Published Apr 7, 2025
By Team Warden
SPEX: Combating AI System Manipulation - How SPEX Enables Verifiable Integrity
TLDR
AI systems can fail by being unfit for purpose or being actively dishonest. While Explainable AI (XAI) and guardrails help evaluate fitness and honesty, they can be compromised if controlled by bad actors. SPEX is a cryptographic protocol that allows users to independently verify whether AI outputs have been altered or misrepresented—providing transparency and integrity, regardless of the operator's intent. It ensures outputs are delivered as claimed, though it doesn’t evaluate fairness or utility.
Assessing AI Outputs: Fitness and Integrity
When working with AI systems, as a consumer, you may encounter different types of behavior:
Fit for Purpose: The outputs are sufficiently accurate and valuable for their intended use.
Not Fit for Purpose: The outputs may lack accuracy, exhibit bias, or otherwise be unsuitable for their intended use, even if built with good intentions.
Honest: The system operates as designed and with good intentions, regardless of whether it is fit for purpose.
Dishonest: The system does not operate as intended; a bad actor manipulates outputs, either to cut costs or gain a strategic advantage.
Challenges in AI System Integrity
AI systems present two main classes of challenges:
Limitations in Fitness for Purpose: AI systems often tackle complex, uncertain problems where perfect accuracy cannot be guaranteed. The inherent nondeterminism of AI models means that even well-intentioned systems may sometimes fail to be fit for purpose. To mitigate this, Explainable AI (XAI) and guardrails can help assess whether an AI system meets expectations. However, even with these safeguards, full confidence in the safety and accuracy of outputs may not always be possible.
Dishonest Operators: Some AI system operators may manipulate outputs—either to reduce costs or gain a strategic advantage. XAI and guardrails can provide some protection, but only if they are not controlled by the same dishonest actors. In many cases, AI systems operate end-to-end, integrating explainability and monitoring tools in a way that leaves consumers dependent on trusting the operator itself.
Ensuring Integrity with SPEX
SPEX is a cryptographic verification protocol built to empower users of AI systems. It enables consumers to independently verify that AI system operators are acting honestly, providing strong guarantees that outputs have not been tampered with or misrepresented. By making AI-generated content verifiable and auditable, SPEX brings much-needed transparency to AI deployments.
Importantly, SPEX verifies integrity, not intent—it does not assess whether an AI system is fair, safe, or appropriate for a specific use case. Instead, it ensures that what was claimed is what was actually delivered.