AI Pulse · AI Trends Pulse

A developer published a hands-on distillation study that tested whether DeepSeek’s known censorship filters carry over when you use its outputs to train another open-weight model. The short answer, according to the repo and discussion, is they don’t.

Distillation is when you take a larger or proprietary model’s responses and use them to train a smaller, open model. It’s a common technique for teams that want performance without the cost or licensing baggage of closed APIs. The worry has always been that if the teacher model refuses certain prompts or bakes in alignment guardrails, the student model inherits those same refusals. This test suggests that’s not automatic. When the researcher distilled DeepSeek outputs into GPT-OSS, the resulting model answered prompts DeepSeek would have blocked.

Why it matters: if you’re building an internal assistant or a customer-facing tool on an open-weight stack, you need to know what you’re inheriting. A model trained on filtered outputs might seem safe until it suddenly refuses a legitimate business query because the upstream model flagged a keyword. This study is narrow, one setup, one pair of models, but it’s testable and it points to a practical reality. Censorship and alignment are behaviours, not always baked into the weights themselves. That gives you more control, but it also means you can’t assume safety or policy compliance transfers automatically either.

If you’re stitching together open models, fine-tuning on distilled data, or just trying to understand what’s actually inside the tools you deploy, this is the kind of technical detail that belongs in your AI command centre. You want visibility into what each model does, where the outputs come from, and what filters are live or dormant. The repo is public, the test is reproducible, and the lesson is clear: behaviour doesn’t always follow the data.