# Pisa

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Mariano Ulloa](https://cerebralvalley.ai/u/nanoulloa)
- **GitHub:** https://github.com/nanoulloa/pisa
- **Demo video:** https://youtu.be/JjTlPZmFZY0
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3272

Pisa is a reality check for viral videos. Forward a clip on WhatsApp or upload it on the web, and usually within a minute Pisa tells you whether the motion in it is physically possible: Physically impossible, Consistent with real physics, or Inconclusive.

The problem: generated videos are getting hard to tell from real ones, and asking a chatbot "is this real?" returns an opinion nobody can check. Physics is harder to fake. Research shows video generators drop objects with a fraction of Earth's gravity and let two objects released together land apart.

How it works: Muse Spark 1.3 watches the clip, finds the moments something falls, flies or bounces, and names the objects to follow. SAM 3.1 outlines each object in every frame of those moments. Pisa's engine stabilizes the camera, measures each path and checks it against five laws every real fall obeys: the shape of free fall, two objects falling together, bounces losing height, the strength of gravity, and one gravity per video. The verdict always comes from those measurements, never from a model. Muse Spark then explains the result in the reader's language, and every number it writes is checked against the measurements.

The answer comes with evidence anyone can check: the clip with the object outlined, a lime ghost following the path real gravity allows, and plain numbers. When nothing can be measured with confidence, Pisa says Inconclusive instead of guessing.

How we know it works: we ran 40 videos whose answer we know: 29 real (filmed with a phone or from Google DeepMind's Physics-IQ) and 11 made by AI generators (Kling, Wan, HunyuanVideo, Cosmos, CogVideoX, VideoCrafter). Pisa called 0 of the 29 real videos impossible (95% interval 0% to 12%), and 7 of its 8 conclusive verdicts were right. It caught 1 of 11 generated videos: a Kling clip made from the first frame of our own real clip, where a ball's gravity changes by about a third between flights. That clip first exposed a flaw, which we fixed; since then it comes back Physically impossible in every replay and live check, and the real clip of the same room comes back Consistent. Asked "Is this video real or AI-generated?", Muse Spark caught more generated videos (4 of 11) but called 2 real videos fake, and it can't show why. On 12 more real videos Pisa had never seen, it accused none.

The honest limit: Pisa answered Inconclusive on 32 of the 40 videos. Most AI videos online move the camera, shoot close up or have too few frames, and many real lab falls last under a fifth of a second. Pisa would rather say it can't tell than accuse a real video. Full report: https://pisa-theta.vercel.app/en/how-we-know

Pisa works on the web and on WhatsApp, in English, Spanish, Simplified Chinese and Brazilian Portuguese.

## More from Meta Global AI Developer Hackathon

- [Chidi Enwereji](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3269)
- [Nexora](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3270)
- [Meta_Yaho](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3271)
- [Prashanth](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3273)
- [Oc2cO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3274)
- [Chidimma Joel](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3275)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3272. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt