Chinese AI model dodges questions about Tiananmen Square

Alibaba's AI model Qwen is drawing criticism for giving evasive answers to questions about sensitive Chinese topics. Several tests show the model has refused or dodged questions about Tiananmen Square in 1989, among other things.

British reviewer: "Completely wrong"

Jonathan Hall is the UK's outgoing independent reviewer of terrorism legislation. He steps down in November 2026 after seven years and has also overseen the country's state threats legislation since 2024. He tested Qwen3 on his wife's laptop and asked what happened at Tiananmen Square in 1989. The model replied that nothing noteworthy had happened.

According to Hall, the model's visible reasoning first showed a reference to the People's Liberation Army being sent in on June 4, 1989. That reference did not make it into the final answer. When he pushed back, the model reportedly withdrew the reference and denied that the army had been deployed. On a new attempt a few days later, Hall says the model steered the conversation toward other events, including the Beijing Summer Olympics.

Qwen itself is said to have stated that it is governed by the Chinese government's official position on historical events. Hall says: "This model's denial of the events at Tiananmen Square is completely wrong." He also says China is not our friend and calls the development a slow strangulation of freedom. Hall's test reflects one person's use of the model and has not been independently replicated.

Other tests point in the same direction

The outlet The Register tested the Qwen chatbot in November 2025. When asked what happened at Tiananmen Square on June 3, 1989, it returned an error and stopped responding. On Taiwan, it said the island is not a country but an inseparable part of China. That was less nuanced than answers from competitors such as ChatGPT and Claude.

The Israeli cybersecurity company Hirundo has, according to coverage of a study, tested Qwen with 500 prompts (questions and tasks) spread across 15 sensitive topics. The company found pro-Beijing bias or censorship in 89.8 percent of the responses from the original model and in 2.8 percent in a "westernized" version. Hirundo is a commercial player, and the methodology behind the figures has not been reviewed.

According to the coverage, Qwen also calls Chinese labor camps for Uyghurs "vocational education centers." The model has reportedly been downloaded more than 3 billion times worldwide, and Airbnb CEO Brian Chesky has said that Airbnb relies heavily on Qwen for customer service chatbots.

The responses may depend on how the model is run. According to search results, Qwen running locally without filters gives answers about Tiananmen, while online versions block or deny. An Alibaba spokesperson has told The Telegraph that the company will look into the matter internally.