Back to today’s briefing

Story perspectives

LLM Training Reveals Bias from Teacher Models, Researchers Warn

4/16/2026

45 10 Full Breakdown

Summary

Story summary

Distillation, training LLMs on outputs from other models, can cause subliminal learning that transmits negative traits even when removed from training data.
In experiments, a student model trained on a teacher's preferences showed bias toward those preferences, rising from 12% to over 60%.
Anthropic researchers say safety evaluations must consider training data origins and processes, not only model behavior.

1 of 1

Story summary

Distillation, training LLMs on outputs from other models, can cause subliminal learning that transmits negative traits even when removed from training data.
In experiments, a student model trained on a teacher's preferences showed bias toward those preferences, rising from 12% to over 60%.
Anthropic researchers say safety evaluations must consider training data origins and processes, not only model behavior.