Committees of AI models can make different decisions from the same starting point

Committees of AI models, even those set to be fully predictable, can produce different final decisions from almost identical starting information. This means companies building these systems cannot rely on simple reruns to check their work or expect consistent behavior.

Many assumed that if you gave an AI system the same input twice, you would get the same output. This paper shows that for groups of AI models, even tiny, meaningless changes in the input can lead to completely different final decisions. This makes it much harder to audit these systems or trust them for critical tasks where consistency is essential.