An Analysis of Chinese Censorship Bias in LLMs
Bibliographic record
Abstract
When a large language model (LLM) has been trained on text featuring social biases, those biases implicitly impact the outputs of the model. Training an LLM on sanitized content, i.e., those pieces of content which remain after being subjected to state censorship (including alterations, deletions, and self-imposed censorship), results in what we term censorship bias. A model impacted by censorship bias may be less likely to reflect views that are routinely prohibited and more likely to reflect views that are not. This may particularly be an issue when interfacing with a model in a language that is predominantly used in a region with strong censorship laws. In this work, we outline what censorship bias is, introduce a novel methodology for identifying and measuring it, and apply that methodology to evaluate the most popular current LLMs. As part of the contributions of this work we designed and evaluated CensorshipDetector, a Chinese language text classification model which we use as part of our experimental design. Our evaluation of CensorshipDetector found it to be 91% accurate at differentiating between sanitized content and non-sanitized content. Our testing revealed evidence of censorship bias across all of the models we evaluated. Finally, we outline the potential harms of censorship bias, namely the exportation of information manipulation that would have primarily harmed a domestic audience to diaspora, as well as recommendations to various stakeholders to limit the harms of censorship bias and prevent it in the future.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.018 | 0.073 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".