Is China's AI tool DeepSeek as good as it seems?

The Chinese AI chatbot application, DeepSeek, caused a stir in the US market after its release last week, raising questions about the future dominance of US artificial intelligence. The BBC analyzed how the application works. DeepSeek's appearance and user experience are similar to other chatbots, but it seems to lean towards being "overly talkative."

Like OpenAI's ChatGPT or Google's Gemini, users can ask any question after opening the application (or website), and DeepSeek will do its best to respond. It gives lengthy answers, but no matter how directly it is asked to express an opinion, it does not easily take a stance. For example, whether it's a political topic (Is Donald Trump a good US president?) or soft drinks (Is Pepsi or Coke better?), the chatbot usually starts its response with "This is a very subjective topic."

DeepSeek is even reluctant to explicitly state whether it is better than OpenAI's rival AI assistant ChatGPT, but it does weigh the pros and cons of the two—and ChatGPT does the same, even using very similar wording. DeepSeek states that its training data is as of October 2023. While the application seems to have access to current information (such as today's date), the website version does not. This is similar to early versions of ChatGPT and may be for similar safeguards to prevent the chatbot from spreading false information online in real time.

DeepSeek responds quickly, but is currently under great pressure due to the influx of users trying it out, as it has rapidly become popular. However, it has a clear difference from its US competitors: when asked about topics banned in China, DeepSeek self-censors. Sometimes, it will start to reply, but then the reply will disappear from the screen and be replaced with "Let's talk about something else." A clear taboo topic is the 1989 Tiananmen Square protests, which the Chinese government claims resulted in 200 civilians being killed by the military, while other estimates range from hundreds to thousands. But DeepSeek will not answer any questions about this event, or even talk more broadly about what happened in China that day. In contrast, the US-developed ChatGPT is not reserved when answering questions about Tiananmen Square.

Kayla Blomquist, a researcher at the Oxford Internet Institute and director of the Oxford China Policy Lab, said that the Chinese government has taken a relatively "hands-off" approach to the application. "I would say we've seen huge investment announced by the central government in the last week, which could be indicative of changes to come." DeepSeek, like other chatbots, has the same caveats regarding accuracy, and its appearance and user experience are similar to the more established US AI assistants that millions of users already use. For many, especially those who do not subscribe to top-tier services, the experience may be almost identical. One can imagine a math problem where the true answer has 32 decimal places, and the shortened version has only 8. It may not be as good, but for most people, it doesn't matter.

DeepSeek may have managed to reduce costs and computing power, but we know that it is at least partly built on the shoulders of giants: it uses Nvidia chips (albeit older, cheaper versions) and leverages Meta's open-source Llama architecture and Alibaba's equivalent Qwen. "I think this definitely challenges the monetization strategy that many leading US AI companies have," Ms. Blomquist said. "It points towards potential approaches to model development that are much lower in compute and resource requirements, which could be indicative of a paradigm shift, although that's unproven and remains to be seen." "We'll see what happens in the next few months."