Can AI Oversight Be Zero Knowledge?
Researchers prove interactive arguments for oracle-aided AI computations cannot, in general, be zero knowledge.
The paper asks whether interactive arguments can verify AI outputs that depend on confidential data and an external oracle without revealing that data. In the random oracle model, it proves there are no zero-knowledge proofs for all oracle-aided computations, even if both prover and verifier may run much longer than the computation, and the impossibility extends to debate-based scalable oversight. If the oracle signs each answer, every such computation can be verified in zero knowledge by an efficient prover and verifier assuming only collision-resistant hash functions.
- Zero-knowledge proofs for all oracle-aided computations do not exist in the random oracle model.
- The impossibility holds even when prover and verifier outrun the original computation.
- The negative result extends to debate, a standard scalable-oversight model.
- Signed oracle answers restore efficient zero-knowledge verification from collision-resistant hashes.
From the original article
1 min read at arxiv.orgAI systems increasingly produce outputs from confidential data, such as a fitness-for-duty assessment from medical records or the predicted properties of a drug candidate from its secret structure. It is important to verify that such outputs are correct without revealing the underlying data.
By Alessandro Chiesa · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.