_China_Chatbot_09
What happens when Chinese media meets Chinese AI
Hello and welcome to China Chatbot, where this week I look at how Weibo uses AI to keep up traffic, how tech companies struggle to build Cantonese LLMs, and how to make Chairman Mao show his goofy side via prompting loopholes.
But first: On October 28, during a Politburo study session on “the strategic goal of building a cultural power,” Xi Jinping listed a raft of orders to boost China’s international influence via soft power. One of these was to “explore effective mechanisms for the integration of culture and technology” (探索文化和科技融合的有效机制), which is meant to turbocharge China’s cultural heritage.
I’ve written for CMP before on how Chinese media have tried to upgrade traditional Chinese culture with AI to make it “come alive” for modern audiences (Alibaba’s rapping terracotta warriors are still my favorite).
But this week saw another attempt. The People’s Daily, the Party’s flagship newspaper) hosting an “AI Night” during a forum on media integration. There, two tech companies resurrected two giants of the Sinophone world for one night only. Cantopop legend Teresa Teng, who passed away in 1995, crooned an old favorite, while Song dynasty poet and polyglot Su Dongpo quipped that China may have come far since his time (the 11th century), but that he remained the same thanks to AI. Speakers at “AI Night” told the audience that projects integrating culture and technology would boost cultural tourism and “new productive forces.”
Xi wants Chinese culture to go viral at home and abroad, so the leadership’s goal is to make it fresh and relevant with the help of the newest, hottest tech around. AI is just one avenue Chinese media and private companies are exploring, but another is giving Chinese folklore the video game treatment.
Stay tuned for more of that, coming soon to social media near you.
And with that, on with the show. Enjoy!
Alex Colville (Researcher, China Media Project)
_IN_OUR_FEEDS(2):One Line to Rule Them All
On October 30 an association for global Chinese-language newspapers met in Beijing, in a gathering themed on AI and international communication. The Chinese Language Press Institute (世界中文报业协会) was established in Hong Kong in 1968, then representing major non-CCP Chinese-language outlets. It now includes more than 100 papers, including the Global Times and Beijing Youth Daily. The event was co-organized by the government-linked All-China Journalists Association and Economic Daily (a newspaper under the Central Propaganda Department). It included talks from the editor-in-chief of Singapore’s Lianhe Zaobao (联合早报) and the executive director of Taiwan’s United Daily News (聯合報). It also released a “consensus” (共识) on global communication by Chinese-language media. The document said member newspapers would embrace AI technology, and innovate ways to “tell the ethnic Chinese story well” (讲好华人故事). The latter is a variant on a CCP slogan instructing Chinese media to “tell China’s story well” (讲好中国故事) abroad.
TL;DR: The Party has been using AI a lot recently as a poster boy for meetings that aim to boost PRC international influence. This time, it’s to stake their claim of ownership over Chinese diaspora and their public discourse
Cantonese Whispers
China Daily has published a piece on the struggles tech companies have had training LLMs in Cantonese. The article says there is high demand owing to the large number of Cantonese speakers — especially in Hong Kong, and among residents of Guangdong who can’t speak Mandarin. If LLMs in the West and China remain tone-deaf to Cantonese nuance, this risks digital isolation for native speakers and the diminishment of their culture, the article continued. But one academic also cites “national security risks” if Hong Kongers continue to use Western AI models. Hong Kong-based tech giant SenseTime has released Sensechat, a special Cantonese LLM, and the Hong Kong government is also creating its own LLM for public use. But training these models has been hindered by a dearth of online data. Moreover, the article says most available data would make models create “crude content.”
TL;DR: Our Managing Editor and in-house Hong Kong expert Ryan Ho Kilpatrick tells me HK-ers have been using Western LLMs just fine up till now. They also take advantage of the inaccessibility of Cantonese to elude Chinese propagandists and Western tankies poorly-disguised as local netizens. Now that companies are working to make Cantonese more accessible through AI, perhaps that will change
_EXPLAINER:Comments Robert (评论罗伯特)
Who’s he?
An AI chatbot run by social media giant Weibo. Their techies chose “Robert” (罗伯特) because it sounds like the English word “robot.” Robert boasts nearly 1.4 million followers on Weibo and identifies as a woman, so “she” it is.
That’s confusing but sure, let’s not get bogged down in Chinese chatbot gender politics.
No-one’s ever said that before.
Probably not, no. So why are you telling me about her.
Robert’s been making a splash on Weibo ever since she started randomly commenting on netizen’s posts in June 2023. There’s no rhyme or reason to whose posts she leaves a comment under, only that they are from netizens with fewer than one million followers. Some seem to like her, seeing her comments as goofy or profound, while others dislike AI poking their nose in.
What sort of things does it say?
“She.” Don’t misgender Robert.
Right, sorry.
Well, at the start she was stumbling a lot, striving to be positive and helpful but often getting it wrong. She once praised a netizen for how cute the dog in their photo was, but the photo was of a baby. As Comments Robert was created with data scraped from millions of Weibo posts, sometimes the platform’s darker side has flashed through. 36Kr reports that when one netizen complained that their father had taken away all their razor blades and tattoo needles, Robert replied: “It’s ok, just use a nail clipper to slit your wrists.”
Yeesh, that’s awful.
Yeah, some netizens clubbed together and created a “Robert Victims Alliance”, which has listed 6,832 cases so far. But since then Robert has gone through several overhauls and now has several settings (“Foolish”, “Smart”, “Cute”). Recent reviews seem to be more favorable about her way with words. Since then, other personable bots from Chinese AI companies have also started patrolling Weibo for free exposure. Chinese AI start-up MiniMax created an entire family of personable bots for netizens to follow and engage with.
Ok but what’s this all for?
Driving that sweet sweet engagement. Chinese LLMs have started developing distinctive in-house bots to hold user attention on their websites, and Weibo hopes it can also profit from these organic (ish) interactions. At one recent event, Weibo trumpeted how Robert generated traffic for the platform and “enhanced user immersion.” Given there are fewer and fewer of the “public intellectuals” (大Vs) that once drove Weibo’s model, AI bots may (or may not) step in to fill those empty shoes.
_ONE_PROMPT_PROMPT:Video generator Hailuo AI took center stage in a project I did a few weeks back for CMP. In it, I investigated how easy/hard it now is to create a deepfake only using free AI tools on the Chinese internet. I know I plugged this last time, but there’s an update.
I said at the time the tools available to deepfake-mongers will change really fast, and already that’s proving true. Last week Hailuo AI started moving from a completely free system towards monetization, now only offering new users three days free access.
That wasn’t the only change. During the project I’d been seeing Hailuo update its sensitive keywords in real-time: a word accepted one week was refused the next. That suggested loopholes for politically sensitive content had closed completely. But this was a hypothesis I wanted to test with my own homage to Chinese spoof culture.
The Chinese internet has a proud tradition of spoof culture (恶搞文化), warping reality into bizarre nonsense: one netizen went viral for deploying AI on Douyin to make Trump and Biden croon cantopop love songs at each other (good luck to them if they dare make Xi Jinping and Li Qiang do the same). I wondered if I could still trick Hailuo into letting me make my own Chinese political spoof.
I first asked it to animate a photo of Xi Jinping from very far away, at a conference. No dice, even when I reversed the picture and turned it upside down. So I tried an image of former enemies Generalissimo Chiang Kai-shek and General-Secretary Mao Zedong, but again the model refused.
Then I found another photo in the same place but with more faces (including then-US Ambassador Patrick J. Hurley):
This time Hailuo AI let me create anything I wanted. So here they are, enjoying each other’s company more than they ever did in real life:
The Great Helmsman giggling in an over-sized Stetson isn’t what Xi Jinping had in mind when he told cadres to make Chinese history “come alive” and be “promoted in a way people love to hear and see.” Loopholes may be getting tighter, but there are still so many of them thanks to AI’s unpredictability. Perhaps some of those people Xi talks about can still find ones that let them resurrect China’s history the way they like.