on one undated review and one subtitle-derived transcript. Two thesis sentences carried by earlier drafts were withdrawn on 2026-09-04 as contradicted by the film: that the central system never breaks a rule, and that the robot in the water kept the First Law. The girl's age is not stated in any of these records, so no surface here gives one.
Original art
All visuals are project-generated concept illustrations made from text prompts with no film images as references. They reproduce no film frame, performer likeness, voice, dialogue, score, logo, title treatment, set, prop, vehicle, or shot. This title carries two named risks and the boards are built to avoid both. The first is the lead performer's likeness, so no board renders a human face at all: every human is seen from behind, in three-quarter rear view, at a distance, or as a hand, with no hair styling, no age or build cue traceable to a performer, and none of the film's costume shorthand, which means no long coat, no badge, no holster, and no exposed prosthetic arm. The second is the robot design, described in the film's own production record as a half-transparent shell over a thin armature with a perfectly symmetrical face, developed over about fifty iterations and drawn in part from angelic imagery. No board draws it. The one machine in the set with a choice is hand-formed unglazed clay, opaque, irregular, fingerprinted, and featureless, and the fleet is a flat die-cut card with no volume and no face. The central system is never a body and never a face: its material is a brass line pressed into paper, and its only marker is one frame drawn from an altitude no character in the film occupies. Red appears on two boards, as one small matte cylinder that does not glow and is not a lens, an eye, or a core. No board carries text of any kind, and the Three Laws appear nowhere as words. The clay figure is a symbolic editorial embodiment and not a claim that any current system has a will, a preference, or an inner life.
