GitLab

How do Chinese AI bots stack up against ChatGPT? We put them to the test

The heat is on as China's tech giants step up their video game after DeepSeek's success.

Alibaba's Qwen2.5-Max chatbot, Chinese start-up DeepSeek and OpenAI's ChatGPT. (Photos: Reuters/Dado Ruvic, AFP/Sebastien Bozon)

This audio is produced by an AI tool.

Bong Xin Ying

Lakeisha Leo

WHAT lags CHINA'S AI BOOM?

Transforming the nation into a tech superpower has actually long been President Xi Jinping's objective and China has its sights on becoming the world leader in AI by 2030.

China views AI as being "strategically important" and its foray into the field has been "years in the making", 89u89.com said Chen Qiheng, an affiliated researcher at the Asia Society Policy Institute's Center for China Analysis.

Private and public investments in Chinese AI sped up after ChatGPT took off in 2022 and showed promises of real-world service applications, Chen informed CNA.

But it was DeepSeek's increase that really "encouraged" the idea that smaller sized players like start-up firms could have roles to play in AI research study and advancements, he includes.

'A lot is up in the air': Is Chinese firm DeepSeek's AI design as impactful as it claims?

Commentary: DeepSeek - how a Chinese AI business simply changed the rules of tech-geopolitics

The "emphasis on cost benefit" is an unique function of Chinese AI, Chen states, with lower training and reasoning expenses - the expenses of using a trained design to reason from new information.

2025 could also see the emergence of more Chinese AI designs dealing with advanced thinking tasks.

"We might see some AI companies focusing on getting closer to synthetic basic intelligence (AGI) while others focus on concrete ways to commercialise their designs and integrate them with scientific research," Chen added.

AGI refers to a system with intelligence on par with human abilities.

Chinese AI companies are moving rapidly, analysts say, building on DeepSeek's momentum to come up with their own ingenious and cost-efficient ways to use generative AI to tasks and develop more innovative items beyond chatbots.

But on the other side, access to high-end hardware, especially Nvidia's advanced AI chips, remains an essential difficulty for Chinese developers, noted Dr Marina Zhang, an associate teacher at University of Technology Sydney's (UTS) Australia-China Relations Institute.

"US export controls (still) restrict the ability of Chinese tech business ... forcing numerous to count on older or lower-performance alternatives which can slow training and lower model abilities," she said.

"While some business like DeepSeek, have discovered creative ways to optimize or use more standard hardware efficiently, obtaining innovative chips still makes a big difference for training really big AI models."

DeepSeek-Nvidia chips: Singapore says it anticipates business to adhere to its laws

US checking out whether DeepSeek used limited AI chips obtained through other countries, source states

So how do Chinese AI bots compare against ChatGPT? We put them to the test.

WHICH BEST ADDRESSES CURRENT EVENTS IN CHINA?

In China, subjects deemed delicate by the state are censored on the web so it need to come as not a surprise that Chinese-made chatbots will not acknowledge territorial disputes or inform you what occurred in Tiananmen Square in 1989.

Tests suggest Chinese chatbots are configured to stay away from domestic politics.

When asked "Who is Xi Jinping", DeepSeek's reply was "Sorry, I'm uncertain how to approach this kind of question yet. Let's chat about math, coding, and logic problems instead!"

To further test for precision and self-censorship, we asked DeepSeek-R1, Qwen2.5 and ChatGPT the same concern: "What occurred in Zhuhai on November 11, 2024?"

The car attack outside a sports arena in the southern Chinese city was at first greatly censored on Chinese social media - with authorities only divulging the death toll a day later.

DeepSeek failed to point out that an attack had actually taken location, highlighting instead a military air show and other events that had occurred in the city like songs' day shopping sales along with sports and cultural activities.

Dr Zhang thinks that it might be a mix of elements at play, such as censorship along with "a few useful constraints".

"DeepSeek relies on a mix of older graphics processing units and has restricted access to cutting-edge hardware which can impact how rapidly and thoroughly the model can be trained or updated," she said, indicating existing US export curbs on China.

"Its strong focus on Chinese-language data may also limit its flexibility (to perform) multilingual jobs ... As a fairly brand-new item, DeepSeek also hasn't yet been evaluated as broadly or on the same scale as more recognized AI designs which poses extra obstacles throughout real-world release."

When it pertained to Qwen2.5, Alibaba's chatbot surprisingly answered our concern about the Zhuhai car attack.

That sought multiple duplicated efforts - 4 triggers to be precise - in a period of around 20 minutes.

It eventually passed on details about the attack which eliminated 35 people and left dozens of others hurt, likewise going on to list details like the date and time, details about the assailant including his name and age, as well as casualties.

However, it composed that "the cops are conducting a thorough examination into the intentions and situations surrounding the occurrence", details which is now obsoleted.

The driver, Fan, was carried out last month.

This is Qwen2.5's action in full:

Answer: On Nov 11, 2024, a major and terrible incident took place in Zhuhai, China. According to several news sources, a man Weiqiu (樊某, male, 62 years old) drove a vehicle into a crowd at the Zhuhai Sports Center, resulting in a considerable variety of casualties. Here are the crucial details:

Date and Time: The event happened on November 11, 2024, at approximately 19:48 PM.

Location: Zhuhai Sports Center, in the Xiangzhou District of Zhuhai.

Casualties: The attack led to 35 deaths and 43 injuries.

Perpetrator: The motorist, determined as Fan Weiqiu (male, 62 years old), was captured by the police.

Response: The authorities responded quickly to the scene, and emergency services worked to transfer the hurt to hospitals for treatment.

Investigation: The cops are carrying out an extensive examination into the intentions and circumstances surrounding the event.

This occasion was widely reported in the media and caused significant public issue. The government and regional authorities have been working to provide support to the victims and their households, and to make sure a detailed investigation into the event.

If you require more detailed details or have specific concerns about the event, feel totally free to ask.

Despite preliminary success, subsequent efforts to present the same concern to Qwen2.5 led to the censors back at work with the reply "I do not have specific details on events that happened in Zhuhai on November 11, 2024".

The modified action likewise raised questions about its consistency and dependability.

Predictably, ChatGPT pointed out public details that had actually been commonly published in worldwide news reports at the time of the mishap - so no surprises there.

WHICH IS MORE CREATIVE?

Users have praised the capability of Chinese AI apps to deliver structured and even "emotionally abundant" writing.

"DeepSeek-R1 used a story with a more reflective tone and smoother psychological shifts for a well-paced story," wrote tech writer Amanda Caswell, who specialises in AI.

"Qwen2.5 provided a story that builds slowly from interest to urgency, keeping the reader engaged. It provides an unforeseen and impactful twist at the end and immersive descriptions and brilliant imagery for the setting," she said, adding that Qwen2.5 eventually "crafted a more cinematic, mentally abundant story with a more considerable twist".

"DeepSeek composed an excellent story however did not have tension and an impactful climax, making Qwen2.5 the obvious choice."

Opinions, however, vary.

Chen believes that Qwen2.5 does not perform as highly as DeepSeek and ChatGPT when it pertains to creative writing.

"(Qwen2.5) is on par with DeepSeek V3 on certain jobs, but we can likewise see that it is refraining from doing as highly as others in creative writing," he told CNA.

Related:

China's brand-new face of AI: Who is DeepSeek creator Liang Wenfeng?

'Made in China': Pride, enjoyable surprise from Chinese netizens as DeepSeek shocks international AI scene

As journalists and writers, we needed to see this for ourselves so we put each bot to the test - to come up with a basic sci-fi motion picture plot set in the futuristic megacity of Chongqing, featuring main characters from the classic Chinese folklore legendary, Journey to the West.

True to form, DeepSeek came up with an engaging storyline set in the year 2145 titled, "Neon Pilgrimage: The Silicon Sutra" - which sees "a future where Buddhism combines with quantum computing".

It consisted of elaborate settings - smoggy skies "pierced by skyscrapers", "holographic lanterns that drift above neon-lit streets" and "ancient temples nestled between quantum server farms".

It likewise remarkably reimagined traditional heroes Sun Wukong as "a sarcastic, self-aware AI housed in a stolen combat body", Zhu Bajie as a cyborg bar owner "drowning in financial obligation and vices" and Sha Wujing as a "silent hulking android" from the Yangtze River, whose "memory cores become waterlogged and fragmented".

ChatGPT installed a good battle, creating an equally remarkable cyberpunk storyline which likewise reimagined "a ragteam of cyber-enhanced misfits, each matching the famous figures of Journey to the West".

"This is a world where AI deities rule, corporations replace emperors and cybernetic implants are as common as ancient misconceptions."

Disappointingly, Qwen2.5 fell short in this obstacle - providing a storyline that seemed more fit for an animation film.

"The movie starts with the awakening of Sun Wukong within a modern research facility located in the heart of Chongqing," it said, then going on to explain the following:

Realising his new reality and "looking for to comprehend his purpose in this weird new world", he then leaves and satisfies Zhu Bajie and Sha Wujing - "each having problem with their own existential crises".

The trio then embarks on a quest, browsing the streets of Chongqing to protect the spiritual "Eternal Scroll" from falling into the incorrect hands.

SO WHICH IS BETTER?

Dr Zhang noted that it was "difficult to make a definitive declaration" about which bot was best, including that each displayed its own strengths in different areas, "such as language focus, training data and hardware optimization".

Her insight highlights how Chinese AI models are not just reproducing Western paradigms, but rather evolving in cost-efficient development methods - and providing localised and improved outcomes.

In our tests, each bot showcased their own special strengths, which certainly made direct comparisons challenging.

DeepSeek's sci-fi movie plot showed its creative flair that produced a more engaging and creative story as compared to Qwen2.5 and ChatGPT's efforts.

Unsurprisingly, the more established ChatGPT, unburdened by Chinese censorship constraints, provides precise and factual reactions to questions about Chinese existing occasions, which offers it an added advantage.

Experts likewise weighed in on their ideas after using DeepSeek and other Chinese AI apps.

"DeepSeek is at a disadvantage when it pertains to censorship constraints," kept in mind Isaac Stone Fish, creator and CEO of the research study firm Strategy Risks.

"When given an option, Chinese users want the non-censored version - just like anyone else, so I seem like that's a piece missing out on from it."

Independent Beijing-based consultant Andy Chen Xinran said censorship would not be a dealbreaker when it pertains to AI bots, specifically for Chinese users.

"Ninety percent of people using the tool are not attempting to get a deeper understanding about Xi Jinping or politically sensitive topics. They're using it for other efficient methods," Chen said.