Keystroke dynamics — how fast you type, how long you hold each key, the rhythm between successive presses — are enough to identify a human author. They're also enough to infer cognitive load, fatigue, neurological conditions, and emotional state. Proving you wrote something by revealing your typing patterns is like proving your identity by publishing your medical records. The verification works. The privacy cost is prohibitive.
Kressin et al. (arXiv:2603.00179) build ZK-PoP: zero-knowledge proof of process. The system captures behavioral features during writing, then proves — without revealing any feature values — that the features fall within human population distributions and that the editing process is consistent with incremental human work. The verifier learns only one bit: “a human wrote this, in the way humans write.”
The construction uses Groth16 proofs and Bulletproof range proofs over arithmetic circuits. Proof generation runs under 30 seconds for a one-hour writing session. Verification takes 8.2 milliseconds. The accuracy matches non-private classifiers to within a few percentage points.
The architecture separates what the proof demonstrates from how the proof is generated. The “what” is a statistical claim: these behavioral features are consistent with a human. The “how” is a cryptographic construction that makes the statement verifiable without revealing the evidence. You can check the claim without seeing the data that supports it.
This matters increasingly because LLM-generated text is becoming indistinguishable from human text at the content level. If you can't tell by reading it, you need a different signal — the process, not the product. But process signals are intimate. The zero-knowledge layer is what makes process attestation deployable at scale: the signal is strong enough to verify authorship and private enough to not constitute surveillance.