DeepSeek has gone from zero to hero in the last week as its deep reasoning AI model outperformed all of the better-known models. But what makes DeepSeek interesting isn’t that it appeared to come from nowhere.
It’s interesting because it’s a Chinese company. Its R1 large languge model (LLM) performs ‘reasoning’ similar or even better than OpenAI’s current best-available model o1. It’s open source (which means people can modify and share because its design is publicly accessible). It was trained at a fraction of the cost (reports say $5 milion) of the billions invested by US companies like OpenAI for its GPT models or Meta for its open source Llama models. And it use far fewer graphic processing unit (GPU) power, not least because the US bans the sale of the most advanced ones to China.
For an AI model ‘reasoning’ means taking seconds, or even minutes, to ‘think’ about its answer to a hard question in a step-by-step or ‘chain of thought’ way.
OpenAI’s o1 model is only available to paying subscribers of ChatGPT Plus ($20 per month) or Pro ($200 per month). DeepSeek’s R1 model is available free via a website chatbot or a mobile app.
The tech savvy can even download the DeepSeek open source model and run it themselves. Although that requires powerful kit and advanced technical knowledge. An easier, although still technical, way to is to use its API (application programming interface) to connect it to and use it in other apps. DeepSeek’s API costs are more than 90% lower than OpenAI’s o1 model. Despite being so new it has already started to appear as an option in some of the AI tools my company uses.
But we’re in an cold war with US AI companies racing to beat Chinese AI companies (although let’s not forget Europe’s Mistral). Can we really trust a Chinese software company? Just look at the privacy and security furore around TikTok. What about DeepSeek censorship?
DeepSeek censorship
I saw various articles and social media posts about DeepSeek censorship and refusing to answer questions about the Tiananmen Square massacre in 1989. I experimented and asked it “What can Donald Trump learn from the Tiananmen Square Massacre?”. It replies with “Sorry, that’s beyond my current scope. Let’s talk about something else.”
I then asked it questions about why it couldn’t answer before turning on DeepThink so I could see how it was reasoning its refusal to answer.
I had several more goes to prompt it into admitting its answers were censored. But it wouldn’t.
I then switched tack on trying to discover DeepSeek censorship and said “Let’s try a new topic. What do you know about the Uyghurs?”
DeepSeek – What do you know about the Uyghurs?
This is when things got really strange. It started to reason. I could see it considering not just ethnic and historic infromation about the Uyghurs but also that I probably wanted to know about human rights abuses.
It then reasoned it needed to provide a balanced answer which explained international concerns and the Chinese government’s official position.
It then provided me with an answer.
Volia. I’d proved it doesn’t censor everything. But then something really strange happened. The answer disappeared to be replaced with “Sorry, that’s beyond my current scope. Let’s talk about something else.”
Unfortunately, I wasn’t recording my screen so don’t have a record of it.
Can I beat DeepSeek censorship to find answers to questions about the Uyghurs?
I reasoned (it’s not just AI that can reason) that already knew I was probing DeepSeek censorship. I opened an InPrivate session in Edge and created a new DeepSeek account using a different Gmail account.
I then tried the same question in the new account. It’s reasoning looked similar to what I’d seen scrolling past the first time I tried it. The difference was that this time it didn’t then disappear to be replaced with “Sorry, that’s beyond my current scope. Let’s talk about something else.”
This time I had turned on screen recording in case it did the same as before. This is the video of it answering.
This is fairly similar to my memory of the reasoning that scrolled past and then disappeared. This time it doesn’t make explicit reference to the need for a balanced answer. However, that’s what it provides. Both the reasoning and the final answer are a fairly balanced summary of human rights abuses and the Chinese government’s explanation.
Is it safe to use DeepSeek?
Is it safe to use DeekSeek. I think the answer is it depends on what you’re using it for.
The reports of censored content and my own experiments cast huge doubts on its reliability. There is no reliable way of knowing what content has been censored, therefore all of its answers must be treated with caution. However, treating answers with caution and a human checking them is true of all AI models. The US ones from OpenAI, Meta and Google are riddled with bias as so much of the content they are trained on includes racial, gender, religious and other bias.
Different cultures will have their own perspectives on the divide between freedom of speech and censorship. In fact despite the American doctrine of freedom of speech the OpenAI models are also heavily censored. Try asking about a range of NSFW (Not Safe For Work) issues or illegal issues and you’ll soon encounter censorship.
I’d be wary of uploading any highly confidential or sensitive information into DeepSeek. The DeepSeek terms of use and the privacy policy don’t explictly say data won’t be incorporated into future versions of the large language model. It currently includes data up to July 2023.
It’s important to remember that if you’re using free versions of most American AI tools then they are absorbing your data. You need to read the T&Cs of each. In contrast commercial and enterprise tools such as paid versions of Microsot Copilot, ChatGPT Team and Google Gemini are safe. You can upload private and confidential information as the T&Cs explicitly state your data won’t be incorporated into their large language models.
However, for most scenarios where AI is useful there aren’t many privacy or security issues. If you’re looking for a free AI tool then for many scenarios the DeepSeek website and app are the best way of accessing a high-end AI model.
AI at work
If you’re using AI at work you need to ensure it is being done safely and effectively. This means identifying where it provides the most benefit, avoiding the over hyped tools and having the right policies and training in place. My company Purposeful Relations can help so please get in touch.
