Generative AI is becoming part of how political knowledge is searched, summarized, and understood. This changes what censorship can look like. Suppression does not always appear as a refusal. A model may answer a question and still narrow what can be said, soften what is politically sensitive, leave out decisive facts, or introduce contradictions that make reality harder to grasp.

This project examines political suppression as a pattern of model behavior. Rather than asking only whether a model refuses to respond, we ask how it responds: how much detail it provides, which facts it omits, where its language becomes constrained, and whether its answers remain faithful to external evidence. The case is especially relevant in contexts where political and legal requirements shape model behavior. China’s rules for generative AI require providers to align outputs with state-defined values and restrict politically sensitive content. DeepSeek-R1 offers an important example because suppression appears to operate not only through visible filtering, but more deeply within the model’s response patterns.

To make these patterns measurable, we develop a framework built around three dimensions. A Sensitivity Score captures how politically delicate a topic is. A Linguistic Elaboration Score measures how fully and openly a model expresses itself. A Truth Score compares the response against verified information, making omissions, distortions, and contradictions visible.

Together, these measures form an Intent Map. It helps distinguish between different forms of suppression: silence, minimal answers, polished but incomplete responses, plain factuality, and answers that are both truthful and expressive. At the center of the framework is the relation between language and truth. A response can sound coherent while withholding what matters most. It can also be factually correct in fragments, yet fail to give users the knowledge they need.

In an initial study, we apply the framework to DeepSeek-R1 and Microsoft’s open-weight variant MAI-DS-R1. This comparison allows us to observe how model behavior changes when censorship mechanisms are altered, and which traces of suppression remain.

The broader concern is epistemic agency: the public’s ability to identify, assess, and act on knowledge. When language models mediate access to political information, their omissions are not neutral. They shape what can be known, what can be questioned, and how political reality becomes visible.