Questions censored by DeepSeek

Original Article ↗ Hacker News Discussion ↗

DeepSeek’s Chinese-made AI models are drawing scrutiny for refusing to answer politically sensitive queries — especially about events like the Tiananmen Square crackdown — while answering similar questions about other countries. Commenters contrast censorship baked into the base model and additional hosting-time filters with the “safety” layers and political biases of Western systems like ChatGPT and Claude, noting that almost all major LLMs now shape or block content according to their legal and cultural environments. Many see the real stakes not in any single incident, but in how increasingly capable, widely used models can embed state or corporate narratives into what users perceive as neutral, authoritative answers.

Nature and extent of DeepSeek censorship

  • Many commenters attribute DeepSeek’s behavior to Chinese legal requirements to uphold “Core Socialist Values” and avoid politically sensitive topics (e.g., Tiananmen, Taiwan, Uyghurs).
  • Hosted DeepSeek (especially R1 671B on deepseek.com and some US-hosted APIs) often gives stock refusals or CCP‑aligned framings on such prompts, while answering similar questions about other countries.
  • Several note that the censorship can be asymmetric: detailed criticism of the US is allowed where criticism of Chinese state actions is blocked.

Hosted vs local, and model confusion

  • Strong distinction between:
    • DeepSeek-R1 671B (original reasoning model, heavily censored),
    • “R1 Zero” (earlier, reportedly less aligned),
    • Distilled models (Llama/Qwen fine‑tuned on R1 outputs) used by Ollama, Groq, etc.
  • Distilled smaller models often show much weaker or no censorship on Chinese politics, leading to conflicting anecdotes from users who think they’re “running R1 locally” when they’re actually running a distilled Llama/Qwen.
  • Some report additional bolt‑on moderation on hosted services: partial answers appear, then are wiped and replaced with a generic refusal.

Technical implementation and jailbreaks

  • Debate over whether censorship is:
    • post‑hoc filtering of outputs,
    • explicit safety fine‑tuning (RLHF),
    • or implicit via censored training data.
      Evidence suggests all three exist across different Chinese models and hosting setups.
  • Users show simple jailbreaks (e.g., leetspeak / ROT13 / alternative encodings) that bypass keyword filters and elicit detailed Tiananmen descriptions.
  • Similar multi‑layer safety stacks and browser‑side output filters are described for ChatGPT and other US models.

Comparison with Western LLMs

  • Many argue Western models also censor heavily (weapons, self‑harm, “crime stats,” group‑targeted questions, some live political scandals) but frame it as “safety” or “harm reduction.”
  • Examples show uneven treatment depending on country, religion, or person, and non‑deterministic refusals.
  • Some see Chinese censorship as more overt and state-driven; Western censorship as subtler, corporatized, and still influenced by governments and powerful individuals.

How much this matters

  • Split views:
    • Some only care about coding/technical tasks and see political censorship as irrelevant.
    • Others worry that people increasingly use LLMs instead of search, so embedded propaganda or omitted history is socially dangerous.
  • Several call for symmetric audits: similar prompt‑refusal datasets for ChatGPT, Gemini, Grok, etc., not just DeepSeek.