Reasoning AI agents do not select at random with the same inputs
An instruction to randomize independently decreased the correlation between agents with the same input, but did not remove it.
Claimed, not confirmed
Researchers did tests on 6 reasoning models for random selection. For GPT-6 Sol and Gemini 3.8 Flash, a test on one agent predicted correlated selections. This occurs when agents have the same identifiers. In a different test, 4 models had to select customer requests for human inspection. GPT-6 Sol came near the target rate, but its selections were predictable from identifiers. The other models selected a small number of requests.
Sources
Posted
