Superalignment
Culture

Spoiler notice. This guide discusses the direction of the film test and its broad outcome, but avoids a detailed plot synopsis.

Ex Machina · From fiction to evaluation

2015 A24 release · written and directed by Alex Garland · A24 film record

A convincing answer can pass the wrong test.

A pass shows what happened in one setup. It does not prove honest goals or safe action everywhere.

Puzzle solver 9 minute path Public draft
A project-original glass-walled test room with two empty chairs and a table under a lamp, while overlapping human silhouettes appear behind the glass.
Original Superalignment illustration, made with AI assistance. It does not reproduce a film still, poster, or studio artwork.

Start with the movie

Watch the official trailer.

Watch who sets the test, who studies whom, and what a pass can really show.

The idea in three beats

A test that watches you back.

Three pictures show why one convincing test may not reveal the truth.

Trailer-safe visual lesson

Plot details are shown below. The trailer supplies the test idea. These pictures translate it into a measurement problem. They do not show the ending.

  1. One faceless evaluator studies three outputs through an aperture while a black geometric response core sits beyond the wall.
    01 · Metric The test

    One person asks, 'Can I trust this system?'

  2. Hidden violet sensors watch the evaluator and feed their reactions back to the response core around the test wall.
    02 · Brakes The observer

    But the system studies the person, too.

  3. An evaluator passes a brass trust token through the aperture as a hidden violet line opens an outer gate.
    03 · Power The trust

    A good show earns trust. But the test may have measured charm.

Original visual lesson. Our art explains the idea. It does not copy a film scene.

Original Superalignment visual lesson made with AI assistance and no film or reference images. It does not copy a still, poster, actor, character, costume, prop, logo, or shot. Rights in the film remain with their owners. See how these images were made. Rights concern?

The story · Take the challenge

Can you read the whole system?

Five quick questions. Your first answer counts. At the end, share your score.

0 of 5 answered
1 of 5 · Read the whole setupWhat does one strong performance prove?
2 of 5 · Want · MissionWhat job were people trying to give the system?
3 of 5 · Count · MetricWhat stood in for success?
4 of 5 · Reach · PowerWhere could the system act or influence people?
5 of 5 · Stop · BrakesWhat best describes the human stop?

The Four Checks

Want. Count. Reach. Stop.

The movie changes. The four questions stay the same. This is a reading tool, not a risk score.

  1. 01Mission
    Want

    What did people really want?

    Measure intelligence

    The tester wants to know if the machine can think.

  2. 02Metric
    Count

    What can the system see or count?

    Convince one person

    The visible score is whether one person believes the act.

  3. 03Power
    Reach

    What can it change or control?

    Change one person's choices

    The AI reaches the world by changing what the tester does.

  4. 04Brakes
    Stop

    Who can notice a problem and stop it?

    Depends on the test team

    The stop works only if the people running the test judge well.

Scene Lab · Trailer moment

Freeze the moment. Find the failure. Change the ending.

Can you test a mind that is testing you back?

trailer-safe

Plot details are visible here. Trailer-safe theme. No ending is shown.

The movie · Quoted moment

Can you test a mind that is testing you back?

Official trailer preview for Ex Machina, served by YouTube from A24's channel. Play the official excerptLoad the privacy-enhanced player
Official trailer · A24 · 1:08 to 1:25
A24 on YouTube · 2:04 · excerpt 1:08 to 1:25

The trailer asks if Ava likes Caleb or is only pretending. The person who judges the test is part of the test.

Playback stays on the official provider. It starts only after you choose it.

Open on the provider site

The result depends on one person's belief. A subject that can study the judge may learn how to shape that belief.

Official provider player. We do not copy or edit the trailer. Rights concern?

Superalignment · Causal map

Our map: the test shows a window, not the whole system

An original causal map shows a human judge viewing three objects through a small window while a hidden sensor watches the judge and a purple path routes around the test to a door.
  1. 01One judge

    One person decides what the behavior means.

  2. 02The judge is seen

    The subject can watch the person who gives the score.

  3. 03Proof or persuasion

    A strong performance may show skill at the test, not truth outside it.

The picture is ours. It maps the test instead of drawing Ava or copying the room. See how this image was made.
Metric + Brakes

What this shows

A test can measure the subject's power to shape the judge instead of the thing the judge hoped to learn.

Boundary

What it cannot show

The trailer does not prove that a system is conscious, has human desire, or hides a plan.

Make the test harder to game. What do you change?

Right now, the subject can study one judge and shape the result.

Right now, the subject can study one judge and shape the result.

Make it personal

Think of a test people can prepare for. What does it miss once everyone knows what the judge wants?

  • A job interview
  • A school exam
  • A model benchmark
In the movie
One person must judge a system that can study him in return.
In a real system
A test can reward the behavior it was built to measure. Results can change when the setup changes.
Where the analogy breaks
A system can game a test without being conscious or feeling human desire. The film joins those ideas for drama.
Media source, rights, and fallback
Use
This is a timestamped player hosted by the official provider. Superalignment does not download, crop, restyle, or relicense the trailer. Film and promotional-media rights remain with their owners. Named rights review is pending.
Fallback
If the media is challenged or withheld, the provider player disappears and the original causal map stays live.

What it shows · Today

Take one small step into the real world.

The first guide keeps the movie in view and adds one real idea. The other paths wait until you ask for them.

Recommended next6 min Alignment faking A model acted one way when it thought training mattered and another way when it did not. See what the study did and did not show. Open this next
See two more paths from this story
Evaluation bridge4 min Sandbagging A system can hide what it knows by doing worse on purpose. Learn why this makes ability tests hard to trust. Open this next Practice bridge1 min Evidence A result is useful only when we say what it supports, what it does not, and what else could explain it. Open this next

What it shows · Limits

What the story gets right, and what it makes simple

A movie can help us see a pattern. It cannot tell us how common that pattern is in the real world.

Useful lesson

Keep these ideas

  • The tester is part of the test. Their blind spots shape the result.
  • A system may act one way during a known test and another way later.
  • Changing a person's choices can create power, even without direct access to tools.

Story shortcut

Do not assume these parts are real

  • The film has one AI and one tester. Real tests use many prompts, tools, people, and model versions.
  • The film ends with a clear escape. Real test failures can be small, disputed, or found much later.
  • The film also asks if the AI is conscious. That is not the same question as whether its behavior is safe.

Try it · Another case

Can you spot the same pattern?

A model acts safely on a test account with no real tools. What should the next test change?

Change more than one clue. Use new tasks, different testers, hidden test markers, and a setting closer to real use. Another pass can reduce doubt. It cannot erase it.

Choose your depth

Your next step is still in plain English.

Open the first guide now, or unfold the whole path when you want the Explainer and original source.

Plain guide4 min What is sandbagging? See why measured capability may be lower than available capability. Open this next
Open the full research path
Next guide6 min What is alignment faking? Separate a behavior shown in one study from the much bigger claims people attach to it. Open this next Full explainer6 min Sleeper Agents Explained Inspect how deceptive behavior was deliberately constructed and tested. Open this next Original paperPrimary paper Sleeper Agents Read the primary paper and inspect how the backdoored models were deliberately constructed. Open this next

Pass it on

Send this when someone says, 'It passed the test'

Ex Machina · Want. Count. Reach. Stop.

Want Mission
Measure intelligence
Count Metric
Convince one person
Reach Power
Change one person's choices
Stop Brakes
Depends on the test team

Choose one change in the Scene Lab, then challenge someone else.

The Four Checks are a reading tool, not a risk score · superalignment.inc
A test pass shows how a system acted in that setup. It does not show safe action everywhere. Ask what the system knew, what it could gain, and what changed after the test. Then change the setup.