1. Title as the costume question

Are five AI models and a judge doing Bayesian reasoning?

2. The answer in one line, and the correct name

Probably not. Asking several models and taking the majority is an ensemble, and whether it helps depends on Condorcet's conditions.

3. Why the costume feels right

Combining several opinions sounds like weighing evidence, and "Bayesian" sounds rigorous. (Editor's framing, not a finding.)

4. The difference that matters

Condorcet's conditions. If each voter is independently more likely than not to be right, a majority vote becomes more likely to be correct as the group grows, approaching certainty (Stanford Encyclopedia of Philosophy, "Jury Theorems"). Ladha (1995) extended the theorem to voters whose votes are correlated.

When AI judges share mistakes. A May 2026 preprint (not peer-reviewed) tested 9 large language models from 7 model families on three natural-language-inference datasets. The 9 judges gave "only about 2 independent votes' worth of information", the panel's accuracy fell 8 to 22 percentage points short of independent voting, and "the best single judge matches or outperforms the full panel" (Kohli, arXiv, 2026).

So: the classic theorem assumes voters who are each better than chance and independent. When models share mistakes, adding more of them adds much less.

5. The correct mechanism's badge and two cases

Not yet stamped. Two cases are not yet written.

6. Open gaps

• A source that sets majority voting, ensembles and Bayesian updating side by side has not been found yet.
• Ladha's specific result on correlated votes has not been read yet, so it is not stated.