September 18, 2026
Free China Movement has published a new black-box research audit examining how DeepSeek responded to politically sensitive Chinese historical material under controlled prompt variations.
The report, From Copying to History: A Black-Box Audit of Politically Sensitive Response Restrictions in DeepSeek, tests simple operations such as copying, counting, translation, classification, definition and historical retrieval using material related to Tiananmen Mothers, the 1989 Beijing student movement and Lian Shengde.
The report finds behavior consistent with content-sensitive response restriction in the tested interface. In some trials, mechanically simple operations involving supplied sensitive text were refused while nearby control strings were processed. At the same time, the audit explicitly does not claim access to DeepSeek’s system prompts, training data, classifiers or internal moderation architecture, and it does not claim that any single person or keyword is categorically blocked.
The report also introduces a sensitive-information handling protocol separating four layers that are often conflated in public debate: training data, retrieval access, moderation or policy layers, and final model output.
Read the full report: https://freechina.live/english/deepseek-black-box-censorship-audit-politically-sensitive-response-restrictions-2026/
Why this matters
Black-box testing cannot reveal an AI provider’s internal implementation by itself. It can, however, document reproducible differences in observed behavior. The report therefore emphasizes repeatable testing, preserved prompts and screenshots, exact model/version information and explicit treatment of competing explanations.
Related external context
Recent independent reporting and research have continued to scrutinize politically sensitive output behavior in Chinese AI systems. Separately, Anthropic reported in September 2026 that it observed DeepSeek relaying some user exchanges to Claude without users’ knowledge, including cases involving sensitive data. That disclosure concerns data handling and model-routing practices and should not be treated as proof of the political-response mechanism tested in this report.
Sources: Free China Movement research report; Anthropic, Countering misuse of AI: September 2026; prior published DeepSeek censorship research cited in the report.