We Built an Engine That Detects AI-Generated Human Motion.
GPT-5 and DeepSeek Both Failed.
AI can fake a face. AI can clone a voice. But AI cannot generate your motion — not the specific, irreducibly biological, physically-constrained pattern of your nervous system. Here are the numbers.
CAMERA ──→ SST_18PT ──→ PES_4D ──→ 128D_VECTOR ──→ ZK_PROOF
30fps Skeleton Entropy Motion Presence
Local Topology Scoring Signature Verified
◄─────────────── 0 DATA UPLOADED ─────────────────►
All processing on-device. Nothing stored.The Experiment
We built a Rust-based verification engine that analyzes human motion through four independent feature dimensions. Then we ran a simple test:
- Enroll a human: 20 motion samples → one cryptographic signature.
- Issue a challenge: "Draw a circle with your right hand. Tilt your torso 12 degrees. Keep your head still." — unpredictable, multi-joint, with a coupling constraint that shares a kinetic chain.
- Test three responses: genuine human, AI-generated forgery, and a different human.
The Results
Test Case Presence Score Verdict ───────────────────────────────────────────────── Genuine Human 0.9817 PASS ✓ AI Forgery 0.5857 FAIL ✗ Impostor 0.0000 FAIL ✗ Human—AI Gap: 0.3960
The engine rejected the AI-generated motion across four independent dimensions:
Why This Matters
Every identity system in production today — passwords, KYC, hardware wallets — answers one question: "Does the credential match?"
None of them answer: "Is the human who enrolled that credential physically present right now?"
This gap has existed for decades. AI makes it fatal.
In a world where AI agents hold private keys, where deepfakes bypass visual verification, where GPT-5 can generate convincing video of anyone doing anything — the only signal that cannot be forged is the real-time, physics-bound motion of a living human entity.
Motion is not a file. Motion is not a template. Motion is a continuous, high-dimensional, noise-driven field generated by the irreducible physics of your nervous system. AI can approximate its output. It cannot replicate its entropy.
How It Works
The engine extracts a 128-dimensional signature from four feature groups:
- Kinematics (40 dims): Skeletal ratios between 14 bone segments. Your bone-length ratios are physically unique — AI does not know them.
- Acceleration Profile (25 dims): Statistical distribution including Hurst exponent. Human: H ≈ 0.6–0.8. AI: H ≈ 0.5.
- Jerk Profile (25 dims): The third derivative of position. The single most unforgeable kinematic dimension.
- Jerk Spectrum (30 dims): Frequency-domain analysis. Human: 1/f^α scaling (α ≈ 1.0–1.5). AI: α > 2.0 or α ≈ 0.
The Deeper Truth
Every AI motion model — diffusion, transformer, VAE — is trained with L2 loss. L2 loss penalizes the square of the error. A 1 mm tremor deviation is penalized 100× less than a 10 mm trajectory error. The model learns to suppress high-frequency, low-amplitude signals — exactly the signals that make human motion human.
This is not a temporary AI limitation. It is a structural consequence of neural network optimization. The better AI gets at generating realistic motion, the more aggressively it smooths — and the more detectable it becomes.
The AI Paradox: Every improvement in visual fidelity comes at the cost of spectral fidelity. AI faces an impossible tradeoff: look more real, or be more real. It cannot do both.
Run It Yourself
The core engine is open source:
git clone https://github.com/myshapeprotocol cd cli cargo run --release --bin myshape-demo -- --verbose
25 tests. Zero dependencies beyond the Rust standard library and audited crypto crates. See the live dashboard at myshape.com/developers.
What We're Building
MyShape is a presence verification protocol. Not proof of identity — proof of presence.
World (the orb) proves you are a human. MyShape proves you are this human — physically present, authorizing this specific operation.
We are in active development. If you build in identity, security, agent infrastructure, or applied cryptography — we would like to talk.
To the skeptic: every claim in this article is verifiable. The benchmark code is on GitHub. The threat model is published. Attack it. We invite adversarial review.
Repository: github.com/myshapeprotocol · Threat Model: myshape.com/papers/threat-model · Live Demo: myshape.com/motion-demo
Discuss on HN or GitHub. More technical essays below.