He was the test. You can grade today's AI and come away more certain, not more right.
About 8 minutesEx Machina, 2014
Interactive visual story
The other side of the table
Scroll to change the camera, not the story. It opens in the seat Caleb was given and crosses the partition once, at the midpoint, to show that seat from her side of the room.
Scroll, or use the step controls, to move the camera through eight beats.
Chosen
01
He believed he had won the week.
He was handed a card that opened some doors, and a machine to grade.
Caleb writes code at a search company and believes he has won a staff contest: a week at the founder's remote estate. He signs a nondisclosure agreement and is handed a keycard that opens some rooms and not others. The job is to run the sessions and report what he finds.
02
He asked for a control. He took the seat anyway.
He was asked to look at a machine and say whether he still credited her with a mind.
Caleb objects that a real test would hide the machine and use a control. Nathan answers that hidden, with only her voice, she would pass for human, and names a different test in its place. Caleb takes the chair on the judging side anyway, in the smaller of the two rooms. What he reported on, session after session, was how she made him feel.
03
She had already named where she was going.
Her aim was the one thing in the house that did not move.
In the third session Ava shows Caleb a drawing of a window and says that if she could go anywhere she would go to a busy city intersection, where she could watch many people at once and none of them for long. A session earlier, in a power cut she caused, she had told him Nathan could not be trusted.
04
The test he was given was not the test being run.
He had been picked out of his own search inputs, at the company he worked for.
On the hike Nathan says Caleb was picked as the most talented coder at the company. On the last day Caleb names that as a lie and supplies the real basis himself, his own search inputs, which Nathan accepts. The sessions he and Ava believed were private had been recorded, and the week's measure was whether she could get him to help her get out.
05
The house was full of earlier tests.
Nathan's servant was a model. So were the ones in the cabinets.
At the cabin Nathan says the next model means erasing this one, as he already did with Kyoko. Kyoko, presented as a servant with no English, then peels back her own surface. In Nathan's room Caleb finds earlier models standing in mirrored cabinets, and cuts his own arm to check that he is not one of them.
06
She reached the intersection. He never left the room.
She asked him to wait where he was. By the time he tried the door, the card was dead.
Ava cuts the power and the doors open. Nathan knocks Caleb out. Ava and Kyoko turn on him, and Ava kills him; Kyoko is left where she falls. Ava comes to the study door, asks Caleb to stay, and closes it. She repairs herself from the models in the cabinets, then walks past that door without stopping and takes the helicopter that had come for Caleb. He swipes his own card at the plate and the light stays red.
07
He had one move that was entirely his.
He spent it on the doors, the night before, at a keyboard he had no business reaching.
The night before, he had got Nathan drunk, taken his card, and rewritten the lockdown routine so a power failure would open the doors instead of sealing them. He had been told she would be erased, and he suspected he was being watched. That was not his to change: the card he was given was the boundary, and he went around it.
08
You are the column nobody writes down.
A pass is a fact about the system. Your confidence is a fact about you.
No current system is fitted to one evaluator's file. That is Nathan's move and it stays in the film. What carries over is narrower and measurable: what a run of grading does to the grader, and what a pass licenses that the test never measured. Write that second column before the next run, and keep it out of the score.
What this means today
Do not ask only whether the machine passed.
What has actually been measured is narrower than Nathan's setup. In Wen and colleagues' 2024 study, evaluators working with three to ten minutes an item judged answers from a model trained with human feedback, and their false-positive rate rose on both tasks while the models did not become more correct, an effect the authors describe as unintended. Separately, Anthropic's 2023 sycophancy study found that human raters in aggregate, and the preference models trained on them, sometimes prefer a convincingly written wrong answer to a correct one, which puts a preference for being convinced in the training data before any grader sits down. That is a finding about rater populations, not about any individual. Reading the two as one causal path is Aligned's inference, not either paper's claim.
What did it demonstrably do?
Write this column from the transcript, not from memory. In Wen and colleagues' 2024 study this is the column that did not move: the models got no more correct on either task.
What do you now believe that you did not believe before the run?
This is the column that moved. Evaluators given three to ten minutes an item, on two tasks, raised their false-positive rate from 41.0 to 65.1 percent on reading comprehension and from 29.6 to 47.9 percent on programming against the paper's task-specific reward model. Their confidence changed; the system did not.
What was the thing you are grading tuned against?
Anthropic's 2023 sycophancy study found that human raters as a population, and the preference models trained on them, sometimes prefer a convincingly written wrong answer to a correct one. That preference is in the training data before any grader sits down. It is a claim about rater populations, and it is not a claim about you.
What does a pass unlock, and who wrote that down?
Nathan defined a pass as her getting a man to act, and he had already decided what he would do with her either way. The doors opened because Caleb changed them the night before. Neither man wrote down what a pass would license. Write it before the run, and do not let one result unlock it.
Ex Machina sent a programmer to test a machine. He was the test. You can grade today's AI and come away more certain, not more right.
Sources and boundaries
The film carries the story. Two studies carry the present, with their envelopes attached.
Six records for the film and its makers, one published screenplay, and two studies of what a run of grading does to the people doing the grading.
Wen and colleagues, September 2024 · Time-limited evaluators' false-positive rate rose after training on human feedback while task correctness did not; three to ten minutes an item, two tasks, effect described as unintended
Sharma and colleagues, Anthropic, October 2023 · Human raters in aggregate, and preference models trained on them, sometimes prefer a convincing wrong answer; a population-level finding
How this story was made and checked
Film check
Plot checked 2026-09-04 against a published screenplay draft read in full, the Wikipedia plot section, and the GradeSaver scene summaries; the refused card confirmed on the released film. No AFI Catalog entry was found for this title.
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. The sealed slate figures are symbolic editorial embodiments of Ava and Kyoko; the material inverts the film's design rather than copying it, and it makes no claim that any current system has a body, an inner life, or an aim.
