On the Foundations of Trustworthy Artificial Intelligence

Platform determinism is foundational to AI trustworthiness: The paper proves that deterministic inference—producing identical outputs across different computer systems—is both necessary and sufficient for trustworthy AI. This means that if an AI system produces different results on different hardware, it undermines trust in fairness, robustness, privacy, safety, and alignment.

Current AI systems have a hidden trust problem: Standard floating-point arithmetic (IEEE 754), which powers virtually all AI systems today, fundamentally violates determinism requirements because it can produce slightly different results across different platforms. This creates an unrecognized vulnerability in how we verify and trust AI systems.

Verification becomes dramatically easier with determinism: The paper demonstrates a "Determinism-Verification Collapse" showing that verifying a deterministic AI system requires only simple hash comparison (constant time), whereas verifying non-deterministic systems faces mathematically intractable problems. This has practical implications for auditing and oversight.

A working solution has been demonstrated at scale: The researchers built a pure integer-based AI inference engine in Rust that achieved bitwise identical outputs across different computer architectures (ARM and x86) in 82 tests on models up to 6.7 billion parameters, with zero mismatches verified through blockchain attestations—proving the concept is technically feasible.

AI trust ultimately depends on mathematics, not just policy: The paper reframes trustworthy AI as fundamentally a question of arithmetic and deterministic computation rather than solely a governance or ethics problem, suggesting that technical architecture choices are prerequisites for meaningful trust guarantees.