AI systems are becoming more powerful, more autonomous, and more integrated into real-world workflows. At the same time, a new phrase is appearing everywhere: verifiable AI.
But that phrase is used to describe very different things, including proving that a model ran, proving that a record was not altered, proving that a computation is correct, proving something without revealing data, or proving compliance and auditability.
These are not the same problem. And they are not solved by the same infrastructure.
This article clarifies the distinction between verifiable AI execution and zkML, explains what NexArt actually proves, and outlines the privacy model NexArt supports today.
The Confusion Around Verifiable AI
The term "verifiable AI" is often used as a catch-all. But in practice, it covers at least two distinct categories:
- Execution evidence systems: proving what ran and preserving tamper-evident records of it
- Computation proof systems: proving that a specific computation was performed correctly
NexArt and zkML sit in different parts of this landscape. Understanding that difference is critical.
What NexArt Actually Does
NexArt focuses on verifiable execution records. It produces Certified Execution Records (CERs), which are cryptographically sealed execution artifacts: structured records of inputs, outputs, parameters, and context that are tamper-evident, independently verifiable, and optionally signed through attestation.
These records are designed to capture AI execution evidence.
Definition: Certified Execution Record (CER)
A tamper-evident, cryptographically verifiable artifact that captures the essential facts of an AI execution, including inputs, parameters, runtime context, and outputs, in a form that can be independently validated later.
What a Certified Execution Record Proves
A CER allows a system to prove:
- That an execution record has not been modified
- What inputs and parameters were recorded
- What output was produced
- What execution context existed
- The integrity and chain of custody of the record
This provides execution integrity and supports AI auditability.
What NexArt Does Not Prove
It is important to be precise. NexArt does not:
- Guarantee LLM determinism
- Prove that an output is correct
- Prove hidden computation correctness
- Provide zero-knowledge privacy by default
NexArt is not trying to prove that a computation is correct. It is proving that a record of execution is authentic, tamper-evident, and intact.
What zkML Proves Instead
zkML, or zero-knowledge machine learning, focuses on a different problem. It aims to prove that a specific computation was executed correctly, that a model produced a result according to a defined circuit, and that certain properties hold without revealing underlying data.
Definition: zkML
zkML refers to techniques that use zero-knowledge proofs to verify that a machine learning computation was performed correctly, often without revealing the underlying data or model details.
zkML Is About Computation, Not Execution Records
This is the key distinction:
zkML is computation-proof infrastructure.
NexArt is execution-evidence infrastructure.
zkML answers: Can we prove this computation is correct?
NexArt answers: Can we prove what actually ran?
These are different trust problems.
Transparent Evidence vs Private Proofs
These two approaches represent different trust models.
NexArt: Transparent by Default
- Designed for auditability
- Supports debugging and investigation
- Captures full execution context
- Produces tamper-evident execution records
Best suited for: enterprise AI workflows, governance, agent execution tracking, incident analysis
zkML: Private Proof by Design
- Proves correctness without revealing full data
- Supports confidential computation
- Minimizes information disclosure
Best suited for: privacy-sensitive environments, on-chain verification, hidden model or data scenarios
These models are not mutually exclusive. They can be combined.
Privacy in NexArt: The Levels That Exist Today
NexArt is transparent by default, but supports selective privacy through structured mechanisms. Here is a practical privacy ladder.
Level 1, Full Transparency
The execution record contains the full data. Best for internal systems, debugging, and full audit visibility. Maximum auditability, minimal confidentiality.
Level 2, Verifiable Redaction
Sensitive fields are removed, but the resulting record remains verifiable. Best for external sharing, customer-facing verification, and controlled disclosure. The redacted artifact becomes the new verifiable record.
Level 3, Hash-Based Evidence
Sensitive values are represented as hashes or envelopes. This allows later proof without revealing the data immediately. Best for selective disclosure and proving a value existed. Preserves integrity but does not provide full privacy guarantees.
Level 4, External Evidence Reference
Sensitive data remains outside the CER, referenced through hashes or metadata. Best for enterprise-controlled environments, restricted access systems, and compliance workflows. Stronger operational privacy, but depends on external systems for full verification.
Key principle: NexArt is transparent by default, but selective privacy can be applied without breaking execution integrity.
What NexArt Privacy Is Not
To avoid confusion, NexArt privacy is not:
- Zero-knowledge proof of computation correctness
- Full confidential inference
- Hidden-model verification
- zk-style privacy without zk complexity
NexArt's privacy model is based on selective redaction, integrity preservation, and structured execution evidence. It does not attempt to replace zero-knowledge systems.
Why Execution Evidence Still Matters
Many real-world AI systems need tamper-evident execution records, auditability and governance evidence, structured context around decisions, signed execution artifacts, and independently verifiable records. These needs exist even without privacy-preserving computation proofs.
This is especially important in enterprise AI systems, agent execution workflows, governance pipelines, incident investigations, and regulatory reporting. Execution evidence is often the first requirement.
Where This Fits in AI Regulation
Regulation is increasing the demand for verifiable AI systems. Frameworks like the EU AI Act emphasize traceability of decisions, documentation of system behavior, auditability of AI workflows, and accountability in high-risk systems.
These requirements do not necessarily mandate zero-knowledge proofs. In many cases, they require something more practical: structured execution records, tamper-evident execution evidence, and the ability to reconstruct and review decisions. This is where verifiable AI execution becomes relevant.
Where NexArt and zkML Can Work Together
These systems can be complementary. A practical architecture could look like:
- NexArt records execution context, inputs, outputs, and provenance
- zkML proves correctness of specific sensitive computations
- Together, they provide both auditability and privacy where needed
For most systems today, execution evidence is the practical starting point. Computation proofs can be added selectively.
What This Means for Builders
If you are building AI systems, ask:
- Do you need tamper-evident execution records?
- Do you need auditability and governance evidence?
- Do you need to track agent execution and decisions?
- Do you need selective privacy for certain fields?
- Do you truly need zero-knowledge computation proofs?
In many cases, NexArt provides the execution evidence layer. zkML or similar systems may be added for specific use cases.
Conclusion
Verifiable AI execution is not the same as zero-knowledge AI proofs. NexArt is built for execution evidence: tamper-evident execution records, attestation, auditability, and execution integrity. This is different from proving hidden computation correctness.
Both categories matter. But they solve different problems.
Not every trust problem in AI is a zero-knowledge problem.
This article was originally published on Medium. Read on Medium →