Pisa measures. It doesn't guess.

How well it works

40 videos whose answer we know: 29 real ones, filmed with a phone or taken from Google DeepMind's Physics-IQ benchmark, and 11 made by AI video generators. Most generated videos found online had nothing Pisa could measure with confidence (a moving camera, a close-up, too few frames), so it said Inconclusive instead of guessing.

Measuring versus asking

People ask chatbots “Is this video real or AI-generated?” We asked Muse Spark, Meta's model, the same question about the same videos, and scored its answers the way we score Pisa's verdicts.

Five laws every real fall obeys

Each law returns a measurement and a confidence. Pisa calls a video physically impossible only when a law breaks with high confidence.

1. Law 1

   Free fall

   Once it leaves the hand, an object speeds up steadily until it lands. Air can slow that down, never reverse it. An object that slows down while falling, or speeds up faster and faster, is not falling.

   Needs no ruler and no clock

2. Law 2

   Two objects, one fall

   Two similar objects let go together fall together, as Galileo showed in Pisa. Generated videos often drop them at different rates.

   Needs no ruler and no clock

3. Law 3

   Bounces lose height

   A ball never comes back faster than it arrived, so every bounce peaks lower than the one before.

   Needs no ruler and no clock

4. Law 4

   The strength of gravity

   With something of known size in the frame, Pisa measures gravity itself: 9.81 m/s² on Earth. Slow motion changes this number too, so on its own it is never proof.

   Needs a known size and real speed, so it never convicts alone

5. Law 5

   One gravity per video

   Every fall in a clip must show the same gravity. Slow motion slows every fall equally; generated clips often don't.

   Needs no ruler and no clock

Models help. Measurements decide.

Muse Spark

Watches the clip, finds the moment something falls, is thrown or bounces, and names what to follow.

SAM 3.1

Outlines the object in every frame of that moment.

Pisa's physics engine

Turns the outlines into a path and tests it against the five laws. Plain, tested code: the same video always gets the same verdict.

Muse Spark

Puts the result into words in your language, using only numbers the engine measured. Every number is checked.

What Pisa can't tell you

• A consistent verdict doesn't prove a video is real. It means the motion breaks no law Pisa checks.
• Without a known size and a trusted frame rate, slow motion and weak gravity look the same. Pisa calls both consistent instead of guessing.
• Video generators can learn to fall correctly: one study fixed gravity with 100 examples. Pisa checks several laws at once, so fixing one is not enough.
• Pisa needs something that falls, flies or bounces on its own. It doesn't check faces, voices or lip-sync.
• Our test set is small. The intervals show how far the numbers could move.

The research behind Pisa

• Generated videos fall with too little gravity, between 0.38 and 2.27 m/s², and objects dropped together don't land together.

  arXiv 2512.02016

• Asking a model is not enough: a vision model tuned for physics rejected clean, real throws and bounces that measuring accepted.

  Morpheus, ICML 2026

• Physics-based detection is an active field. Pisa's difference is that it shows its measurements instead of a score.

  MotionPhys, arXiv 2608.20770

• Pisa's source code, tests and eval

  github.com/nanoulloa/pisa