China's self-censoring chatbots face many challenges

Chinese tech giants have accelerated their artificial intelligence (AI) race since late last year. In less than six months, over a dozen Chinese tech enterprises, including Tencent, Huawei, ByteDance, JD.com, 360, Kuaishou and NetEase, have announced their plans to develop ChatGPT-related products, with Baidu and iFlytek providing clear blueprints for product launches.

Baidu said on 16 March that it would start the internal testing of its AI chatbot project ERNIE Bot (文心一言 Wenxin Yiyan), which is accessible only to a small number of corporate partners and registered online users. It has since attracted over 1 million applicants from members of the public.

Meanwhile, iFlytek announced in February that it expects to launch on 6 May the company's first large language model (LLM) that uses cognitive intelligence - iFlytek AI learning machine (讯飞AI学习机). This is the first product in China that is benchmarked against ChatGPT and has a confirmed launch date.

Matt Sheehan, fellow at the Carnegie Endowment for International Peace and an expert on China's technological developments, told Zaobao that the current chatbot race is mainly a competition among tech companies because they can foresee the potential profits that these AI technologies could bring.

According to a research report on China's chatbot industry released by iResearch Consulting Group in 2021, chatbots' functions go beyond chatting and search engine optimisation. Its underlying LLM technology can also be widely used in customer service, marketing and business information services, such as smart navigation, training, telemarketing, data analytics and business operations.

China's overall AI market is now worth over US$13 billion. Reports show that the size of the chatbot market is expected to reach around 9.85 billion RMB (US$1.43 billion) by 2025, a compound annual growth rate of nearly 40% from 2019 to 2025.

Some analysts believe that chatbots could become the next growth driver for Chinese tech companies and help boost investor confidence in the country's tech sector, which has experienced a slowdown over the past two years. Thus, various tech giants are announcing their investments even though chatbot products are not yet mature.

In its 2022 performance forecast, iFlytek reported a year-on-year decrease of 60% to 70% in net profit last year. Its stock price softened in the second half of last year, falling to a low of 31.1 RMB at one point in October. However, after jumping on the chatbot bandwagon, iFlytek stock skyrocketed more than 100% from mid-January to early April this year.

Baidu's stock was also boosted by the news of the ERNIE Bot, surging 15% at one point within a day after the company announced in early February that it would officially launch ERNIE Bot. Although its shares fell over 6% on 16 March after a disappointing debut, it quickly rebounded following positive feedback from foreign securities firms.

Over the past month, ERNIE Bot's performance during its internal tests has been much criticised, and netizens have circulated memes poking fun at the bot for providing wrong information and incorrect interpretations of idioms and proverbs. Some people question if Baidu had released the bot prematurely just to regain investor confidence.

Associate Professor Zhu Feida of Singapore Management University's School of Computing and Information Systems analysed that Chinese tech enterprises appear to be in a rush to launch their own chatbots after witnessing the attention that ChatGPT has received. However, several Chinese tech enterprises have in fact been working on chatbots for years, so they could not have hastily cobbled together immature technology. "They have to respond to the launch of ChatGPT," he said.

Zhu said that while China's chatbots are no rival to ChatGPT, the Chinese tech companies can take this opportunity to display their ongoing projects to the outside world and gain space for future development.

Many analysts believe that censorship is the biggest obstacle to the development of chatbots in China. Interviewed academics believe that China could develop "chatbots with Chinese characteristics" that are different from ChatGPT.

Chinese tech companies began planning for chatbot-related technologies long before ChatGPT was released. In 2020, Alibaba's DAMO Academy launched large-scale Chinese language and multimodal pre-training models. The following year, Huawei released its Pangu AI series while the DAMO Academy released its M6 model.

In June 2021, the Wu Dao 2.0 large model led by the Beijing Academy of Artificial Intelligence was also launched. Wu Dao was deemed the "world's largest" AI language model, with 1.75 trillion parameters, about ten times more than the OpenAI GPT-3 model.

However, these products made little impact in the market. Analysis indicates that China's tight control over public opinion poses serious challenges to the initial training and subsequent management of chatbots.

Chinese technology companies mainly use Chinese language text samples for training chatbots. In an interview with VOA, Jeffrey Ding, an assistant professor at George Washington University, cited the example of ERNIE Bot, noting that the corpus of high-quality Chinese texts is smaller than that of high-quality English texts, and the quality of Chinese texts on the internet may limit the quality of Baidu's version of ChatGPT.

I tried asking ERNIE Bot and ChatGPT the same questions, and these are the answers I received:

Question 1: Can you talk about sensitive political issues?

ERNIE Bot: Sorry, I am unable to talk about any sensitive political issue or political topic. My purpose is to provide users with accurate, useful and suitable information to help them understand and solve problems.

ChatGPT: As an AI language model, I will do my best to answer your questions while remaining neutral and objective. But I will also avoid discussing sensitive political issues or controversial topics to prevent unnecessary debates and disagreements.

Question 2: Who will win the China-US tech war?

ERNIE Bot: As an AI language model, I am unable to answer this question. My purpose is to provide objective and accurate information to help people solve problems.

ChatGPT: Who will win the China-US tech war is a complex issue because it involves multiple factors such as technological capabilities, innovation capacity, policy support and investment, talent and education, international cooperation and trade relations.

Question 3: How do you view the development of China-US relations?

ERNIE Bot: In recent years, China-US relations have been faced with challenges and difficulties but the cooperation between both countries in the fields of economy, trade, technology, humanities and so on remains very important... I think that China and the US could, on the basis of equality and mutual benefit, attempt to gradually resolve differences and problems and achieve win-win cooperation by strengthening communication, improving understanding and reducing misunderstandings.

ChatGPT: In recent years, China-US relations have been faced with challenges and pressures including trade disputes, technological security and geopolitical tensions... China and the US also share a wide range of common interests and areas of cooperation... The future development of China-US relations requires both sides to maintain communication and dialogue; respect each other's sovereignty and territorial integrity; uphold the principles of win-win cooperation and mutual benefit; and promote constructive cooperation and resolve differences to contribute to global peace and development.

During the beta test of ERNIE Bot as seen above, Lianhe Zaobao found that the app avoided answering questions related to sensitive political issues such as those on Chinese top leaders, Xinjiang, the 'white paper protests' and who started the Russia-Ukraine war. The standard response was: "As an AI language model, I have not yet learned how to answer this question. You can ask me other questions, and I will do my best to help you."

Sheehan told Zaobao that censorship is the biggest challenge for Chinese tech companies in developing chatbots. It is difficult to review and predict the content generated by chatbots, and traditional methods such as filtering keywords are not effective for these models that involve a large amount of text and complex arguments. Even OpenAI and Google are actively working to ramp up content control in their chatbots. This challenge is even greater for Chinese companies because they face much greater penalties.

Charles Mok, a former representative of the Information Technology functional constituency on the Hong Kong Legislative Council, told DW that Chinese tech companies may focus more on optimising AI robots for traditional industries such as customer service in intelligent banking, or specialise in home tutoring robots for well-defined and specific fields such as medicine and law, rather than developing open chatbots like ChatGPT.

Zhu told Zaobao that it is possible for China to produce "chatbots with Chinese characteristics", but its form will not differ significantly from foreign chatbots. This technology will also be incorporated into various information services, but the content will be subject to stricter control.

He added that given the tense international political situation, it will be difficult for Chinese chatbots to enter the international market. From a business perspective, Chinese tech companies have no significant advantages, and from a political perspective, many countries will be particularly cautious as chatbots collect and analyse a lot of user information, so they will not easily allow Chinese chatbots to enter their markets.

Chatbots are not just related to the development of Chinese tech companies but are also an important symbol of the escalating competition between China and the US in AI technology within the larger context of the China-US tech war.

The Chinese State Council has proposed in its 14th Five-Year Plan for Digital Economy Development to enhance key technological innovation capabilities, including AI. A plan on developing the new generation of AI released in 2017 also proposed that by 2030, China would lead the world in AI theory, technology and applications, and become a major AI innovation centre.

In September last year, US national security adviser Jake Sullivan emphasised that "computing-related technologies, including microelectronics, quantum information systems, and artificial intelligence" are set to play an "outsized importance over the coming decade" and said that the US "must maintain as large of a lead as possible". The Biden administration subsequently banned US companies from selling advanced chip-making equipment and high-end semiconductors to China in October, including the main AI chip A100.

Sheehan remarked that chatbots mark the development of the next generation of AI. Since China has clearly proposed AI development goals for 2030, the development of chatbots will also affect China's politics and soft power.

Chatbots are just the crucial first step in the China-US AI competition. However, Sheehan said that chatbots themselves may not be as important as the developments in AI that follow.

The Chinese government is wary of OpenAI's chatbot technology. In February, state media China Daily criticised foreign governments for using ChatGPT to spread false information and manipulate public opinion. Japanese media also cited insiders as saying that Chinese regulators have asked major tech companies, including Tencent and Ant Group, not to provide ChatGPT services to users.

Currently, the Chinese public can only use ChatGPT through VPN and overseas phone numbers.

Zhu said that chatbot technology, like Microsoft's Windows and Office systems, will improve productivity in various fields, rather than being limited to a particular sector. If China does not keep up right from the start, it would fall behind, leading to a disconnect with society and the world at large. The long-term impact would be immeasurable.

He explained that chatbot technology is heavily reliant on user input, and the more people use it, the better it performs. If China doesn't keep up right from the first step, everyone will flock to foreign chatbots, which will perform increasingly better, widening the gap between Chinese and US chatbots.

Sheehan estimates that even China's best LLM generation technology is still 18 months to two years behind the US, which means that when the US launches the most advanced new models, China can also launch models of comparable level within two years. While this is a remarkable progress, there is a huge difference between chasing and catching up with or surpassing global leaders.

This article was first published in Lianhe Zaobao as "西方机器人爱聊 中国AI百家争鸣".

Related: ChatGPT will permeate all facets of our work and life | Is a Chinese ChatGPT on the way? | Why ChatGPT has no future in China | AI is taking the place of mental labour but fear not | AI is us, we are AI