A preprint posted to arXiv argues that a combination of hash-based encryption and error-correcting codes can push AI decision test accuracy toward arbitrarily low error rates.
The paper extends earlier work on Hash-based Homomorphic Artificial Intelligence (HbHAI), a technique that runs AI algorithms on cryptographically secured data without modifying the underlying algorithms. The new results claim two additions: a compression improvement that shrinks data by up to 10x — cutting compute time and energy use proportionally — and a method that borrows repetition codes from classical error-correction theory to reduce validation error on AI decision tests. The authors say this second result holds "arbitrarily," meaning the error floor can in principle be driven as low as desired.
Homomorphic computation has long been the holy grail of privacy-preserving AI: process sensitive data without ever decrypting it. Most schemes carry brutal performance penalties; the HbHAI line of work claims to sidestep that cost. If the 10x compression figure holds under scrutiny, it would matter for anyone running inference on regulated data — health records, financial signals, biometrics — where plaintext exposure is a liability.
This is a preprint and has not yet passed peer review, so the "arbitrary" error-reduction claim deserves skepticism until independent researchers can reproduce it on standard benchmarks.