How DeepSeek Thinks And Censors

Everyone is talking about Deepseek, an AI model from China. I tried it, and boy, was it interesting.

A bit about Deepseek: It is free and it has been benchmarked to rival ChatGPT. It manages to achieve all these with far less energy and far fewer high-end chips. It is a huge achievement and many were very surprised, and even shaken by it.

And yes, it also comes with the uniqueness of being a Made in China application. While all AI models screen and filter information, whether they are made in the U.S. or France or whatever. Chinese apps are just clearer and more consistent in what they screen.

But what makes Deepseek truly interesting is that you actually get to see how it thinks and witness the moment it filters.

I asked it a generic question: What happened in Hong Kong in 2019. This is a question that could potentially lead to a sensitive topic: the protests.

The app showed me, in detailed text, what it was thinking before generating its final answer. The first topic it “thought” about in response to my generic question was the protests. It thought through the concerns surrounding the bill, the protests, the international repercussions, everything.

And then, when it was time to generate the answer, it jumped to the standard answer to sensitive topics “Sorry, I’m not sure how to approach this type of question yet.”

Just to be sure, I asked the same question to other popular AI models and they all picked the protests to answer the question.

The next thing I tried was to ask it to fix this writing. I copied and pasted this text into Deepseek and asked it to fix the grammar. This time, it gave me the standard “I’m not sure how to approach this type of question yet” answer, without going through the thinking process.

Just to be sure, I tried asking similarly generic but potentially sensitive questions, such as “what happened on January 6th,” and it went straight to the what happened to the United States Capitol Building in Washington, D.C., on January 6th, 2021, thought about it, and talked about this in the answer it provided. I also copied and pasted this text to other AI models and they were willing to fix my broken English.

I am sure they will soon change this and will not show users how it thinks through sensitive topics. But for now, it’s fun to poke around and watch how it thinks about tricky questions.


Posted

in

by

Tags:

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *