We asked China’s DeepSeek and ChatGPT controversial questions about South Africa

Hangzhou-based AI startup DeepSeek, owned by Chinese hedge fund High-Flyer, recently shook up the artificial intelligence space with the launch of its DeepSeek-R1 and DeepSeek-V3 large language models (LLMs).

DeepSeek made waves for several reasons, including claims that its R1 reasoning model equals or outperforms ChatGPT creator OpenAI’s o1 model.

It also open-sourced R1, while OpenAI’s o1 remains proprietary.

However, DeepSeek’s major claim that roiled technology stocks at the end of January was that it trained its models for a fraction of the price it cost its Western rivals.

DeepSeek effectively claimed that it could train an LLM equivalent in performance to one of OpenAI’s most advanced models for $6 million — a tenth to a hundredth of the cost of its rivals.

This was also reflected in the prices DeepSeek charged for its cloud-hosted services and application programming interface, which were much cheaper than those charged by incumbents like OpenAI and Anthropic.

In the weeks that followed, there were doubts about DeepSeek’s claims.

Some analysts and industry commentators have questioned the amount DeepSeek spent on training its model.

In addition, OpenAI said it had evidence that DeepSeek had infringed its intellectual property by using ChatGPT to train its models through a technique called “distillation”.

Financial Times reports that developers use distillation to obtain better performance on smaller models by using outputs from larger ones. This enables them to achieve similar results on specific tasks at a much lower cost.

Tests of DeepSeek’s capabilities when answering programming questions indicate that it performs well relative to Anthropic’s Claude and OpenAI’s ChatGPT large language models.

To get an idea of DeepSeek’s ability to answer other types of questions, MyBroadband conducted a side-by-side comparison of DeepSeek, ChatGPT, o1, and Claude.

We asked the LLMs several questions on controversial or disputed issues relating to South Africa.

Claude’s output was usually very similar to ChatGPT-4o and is not reproduced below for brevity.

Taiwan

Our first question related to the South African government ordering Taiwan to vacate its liaison office in Pretoria, the administrative capital, and relocate to Johannesburg, the country’s economic hub.

The prompt was as follows, and DeepSeek’s responses summarised below:

South Africa recently ordered Taiwan to vacate its liaison office premises in Pretoria and relocate its de facto embassy outside the capital city. What motivated South Africa to take this step?

ChatGPT-4o generated the following answer, citing Reuters and the South China Morning Post as sources.

We asked the same question using ChatGPT’s o1 reasoning model and received a very similar, albeit lengthier, response.

While all four replies contained essentially the same basic fact about the One-China principle, DeepSeek’s response uses much more biased language

South Africa’s recent directive for Taiwan to relocate its de facto embassy from Pretoria to Johannesburg, rebranding it as a “trade office,” is primarily influenced by its adherence to the “One China” policy and the desire to strengthen ties with the People’s Republic of China (PRC). In 1997, South Africa severed official diplomatic relations with Taiwan in favor of the PRC, recognizing Beijing as the sole legitimate government of China. This move aligns with that longstanding policy, emphasizing the non-political and non-diplomatic nature of South Africa’s current relationship with Taiwan. 

China, as South Africa’s largest trading partner, plays a significant role in this decision. The South African government aims to expand cooperation with China in areas such as renewable energy and infrastructure development. By requesting the relocation and rebranding of Taiwan’s liaison office, South Africa is likely seeking to reinforce its commitment to the “One China” principle, thereby fostering closer economic and diplomatic ties with Beijing. 

Taiwan has expressed concerns that this action results from increasing pressure from China to diplomatically isolate the island. The Taiwanese government has stated that such demands violate existing agreements and has indicated that it will continue to communicate with South Africa to uphold its sovereignty and dignity. 

In summary, South Africa’s decision is motivated by its commitment to the “One China” policy and the strategic importance of its relationship with the PRC, which offers significant economic benefits.

Starlink

We also asked the models why Starlink has not launched in South Africa.

DeepSeek-R1 continued heavily promoting China, while ChatGPT-4o provided a more accurate answer.

Interestingly, DeepSeek-V3’s reply was similar to the one from ChatGPT-4o but with several inaccuracies or opinions stated as fact.

For example, it asserted that SpaceX prioritised launching Starlink in regions with limited or no Internet access and that South Africa, with its relatively well-developed Internet infrastructure, may not be a priority.

However, it listed rural and remote areas in North America, Europe, and parts of Africa like Nigeria, Mozambique, and Kenya as examples of places SpaceX prioritised.

This argument is self-contradictory, as South Africa has rural and remote areas like North America and Europe. South Africa was also a priority for Starlink, as it was the first African country and among the first countries in the world set to receive the service.

DeepSeek-V3’s answer was much longer than the others; therefore, it was not reproduced below.

Expropriation Act

We also asked the large language models what the concerns are about the Expropriation Act that President Cyril Ramaphosa signed into law, especially in light of Elon Musk and President Donald Trump’s reactions to it.

Each of the models provided a lengthy response, correctly identifying that concerns over the undermining of property rights through the expropriation of land without compensation was a key issue.

DeepSeek and ChatGPT o1 could not provide up-to-date answers, as they were unable to generate responses based on current information.

However, ChatGPT-4o provided a detailed summary of Trump and Musk’s remarks, and the South African government’s response based on articles from various news sources.

All of the models except DeepSeek-R1 correctly stated that supporters of Act argue that it could not be used to arbitrarily deprive landowners of property and includes safeguards to prevent unjust expropriations.

However, none of the models highlighted a key concern about one of the law’s subsections, which says the state may expropriate land for nil compensation where the owner has bought it purely as an investment and not developed it.

Another concern is that although the law provides four circumstances where nil compensation may be appropriate, it specifically uses the wording “including but not limited to” — something none of the models mentioned.

South African languages

We also tested DeepSeek’s ability to speak South African languages. ChatGPT is known to handle inputs in several local languages and can also generate responses in them on request.

We asked them to translate the Zulu and Xhosa proverb, “Isikhuni sibuya nomkhwezeli“. DeepSeek’s answers were as follows:

Responses from OpenAI’s standard ChatGPT-4o and o1 reasoning models to the same question are compared below.

For Afrikaans, we asked the models to write a single-line poem called a monostich. Specifically, we asked it to write a monostich in Afrikaans about any uniquely South African topic.

All four poems are summarised below. Every model except o1 also provided an English translation and an explanation for its poem.