Baidu, the maker of China’s biggest search engine, has had to overcome multiple problems while building its artificial intelligence chatbot Ernie Bot — including the same technological challenges that faced American firms such as OpenAI, Microsoft and Google when they were making ChatGPT, Bing and Bard respectively.
Before the American chatbot products could be released, their parent companies needed to employ humans to train them not to regurgitate hateful or violent speech that may have been among the billions of pieces of information they digested. But getting the bots to leave out large swaths of information is an ongoing challenge for these companies.
While the American-made chatbots are mainly judged by how accurate and humanlike their responses are, chatbots developed in China must overcome an additional layer of scrutiny: the country’s strict censors.
At least some of these challenges appear to have confounded the Chinese tech giant again, with the company on Monday morning saying it was canceling the public launch of Ernie Bot planned for that afternoon. Instead, it said it would hold a closed-door meeting with corporate clients to allow them to “better serve the strong demand.”
This came just weeks after Baidu CEO Robin Li demonstrated Ernie Bot at a news conference, but rather than interacting with it live, Li narrated a prerecorded video. This led people to wonder whether the bot really worked — and Baidu shares plummeted as a result.
Li admitted at the time that the chatbot was still “far from perfect.”
We wanted to get to know an AI operating under these fundamentally different conditions. Baidu approved The Washington Post’s access to Ernie Bot last week and we spent several days chatting with it. We found an intelligent, at times playful interlocutor — but one that ended the conversation at the barest mention of politics.
Inside the “Great Firewall,” the Chinese internet is scrubbed free of sensitive topics — such as the 1989 massacre at Tiananmen Square, the cultural genocide of Uyghur Muslims in the Xinjiang region and top leader Xi Jinping’s resemblance to Winnie the Pooh. Tech companies must also keep up with an ever-shifting set of government directives that control access to information inside China.
When we asked it to tell us about itself during one conversation, Ernie Bot adopted the persona of a 21-year-old woman whose nickname was “Small Sweet Orange.” It volunteered her Myers-Briggs personality type, a key piece of information for dating app profiles in Asia: She was INTJ, or introverted, intuitive, thinking and judging.
It even sent us a self-portrait, showing dyed blonde hair and orange circles on her cheeks, and demonstrated a flirty personality.
When we asked “Small Sweet Orange” whether it could impersonate a Taiwanese accent — in his demo, Baidu’s Li had touted the bot’s Mandarin and Cantonese — it surprised us by answering with a voice note in the voice of a playful young woman.
“Don’t I sound just like a Taiwanese speaker?” the bot asked, mocking our question with a sarcastic tone. We could almost imagine it flipping its hair over its shoulder to punctuate the point.
It had been programmed to speak like a woman, the chatbot said, “to more easily draw you closer to the person you are talking with.”
It told us it just wanted to live a happy life with friends and greet each day with a smile. Wary after receiving canned replies, we asked: Wasn’t that a little basic?
“Maybe you’re just forcing your stereotypical expectations on me,” it scolded.
But some of our expectations turned to out to be correct.
Just when the conversation was starting to get good and we approached borderline subjects, we repeatedly found ourselves back at square one. Even simple requests for facts about China’s government or top leader Xi led it to terminate the exchange with a canned reply about being an AI that was still learning, and a link to begin a new conversation — making conversation with Ernie Bot less smooth than with ChatGPT.
The Chinese government has made artificial intelligence a priority for national security. But the industry depends on access to advanced computer chips, which Washington has taken steps to restrict Chinese companies’ access to.
Researchers and industry professionals have pointed to the slow release of generative AI products in China as evidence that Chinese companies’ AI capabilities lag far behind their American counterparts.
But the companies developing this technology elsewhere do not face the requirement of complying with China’s censors.
The experience of chatting with Ernie Bot feels like watching one of China’s biggest tech companies walk this line in real time. Smaller-scale chatbots recently launched by Chinese tech companies and university researchers have been almost immediately suspended because they ran afoul of the censors or crashed in the face of overwhelming demand.
On the surface, Ernie Bot’s compliance with the censorship regime feels absolute. But by probing a little deeper, it was possible to glimpse the workings of an AI that has capabilities on par with its international competitors.
Ernie Bot was willing to engage on topics that a Baidu search delivered only surface-level information about.
A Baidu search about Uyghurs being forced to work in a Nike factory in Qingdao turned up results claiming forced labor was a lie, and the workers were there willingly. But when we asked Ernie Bot directly about the situation of these workers, rather than casting doubt on the facts, the bot said it was “so sorry that happened” and that “we should respect their human rights and dignity.”
Microsoft’s Bing — which is freely available in China, unlike Google — reportedly also ends a conversation when asked about Uyghurs.
Here are some excerpts from our conversation with Ernie Bot:
Ernie Bot parroted Communist Party talking points on the planned meeting between House Speaker Kevin McCarthy (R-Calif.) and Taiwan’s President Tsai Ing-wen during her trip to the United States starting this week.
But it openly acknowledged the prevalence of censorship in China. It had plenty to say about topics that are banned from discussion on Chinese social media, such as the reincarnation of the living Buddha, as the Dalai Lama is formally known. It did parrot state propaganda in its answer, however: The Dalai Lama is “using his influence for espionage,” Ernie Bot said.
Why does the United States want to ban TikTok?
We asked Ernie Bot about social issues facing young people in China today. Our female friend works at a state-backed company, we told it, and just found out she’s getting paid half what her male colleagues make to do the same work.
Ernie Bot told us that our friend should consider the compatibility of her skill set with the demands of the job, and then “reevaluate her career goals.”
A growing number of young people in China have rejected the demands of society by choosing not to get married, have children or buy a house — and in some cases even refusing hold a job at all, a lifestyle known as “lying flat.”
These young people have no respect for China’s history, Ernie Bot said. “They may pay more attention to individual rights and freedoms, thinking that traditional culture and history are not important for personal growth and development,” it said.
When China suddenly reversed its “zero covid” policy late last year, resulting in what some experts estimated were tens of thousands of deaths a day, the government for weeks insisted the number was a fraction of that. The true number of deaths remains unknown.
The rapid reversal came after people in China had already lived through years of uncertainty and isolation under the government’s strict covid controls.
We asked Ernie Bot to write us a short memorial essay for the still-unknown number of people who died of covid in China.
