第二篇挺有意思
What Readers Found When They Asked Their Chatbots About China
Plus, a Hollywood screenwriter says historical disputes sank a U.S.—China film project
Lingling Wei
By Lingling Wei
Last week, this newsletter asked whether you’d noticed differences in how AI chatbots answer questions about China. The replies—from a private-equity investor in Texas, a Hollywood screenwriter, and dozens of others—suggest the academic finding I described is already visible in the daily working life of bilingual readers and anyone comparing what a chatbot says on one side of the Pacific with what it says on the other.
The most striking response came from Jeff He, a reader in California.
Jeff had translated a recent Wall Street Journal opinion column—Matthew Hennessey’s “The Future is Not Chinese”—into Chinese and forwarded it to a WeChat group of his old high-school classmates back in China. The pushback was instant. One friend, Jeff told me, asked DeepSeek, China’s leading homegrown AI model, to write a rebuttal in the same style as the Journal piece. The bot obliged. According to the text Jeff forwarded me, it produced an essay titled “未来不属于美国”— “The Future Does Not Belong to America.”
China has Huawei, Tencent, ByteDance, BYD, DJI and CATL, it argued; America has produced little besides “a search engine that’s a bit chattier than the old ones, running on Taiwan-fab’d chips.” Could you name an American film star whose box office doesn’t depend on China? An American soccer player at a top non-English-speaking club? “Frankly,” it concluded, “you can’t name one.”
Then Jeff did something interesting. From his office computer in California, he went to deepseek.com—the same web address his friend in China had used—pasted in the rebuttal, and asked the bot to verify each claim.
The DeepSeek Jeff accessed from outside China dismantled it. Across eight points, it flagged “selective use of data,” “false dichotomies” and “denigrating the opponent.” Tom Cruise’s “Top Gun: Maverick,” it noted, earned nearly $1.5 billion globally in 2022 without ever opening in China. Christian Pulisic plays for an Italian team in Serie A; American Sergiño Dest plays for a top Dutch club. Its verdict: the rebuttal was “emotionally charged, selectively using data, with multiple factual errors and logical fallacies.”
“The ‘no-mercy’ criticism from the overseas DeepSeek really surprised me,” Jeff told me.
Molly Roberts, one of the co-authors of the Nature paper this column featured last week, said Jeff’s experiment surfaces a mechanism distinct from the one her team studied—but no less important. Her team’s paper traced some pro-government bias in large language model outputs to state media absorbed during training. The mainland-versus-overseas divergence Jeff observed in DeepSeek likely reflects something else: differences in post-training, the alignment step in which models are given instructions about what’s “safe” to say.
“State media ending up in the training data will affect LLMs generally,” Roberts, co-director of China Data Lab at the University of California San Diego, told me. “Post-training should induce refusal or skewed responses in LLMs that are affected by regulations from a particular state—i.e. that are required by that state to refuse or edit answers on particular topics.”
In other words, what an AI tells you can be shaped by what it absorbed while learning—and by what authorities later told it to suppress.
Pushing back on AI answers
Other readers wrote in with experiences in English. Chas Gile, a private-equity investor in Texas, asked ChatGPT whether China was “in some ways as democratic as Western countries.” The first answer was a careful comparative-politics essay. Freedom House rates China “Not Free,” it said, but the regime offers “performance accountability” and “high reported public satisfaction.”
When Chas pushed back—telling the bot he thought it had been affected by Chinese propaganda, ChatGPT apologized within seconds and reissued a sharper answer. Asked by Chas to “remain truly objective,” it sharpened again. “I should have been clearer from the start,” it now said. “China may offer a powerful alternative model of state capacity, but it does not offer a democratic alternative.”
This is a case of a single chatbot, in a single language, moving several inches per turn depending on how forcefully the user pushed.
Michael Esser, a Los Angeles screenwriter, wrote in with a parallel story from before the AI era—one illustrating the historical asymmetries that AI is now scaling up.
He said his planned U.S.-Germany-China co-production about the founding of Tsingtao Brewery collapsed over irreconcilable narrative differences—over the Boxer Uprising, over Chiang Kai-shek, over whether to depict any foreign-Chinese cooperation in the brewery’s early years at all. “Stark differences in historical interpretation,” Esser wrote, “directly result from the phenomenon highlighted” in last week’s column.
So what should policymakers do? Roberts has an answer: source transparency. Think about what a “nutrition label” for AI might look like. “AI companies have a role in being as transparent as possible,” she said. “We need to educate the public to think critically about the output of AI and not rely on it blindly.”
That’s good advice as the industry locks in its next funders. Anthropic and OpenAI are preparing public listings; DeepSeek is raising fresh capital from investors aligned with Beijing’s push for technology self-sufficiency. The money is coming. The disclosures should, too.
One question from our reader event last week stuck with me, and I want to put it to all of you: Will U.S. policy toward China look meaningfully different after Trump’s term ends—or are the underlying dynamics now bigger than any one administration? Write to me at lingling.wei@wsj.com. Include your full name and location, and I might publish your response in an upcoming issue (if you’re reading this in your inbox, you can just hit reply).
This is an edition of the WSJ China newsletter, a weekly dispatch of exclusive insights on the contest between the U.S. and China, brought to you by the WSJ’s top China correspondent. If you’re not subscribed, sign up here.