How to Verify Probabilistic Consistency of Predictive Models

AI systems need verification mechanisms for honest probability predictions: The paper addresses a critical AI safety concern—ensuring that when AI systems make multiple probabilistic predictions (like estimating risks of harmful outcomes), those predictions are internally consistent rather than contradictory, which is essential for trustworthy AI decision-making.

Consistency can be verified efficiently without examining all predictions: The researchers developed a method where a verifier can check whether a predictive model's answers are self-consistent by only examining a small number of specific cases, rather than checking all exponentially many possible predictions—making verification computationally practical.

The approach uses a "proof oracle" to enable independent verification: The method allows an untrusted AI system to provide evidence (a witnessing probability distribution) that its predictions are consistent, which a polynomial-time verifier can then spot-check through strategic queries, similar to how auditors verify financial records by sampling rather than reviewing everything.

Mathematical foundations exist for probabilistic consistency: Building on decades-old work in artificial intelligence, the paper establishes that approximate consistency of probabilistic claims can be placed in the NP complexity class, meaning solutions can be verified efficiently even if finding them is hard—providing theoretical grounding for practical verification tools.

This enables accountability for AI predictions about risks: By making it possible to verify whether an AI's probabilistic claims about potential harms are logically consistent, this work supports the development of AI systems that can be held accountable for their safety-related predictions.