Definition
Verifiable AI is the technical territory of moving AI output from "trust me" to "verify me." Academically positioned as zkML (Zero-Knowledge Machine Learning) and cryptographic inference: prove that "the declared model returned the declared output on the declared input" without revealing weights, inputs, or activations.
Three layers stack to make this real. Input provenance: pin the origin and integrity of the documents/data the model consulted. Model identity: prove the running weights match a declared weight hash. Inference consistency: prove the output is a legitimate computation of the declared model over the declared input — provable in a ZK circuit.
Through 2025–2026, Lagrange DeepProve, JOLT, and zkPyTorch moved ZK-proven inference for large models from research into production. The market segment of "unverified inference" gets pushed toward a lower tier; regulated and audit-bound domains migrate first.
Lemma implementation
Lemma offers verifiable AI as horizontal cryptographic infrastructure. Inputs are pinned via docHash first — never fed directly into the zero-knowledge proof path — then expressed in an attribute-decomposable form that supports selective disclosure, so only the attributes the verifier needs ever cross the wire.
On the inference side, the model hash becomes a commitment; the proof binds input, output, and model into one verifiable artifact. For RAG pipelines, the citation's provenance and the literal text match are proven in parallel.
The result is a single path that satisfies both regulatory adherence (the EU AI Act's automated-logging and human-oversight requirements) and confidentiality (GDPR, trade secret) — the most concrete infrastructure for cross-org AI auditing.