DeepSeek's political censorship doesn't stay confined to direct questions about sensitive topics. It surfaces in code reviews too.
A documented case shows a developer asking DeepSeek to improve their code. The model completed the task - and when the code or context touched on the 1989 Tiananmen Square massacre, DeepSeek told the user nothing happened there. The censorship wasn't triggered by a political question. It was triggered by content that appeared incidentally during a technical task.
How the Censorship Works
DeepSeek is built in China and operates under laws requiring AI companies to block content the government considers politically sensitive. The 1989 Tiananmen Square protests and their violent suppression are among the most heavily restricted topics. Any reference to the events, the date June 4th, or related context typically produces a denial or deflection from DeepSeek models.
The restriction is baked into the model's weights - the core numerical parameters that determine how the model generates responses - not just layered on as an API filter. Running DeepSeek locally through tools like Ollama doesn't remove it. The model produces the same altered output regardless of how it's accessed.
The Practical Problem for Developers
DeepSeek V3 and R1 are capable models. On coding benchmarks they compete with top western alternatives, and running them locally costs nothing. That combination has made them popular with developers who want strong code assistance without paying for API access.
But capable doesn't mean neutral. When DeepSeek is reviewing code, writing documentation, or summarizing content that happens to include politically censored topics, it may alter the output without flagging the change as censorship. From the developer's side, they see improved output. They don't see that a historical fact was silently rewritten.
For most code work - writing functions, debugging, refactoring - this doesn't matter. For anyone building applications that involve historical data, educational content, or documentation where factual accuracy matters, it's a real risk. The model won't tell you when it's making a politically motivated edit.
DeepSeek is a capable technical tool operating under constraints that most western developers aren't aware of until they run into them. Any work involving historical content, geopolitical context, or incidental references to politically sensitive events is likely to produce altered output.