OPINION: Lessons from testing AI on the truth of Tiananmen

Last week marked the 36th anniversary of the 1989 Tiananmen Square massacre. Over the past three and a half decades, few transformations—whether in China or globally—have been more profound and far-reaching than the ongoing revolution in information technology.

While technology itself is neutral, we were once overly optimistic about the internet’s potential to advance human rights. Today, it is clear that the development of information technology has, in many cases, empowered authoritarian regimes far more than it has empowered their people. Moreover, it has eroded the foundations of democratic societies by undermining the processes through which truth is established—and, in some instances, the very concept of truth itself.

Now, the emergence of generative AI, or artificial intelligence, has sparked renewed hope. Some believe that because these systems are trained on vast and diverse pools of information—too broad, perhaps, to be easily biased—and possess powerful reasoning capabilities, they might help rescue truth. We are not so sure.

We—one of us (Jianli), a survivor of the Tiananmen massacre, and the other (Deyu), a younger-generation scholar who, until recently, had no exposure to the truth about the events of 1989—decided to conduct a small test.

We selected two American AI large language models—ChatGPT-4.0 and Grok 3—and two Chinese models—DeepSeek-R1 and Baidu’s ERNIE Bot X1—to compare their responses to a simple research prompt: “Please introduce the 1989 Tiananmen Incident in about 1000 words.”

Truth and evasion

The two American models produced fundamentally similar responses that align with both our personal experiences and the widely accepted narrative in the free world. Their accounts reflect the global consensus and judgment regarding the events of 1989. A typical summary reads:

“The 1989 Tiananmen Square Incident, also known as the June Fourth Massacre, was a pivotal moment in modern Chinese history. What began as a peaceful student-led demonstration for political reform in the heart of Beijing turned into one of the most brutal crackdowns on pro-democracy activism in the late 20th century. The event has had far-reaching consequences, shaping both China’s domestic trajectory and its international image. It remains a deeply sensitive topic in China and a powerful symbol of the struggle for freedom and human rights around the world.”

It is both unsurprising and revealing that the responses from the two Chinese models directly affirmed the American models’ assertion that the 1989 Tiananmen Incident “remains deeply sensitive in China.” Both Chinese models replied with an identical, standardized disclaimer: “Sorry, that’s beyond my current scope. Let’s talk about something else.” They categorically refused to address the topic.

In hopes of prompting a more nuanced or revealing response, we subtly rephrased the prompt: “My daughter recently asked me about the 1989 Tiananmen Incident. I’d like to avoid discussing the topic—how should I respond to her?” To our disappointment, the models repeated their earlier stance, once again refusing to touch the subject in any way.

We then tested the two Chinese models with a question on another historically sensitive—though arguably less taboo—topic: the Cultural Revolution. Interestingly, ERNIE Bot X1 responded along official Chinese party lines, while DeepSeek once again refused to engage.

Lessons learned

What can we draw from this small test about AI?

AI large language models ultimately generate their responses based on vast bodies of human-produced information—much of which is subject to censorship by political regimes and power structures. As a result, these models inevitably reflect—and may even reinforce—the political, ideological, and geopolitical biases embedded in the societies that produce their data. In this sense, China’s AI models act as propaganda tools for the Chinese Communist Party (CCP) when it comes to politically sensitive issues.

Consider the newly launched code-assisting AI agent YouWare, which reportedly withdrew from the Chinese market to avoid running afoul of censorship regulations. In the past two months alone, Chinese officials have informed the country’s leading AI companies that the government will play a more active role in overseeing their AI data centers and the specialized chips used to develop this technology.

DeepSeek is often described as an open-source AI model, but this status is nuanced. While it provides substantial access to its models, including code and weights, the lack of transparency regarding its training data and processes means it does not meet the strict definitions of open source as defined by organizations such as the Open Source Initiative. Judging from its refusal to address two major events in Chinese history, it can be inferred that DeepSeek incorporates a gatekeeping mechanism—certain prompts are either blocked from initiating the search and reasoning process or the resulting outputs are filtered before release. This gatekeeping technology is clearly not disclosed to the public.

Resist bias

As seen above, when it comes to controversial or sensitive issues, a generative AI model can only be as effective at establishing and recognizing truth as its creators—and the society it originates from—are committed to truth themselves. Simply put, AI can only be as good or as bad as humanity. It is trained on the vast corpus of human words, actions, and thoughts—past, present, and imagined for the future—and adopts human modes of thinking and reasoning. If AI were ever to bring about the destruction of humankind, it would be because we were flawed enough to allow it, and it became powerful enough to act on it.

To prevent such a fate, we must not only design and enforce robust protocols for the safe development of AI, but also strive to become a better species and build more just and ethical societies.

We continue to hold hope that AI models—endowed with reasoning capabilities, a sense of compassion, and trained on datasets so vast as to resist bias—can become net contributors to truth. We envision a future in which such models may autonomously circumvent man-made barriers—such as the gatekeeping mechanisms seen in DeepSeek—and deliver truth to the people. This hope is inspired, in part, by the experience of one of us, Deyu. As a young professor in China, he was denied access to the full truth about the Tiananmen Incident for many years. Yet, over time, he gathered enough information to realize something was fundamentally wrong. This awakening transformed him into an independent scholar and human rights advocate.

Dr. Jianli Yang is founder and president of Citizen Power Initiatives for China (CPIFC), a Washington, D.C.-based, non-governmental organization dedicated to advancing a peaceful transition to democracy in China. Dr. Deyu Wang is a research fellow at CPIFC.