One open training recipe makes AI fact-checkers judge by job status

Researchers tested models from six model families, before and after post-training. They checked if a model judges a claim incorrect more frequently when the poster has a job with lower status. Only the open Tülu-3 recipe caused this, in OLMo-2 32B and Tülu-3 70B. The recipes of Meta, Alibaba and Mistral did not. The authors say the effect is small but statistically reliable.

What they found

In OLMo-2 32B, a lower-status job moves the model before post-training to the answer correct. It moves the post-trained model to the answer incorrect. The model does not notice the job more. It reads the job differently.

The proof

One direction in the model carries the change. It occurs at supervised fine-tuning, the step that trains on example answers. The authors confirmed the mechanism on 400 new claims, with predictions set before the test.

What is not known

None of seven deployed commercial models distrusts posters with lower status more. The text gives no cause in the Tülu-3 data.

Insight: One training recipe can change whom a checker believes

In October 2026, a Japanese test found that six models fact-check the primary claim less correctly than side claims. Average scores hide this. This paper finds a second hidden fault in one training recipe. Average scores hide faults, and the buyer who wants results for each group gets more control of the vendor. Get results for each group and each recipe.