Bias introduced in every step of the way.

Feb 3rd, 2025

TL;DR

Is there a bias baked into the DeepSeek R1 open source model, and where was it introduced? We found out quite quickly: Yes, and everywhere. The open source DeepSeek R1 openly spouts pro-CCP talking points for many topics, including sentences like “Currently, under the leadership of the Communist Party of China, our motherland is unwaveringly advancing the great cause of national reunification.”

We ran the full 671 billion parameter models on GPU servers and asked them a series of questions. Comparing the outputs from DeepSeek-V3 and DeepSeek-R1, we have conclusive evidence that Chinese Communist Party (CCP) propaganda is baked into both the base model’s training data and the reinforcement learning process that produced R1.

Context: What’s R1?

DeepSeek-R1 is a chain of thought (or reasoning) model, usually accessed via DeepSeek’s official website and mobile apps. It has a chat interface like OpenAI and Anthropic. It first “thinks out loud” step by step in an initial area tagged <think>, and then it gives its final answer. Users find both the reasoning and the final answer useful.

This is an opinion we’ve seen a lot online:

Screenshot 2025-02-01 at 9.07.00 AM.png

Other common misconceptions we’ve seen:

With over a decade of experience building NLP and ML products, it is important to us to cut through the misinformation and get to the truth.

At Plum AI, we’ve built products that evaluate LLMs and leverage those evaluations to fine-tune and distill stronger models.

We used an 8xH200 machine, which has over 100x the GPU power of a desktop PC, to run the open source DeepSeek models and perform a structured analysis.

Plum AI’s analysis