Abliteration Removes Model Refusals but Fails to Eliminate Script-Dependent Bias
October 22, 2026
Testing on Nemotron 3.5 Bolt 74B and Qwen3.8-Flash-Next shows that abliteration removes refusal activations but leaves political biases intact. Bias results varied significantly by script, with models affirming specific geopolitical positions in Simplified Chinese that they rejected in Traditional Chinese or English.
HOW THIS AFFECTS YOU
●
builderYou cannot assume that an 'uncensored' model is free from underlying political or linguistic biases.
●
policyThis highlights the critical need for safety evaluations that account for multilingual and script-specific bias.