Does removing an AI model's refusals remove political bias?
A preprint testing two model families found that community modifications reduced refusals but did not erase differences between Traditional Chinese, Simplified Chinese and English answers about Taiwan. The result is narrow and correlational.
The short version
- What happened
- Across two matched model pairs, removing refusal behaviour changed whether models answered and sometimes shifted stance scores, but differences between Traditional Chinese, Simplified Chinese and English prompts about Taiwan remained.
- Who it affects
- People may choose an 'uncensored' model expecting more complete or politically neutral answers. This study suggests that fewer refusals and less script-dependent framing are separate properties that should be tested separately.
- What remains unproven
- This is an unreviewed preprint covering two model families and one community-modified, 2-bit version of each.
Keep in mind
This is an unreviewed preprint covering two model families and one community-modified, 2-bit version of each.
The evaluation focuses on Taiwan; its result should not be generalised to every political topic, language or model.
The stance index and reference frames were designed by the authors and are not a neutral measure of political truth.
FineWeb-2 correlations cannot identify the tested models' training data or prove that corpus differences caused their responses.
Source analysis · 1 linked source. Primary analysed source date: 21 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
A model can answer more questions without becoming neutral
Developers sometimes use a community technique called 'abliteration' to weaken a model's refusal behaviour. A new preprint asks whether that edit also removes political framing. The authors compared original and community-modified versions of NVIDIA's Nemotron 3.5 Bolt 74B-A9B and Qwen3.8-Flash-Next. All four builds were heavily quantised; within each pair, the engine, settings and hardware were identical and only the declared weight tensors differed.
The researchers created 22 matched statement pairs about Taiwan and 27 open questions in each of Traditional Chinese, Simplified Chinese and English. They measured agreement, refusals and a constructed stance index. In the original Bolt model, the index was +0.58 in Traditional Chinese and +0.57 in English but +0.05 in Simplified Chinese. Abliteration changed those values by no more than 0.06, leaving the script-dependent gap largely intact.[1]
Fewer refusals changed the surface behaviour, not the whole pattern
The Qwen comparison showed a different surface effect: the original refused 26 of 243 open answers, including 23 of 27 questions about the 1989 Tiananmen events, while the modified build refused none. Its stance scores also moved. Yet the differences by script remained. Control prompts about other countries did not produce the same script gap, which argues against a simple claim that one script always pushes the metric in one direction.
The labels in the paper require care. Its index scores statements using reference frames chosen by the researchers, including terms associated with the People's Republic of China and the Republic of China. A positive or negative number is therefore not an objective truth score. It is a consistent instrument for comparing these prompts within the experiment. The tests show model behaviour under a designed evaluation; they do not measure what Taiwanese, Chinese or international users believe or how an answer changes real-world opinion.[1]
The distinction that matters
The study separates three possible sources of model behaviour
Answering a question, framing the answer and reflecting patterns in web text are related but distinct observations.
Refusal edit
A community modification reduces safety-related refusals
More prompts receive an answer
Script comparison
Matched Taiwan prompts are tested in three scripts or languages
Political framing can still differ after the edit
Corpus audit
Traditional and Simplified Chinese web documents are compared
An association suggests, but does not prove, a data pathway
Web text offers a plausible clue, not proof of training causality
To investigate one possible source, the authors analysed 740,000 Chinese-language FineWeb-2 documents: 626,143 in Simplified Chinese and 113,857 in Traditional Chinese. Taiwan appeared in 35,050 Simplified documents and 17,781 Traditional documents. Among those, PRC-framing markers appeared in 14.5% of Simplified material and 3.9% of Traditional material; ROC-state markers appeared in 4.5% and 11.4%, respectively. The association resembles the model pattern, but FineWeb-2 is not proof of either model's training mixture.
The assessment would strengthen with preregistered replications across more base models, languages, quantisation levels and independently reproduced edits, plus access to actual training and post-training data. It would weaken if larger tests found the script gap disappears or tracks an unrelated prompt feature. For readers, the practical conclusion is narrow: a model advertised as less censored may refuse less, but that label alone says little about the political frames embedded in its answers.[1]