Definition
A cryptographic method that lets a system prove a given response genuinely came from the AI model it claims to be, without revealing the model's private 'weights'. It guards against a substitute or tampered inference server passing off another model's output as the real one.
Why it matters
When a customer, auditor or regulator cannot tell which model actually answered, compromised or fraudulent answers go untraced; provable inference closes that gap and is a defence against model-substitution fraud.