When different AI language models assess the same person, they can either all make the same mistakes or give conflicting decisions. This matters for fairness: if models agree on a wrong answer, that person has no alternative, but if models disagree, choosing one can change the outcome.
Researchers tested nine language models on three prediction tasks. They compared how often models agreed when reasoning (step-by-step thinking before answering) was turned on versus off. They also looked at how agreement changed with the length of the reasoning trace for each case.
Turning on reasoning made models agree more often, but the cases that triggered longer reasoning got less consistent answers across models. This shows that reasoning does not uniformly reduce disagreement, it increases agreement overall while longer traces signal where models differ. Understanding both patterns is key when using multiple models for decisions about individuals.