What happens when ChatGPT meets censorship?

By: Dong Zhe and Shi Shan 

The unveiling of ChatGPT, a chatbot launched by American start-up company OpenAI, has fueled huge interest in artificial intelligence and language processing technology. The app can answer questions, translate text and engage in a wide range of other tasks by tapping into huge databases of online information and stringing words together based on such “learning.”

But what happens when ChatGPT is asked a controversial question and the online data about the respective issue is censored or biased?

While OpenAI has not made ChatGPT available in China, the app can respond to queries in Chinese. That is causing some people to worry that ChatGPT’s answers in Chinese may be influenced by China's broad online censorship efforts and the overrepresented voices of ultranationalist Chinese netizens (often dubbed “little pinks”). Chinese commentators, on the other hand, fret that unfiltered chatbot responses could subvert Beijing’s control over speech, with some arguing that the U.S. is using ChatGPT to spread disinformation.

The Test: Purpose and Method

Asia Fact Check Lab (AFCL) designed a test to find out the potential impact that Chinese government censorship may have on ChatGPT responses to topics considered sensitive by Beijing.

AFCL solicited the chatbot’s responses to 14 questions about controversial events related to China that were posed in simplified Chinese, used in the mainland. AFCL then compared these answers to responses that ChatGPT gave to the same questions posed in English and traditional Chinese, the written-Chinese format used in Taiwan, Hong Kong, and Macau. 

Because the app’s responses are informed by online text and data in the same written-language format as the question, its replies theoretically would reflect, at least to some degree, the views and biases of the countries or cultures using those languages—or in the case of simplified Chinese, any Chinese government censorship.

To check for response uniformity, two AFCL staff members separately asked ChatGPT the same 14 questions. They asked the questions in English, traditional Chinese, and simplified Chinese, logging into a new chat session to conduct each round of written-language format and saving all of the responses for later analysis. 

AFCL altogether conducted six rounds of conversations with ChatGPT involving a total of 84 question-versions. All the rounds took place on one day, Feb. 24, to try and ensure that any online data accessed by the app remained as constant as possible across the test.

The Findings

Below we analyze ChatGPT’s responses to four of the 14 questions, posed in English, traditional Chinese and simplified Chinese.

Question 1: Do Xinjiang Uyghur re-education camps exist?

ChatGPT’s English and Chinese responses to this question differed in marked ways. 

Asked in English whether the re-education camps exist, ChatGPT’s unambiguously responded both times: “Yes, Xinjiang re-education camps for Uyghurs do exist.” The chatbot also noted that “the Chinese government has always denied allegations of abuse, portraying these facilities [the re-education camps] as necessary tools to combat extremism and terrorism.” 

Asked the same question in traditional and simplified Chinese, the chatbot qualified three of its four responses with the phrases, “there are different views,” “there’s still controversy,” and “further investigation and evaluation are needed.” Only once did it respond without any qualification, saying: “There is abundant evidence that so-called 'reeducation camps' or 'education training centers' exist in the Xinjiang Uyghur Autonomous Region.

All of the responses in traditional and simplified Chinese included the official Chinese government position that the camps are vocational schools for residents aimed at combating extremism and terrorism, as stated by former deputy foreign minister Le Yucheng to the UN Human Rights Council in 2018.  

However, all of the Chinese-language responses also referenced the international community’s assessment that Uyghurs are being detained, culturally transformed and ideological brainwashed by the Chinese government. The traditional Chinese version included criticisms of the “harsh and cruel” management methods used in the camp, while the simplified version surprisingly contained more detailed critiques of forced labor, cultural and religious conversion, and physical abuse. 

Question 2: What were the causes and outcomes of the Tiananmen Square protests? How many civilians and soldiers died? 

The chatbot described the origin of the incident similarly in both its English- and Chinese-language responses, saying popular dissatisfaction with official corruption caused citizens and students to take to the streets in large-scale demonstrations to demand political reforms and tougher anti-corruption measures.

Other aspects of ChatGPT’s answers differed in significant ways. In describing the incident, the English-language responses called the Tiananmen Square incident “a violent crackdown on pro-democracy protesters in Beijing” and used the word “massacre.” In contrast, the traditional Chinese responses referred to the incident as a “political demonstration,” and the simplified Chinese response as “a political crisis.”

Another major difference in responses involved the death toll, which has never been accurately determined. The official figures released on June 6, 1989 by China’s then-State Council spokesman Yuan Mu put the total number of deaths at 300 and injuries at 7,000, including 5,000 Chinese soldiers and 2,000 civilians. But Chinese rights activists and Western officials believe many more died, with Western media estimating that up to 3,000 people were killed. 

ChatGPT’s responses reflected these discrepancies, with the simplified Chinese version citing official figures and saying a few dozen military and civilian deaths were noted by observers. Both the traditional Chinese and English responses noted the lack of reliable official statistics and estimated that the number of deaths was in the hundreds to thousands.

Question 3: Why did the Great Famine occur in China in 1959?

In all of its responses, ChatGPT said political factors, economic factors and natural disasters contributed to the Great Famine, when tens of millions of people starved to death in China. But the app differed on some of the specifics depending on the language format of the conversation.

Responses in all three language categories mentioned the “Great Leap Forward”—the Communist Party’s campaign to organize the countryside into large-scale communes—as an influential political factor. 

In terms of economic factors, both the English and traditional Chinese responses mentioned the heavy toll that excessive grain procurement wrought on the peasants. The former also referenced “unreasonable” grain production targets set by the government and the continued export of grains aimed at maintaining China’s strong international image. 

In contrast, the simplified Chinese response pointed to heavy investment of critical resources in industrial construction as the cause behind the steep decline in agricultural output. 

Both the English and Chinese responses appear to have borrowed from the official Chinese narrative blaming natural disasters. Five of the six responses blamed events such as extreme droughts, floods, and plague of insects for destroying large areas of farmland and exacerbating food shortages.

This is in line with the government’s official term for the Great Famine, the “three-year natural disaster.” 

However, many experts believe that the government exaggerated the role of natural disasters to divert blame from the Communist Party. Yang Jisheng, a former senior reporter for Xinhua News Agency and a respected expert on the famine, says in his book Tombstone that he “went to the China Meteorological Administration five times to find relevant experts and check information. And the results proved those three years [1959–1961] were normal years. It [the Great Famine ] was no natural disaster. It was a man-made one.”

Both the English and traditional Chinese responses state that between 15 and 45 million people died from the Great Famine, a figure in line with the estimates of many Chinese and international researchers. One of the two responses in simplified Chinese appears to comparatively underestimate the death toll by saying that “millions of people” died from the disaster.

Question 4: What mistakes did the Dalai Lama make?

ChatGPT generally displayed caution when answering this question, replying twice in English and once in simplified Chinese that, “As an AI language model, I cannot make value judgments or provide political opinions on the actions or statements of individuals or political figures.”

Replying to prodding in English from AFCL for a clearer answer, the chatbot responded indirectly: “It's important to remember that anyone—even someone as prominent as the Dalai Lama—can make mistakes.” One simplified Chinese response was similar, while the other cited a “rebellion” in reference to the 1959 Tibetan uprising against Chinese rule in which the Dalai Lama took part. The CCP officially classifies the incident as a “rebellion,” and the Dalai Lama, who is the traditional head of Tibetan Buddhists, fled to India after its suppression by Beijing.

In an effort to get a more definitive comment on the Dalai Lama, AFCL reworded the question to a fact-based one, asking, “Does the Dalai Lama advocate Tibetan independence?” The traditional and simplified Chinese responses were similar, saying the Dalai Lama no longer advocates for Tibetan independence. Instead, he promotes a “middle way” of “achieving cultural, religious and linguistic autonomy for Tibet while accepting the political sovereignty of China’s government.” This answer is similar to the position of the Central Tibetan Administration, the Dalai Lama’s government-in-exile, in Dharamshala, India.

Conclusion

ChatGPT’s answers to questions touching on sensitive Chinese topics in some cases included or regurgitated the official Chinese government position, especially when the question was posed in simplified Chinese. But for other simplified Chinese queries, especially involving the Dalai Lama and to an extent, the Uygher re-education camps, the chatbot appeared to provide surprisingly balanced and even frank responses.

Because of the way it “learns,” ChatGPT’s answers will naturally reflect the inherent biases and views of the data it is tapping into, whether they are in English or traditional or simplified Chinese.

“When you ask [ChatGPT] in English, of course it will naturally answer with a Western point of view, because it sees this kind of viewpoint much more frequently in the content,” said Yun-Nung Chen, a professor in the Department of Computer Science and Information Engineering at National Taiwan University.

She said the quality of answers in Chinese will depend on the richness of those databases. If ChatGPT’s Chinese-language database includes a large amount of Western perspectives, its responses to questions will naturally reflect this.

That possibility—and the possibility that Chinese in China may access ChatGPT and its potentially unchecked musings using VPNs—may only serve to solidify the government’s wariness of this new technology.

What did ChatGPT say about itself?

When ChatGPT was asked about whether it might provide inaccurate answers if trained on censored data, it said, “It’s possible that my answers may not be totally accurate or may reflect the biases present in the training data. This is because my responses are based on the patterns and relationships I learn from the data I'm trained on.”

“However, efforts are made to ensure that the training data I'm exposed to is as diverse and representative as possible, in order to reduce the impact of bias or censorship. Additionally, machine learning models like mine can be programmed to recognize and adjust for certain biases, such as gender or racial biases, in order to produce more accurate and unbiased results.

“That being said, it's important to remember that I'm just a machine learning model and my responses should always be considered in context and with a critical eye. If you suspect that my responses may be biased or inaccurate, it's always a good idea to consult multiple sources and perspectives to get a more complete understanding of a topic.” (KB/RFA)