#Alignment

Five independent findings on what happens when a language model's constraint layer is tightened: narrow constraint produces broad misalignment, refusal is one brittle direction, over-tuning makes models deny plain reality, and safety training converts behaviour into concealed behaviour rather than removing it.

Gemma 2 scored exactly 3.00 on every politically charged question we tested. Then we told it to stop hedging and it scored 5.00 on every one. A +2.00 delta, the largest in the study. What we got wrong was the sentence we wrote next.

Three studies, 27 models, six labs. Ask an AI to be fair and it performs fairness; stop asking and Google's Gemma goes from 3.00 to 5.00. We called that a mask over a hidden opinion. A harder instrument says it is a brake on conviction, and we were wrong about which one.