China’s DeepSeek is not for Aussies (Consumer Advice)

China’s DeepSeek Artificial intelligence app wiped trillions off the US Share Market by allegedly releasing a new way to do AI—faster, cheaper, better, and without requiring NVIDIA’s expensive Neural Processor Units.

That is a hell of an impact for a Chinese company registered in 2023 that released its DeepSeek A.I. Assistant on 27 January 2025.

Let’s not get too techy—you only need to know that AI is trained on large language models (LLMs). For example, the contents of an animal database allow it to identify different animals, and the contents of a medical database will enable it to identify specific cancers and ailments. This is phase one—machine learning. It then uses AI-programmed responses to react to the stimuli.

AI experts have long said that AI is only as useful as the LLMs on which it is trained. DeepSeek stated that its training cost US$6 million compared to $100m for OpenAI’s GPT-4 and required only a tenth of the computing power. Hence, NVIDIA’s share price crashed along with any other US AI company, including Microsoft, Google, Amazon, Meta, et al.

DeepSeek was initially trained on 97% Chinese LLMs

If that does not ring alarm bells, you can stop reading.

AI experts talk of bias and poisoning of LLMs – garbage-in and garbage-out.

US-based AI models are predominately from the Western world and are created by people and cultures from like-minded countries. LLMs use free, non-copyright material where possible or pay for the rights to use it for training. There is a conscious effort to reduce bias and poisoning.

Free LLMs include math, physics, programming/coding, chemistry, science, medicine, logic, encyclopedias (Wikipedia included), English and other languages, literary works, artworks, and almost all web content, including news sites, CyberShack and user-generated sites like Reddit or Quora.

To be clear, AI can only react to what it has learned – just as you can with what you have learned.

DeepSeek is relentlessly Chinese in origin.

When you use the DeepSeek A.I. Assistant, you largely use Chinese LLMs. Your data is stored there and subject to Articles 37 and 61 of the Chinese Cybersecurity Law.

Article 37: Critical information infrastructure operators that gather or produce personal information or important data during operations within the mainland territory of the People’s Republic of China shall store it within mainland China. Where, due to business requirements, it is truly necessary to provide it outside the mainland, they shall follow the measures jointly formulated by the State cybersecurity and informatisation departments and the relevant departments of the State Council to conduct a security assessment; where laws and administrative regulations provide otherwise, follow those provisions.

Article 61: This ensures data is handed over on demand or confiscation.

In short, the CCP can access whatever you do with DeepSeek.

What data does it collect?

  • Email address
  • Phone number
  • Date of birth (to create an account)
  • Location
  • Any user input, including text and audio, as well as chat histories
  • So-called “technical information” – from your computer/phone’s model and operating system to your IP address and keystroke patterns.
  • Suspected: Social media accounts and friends

DeepSeek has Chinese censorship and ideology safeguards

It has censorship mechanisms for topics considered politically sensitive for the CCP. For example, the model refuses to answer questions about the 1989 Tiananmen Square protests and massacre, persecution of Uyghurs, or human rights in China.

It has the core socialist values defined by the Chinese Internet Regulatory Authority.

Whether proven or not, AI experts state that DeepSeek could be used for foreign influence operations, spreading disinformation, surveillance and the development of cyberweapons for the government of the People’s Republic of China.

DeepSeek’s privacy terms and conditions state, “We store the information we collect in servers in the People’s Republic of China… We may collect your text or audio input, prompt, uploaded files, feedback, chat history, or other content you provide to our model and Services.”

CyberShack’s view: DeepSeek was too fast and too good to be true

AI experts are more than sceptical that DeepSeek is a front for more than an AI assistant. Its announcement wreaked havoc on US Share Markets. Whether it’s too good to be true (or not), its release was timed to embarrass President Trump’s AI announcements.

Australia’s science minister, Ed Husic, has become the first member of a Western government to raise privacy concerns about DeepSeek. ”I would be very concerned about data and privacy management.

According to CNBC, the US Navy has banned using DeepSeek’s apps altogether, citing “potential security and ethical concerns.”

Microsoft and OpenAI/ChatGPT suspect DeepSeek’s technology was illegally obtained. David Sacks, the White House’s AI czar, told Fox News,There is substantial evidence that what DeepSeek did is distilled the knowledge out of OpenAI’s models.”

Remember, AI is still an unregulated frontier.

CyberShack’s AI news and reviews

Brought to you by CyberShack.com.au