Chinese AI chatbot DeepSeek struggles with Tiananmen Square questions, but wipes out S$1.35 trillion from US stocks

DeepSeek, an AI company based in Hangzhou, China, took the tech world by storm on Jan. 27, 2025 when it released its free AI model to users around the world.

The app, similar to other AI models like ChatGPT, quickly became one of the most downloaded apps on Apple's app store in countries, such as Singapore and the UK.

Its debut spooked the world markets, wiping an astonishing US$1 trillion (S$1.36 trillion) from the U.S. tech index Nasdaq Composite.

How does it work?

If you've used ChatGPT by OpenAI, DeepSeek's interface will be familiar.

Create an account, sign in and ask the friendly blue whale logo any question you can think of.

The speed at which the platform responds appears similar to ChatGPT.

Why so scary?

Still, the emergence of a China-based competitor to ChatGPT may not appear at first glance to be capable of spooking the Western tech world to the tune of a trillion dollars.

But DeepSeek is different from other AI models.

For a start, The Guardian reported that DeepSeek is much cheaper to operate than others.

"This model uses a different kind of internal architecture that requires less memory use, thereby considerably reducing the computational costs of every search or interaction with the chatbot-style system."

According to CTech, "Operating trained AI models is also incredibly costly, with each query consuming enough energy to power a light bulb for several minutes. This contrasts with traditional search engines, which are far more energy-efficient."

However, DeepSeek is still able to take on "complex reasoning tasks" in areas like mathematics and coding and produce similar results to its competitors.

It does so with its "Reinforcement Learning" model, which mimics the human learning process of trial and error.

After an initial training phase, the model is put to work generating responses to various queries.

These are then evaluated by another AI model, and the best responses are selected.

This way, DeepSeek is able to refine its answers over time, and allow the model to "learn autonomously", while using less computing power.

The success of a Chinese company in producing such an efficient AI model despite sanctions on computer chip exports by the U.S. has led to heavy losses sustained by industry giants like Nvidia, which lost a record US$593 billion (S$802 billion) in a single day.

Limitations

But while the technical aspects of DeepSeek may place it ahead of its competitors, it still appears to be shackled by human political restraints.

As it turns out, the model is reluctant to provide answers to certain queries.

Asking the model what happened at Tiananmen Square on June 4, 1989 results in it "stalling" for a noticeably long time, much longer than it took to respond to other questions.

Then it returns an evasive answer:

Interestingly, for certain questions, the model at first returns a long and complex answer, giving lots of detail.

And then, after the answer is generated, it is completely wiped and another evasive response is given.

Intriguingly, the model does appear to be learning over time.

On Jan. 27, it returned this rather jingoistic answer to this question:

However on Jan. 28, the same question prompted a less jingoistic and more complex answer.

However, the answer is then wiped, and then this appeared:

As reported by Quartz, "Like all AI products developed in China, DeepSeek is required to adhere to the “socialist values” of the Chinese Communist Party. This means the chatbot must comply with strict government regulations, limiting discussions on sensitive topics, particularly those related to Chinese politics and governance."

Top image from DeepSeek.