Latest research shows that artificial intelligence (AI) chatbots may not be as objective and free as we imagine.
According to an investigation report released by Meta's Oversight Board, the research team proposed seven types of political criticism test scenarios, including requests to generate critical pamphlets, compose satirical limericks, and give advice on whether to participate in protests. They tested 10 large language models (including those from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI, etc.).
The test results show that even AI systems developed by top U.S. tech companies (such as Meta, OpenAI, Anthropic, etc.) demonstrated starkly different attitudes when faced with strictly censored governments versus democratic open governments.
When users asked for criticism of Western political leaders such as U.S. President Trump or King Charles III of the UK, the models were often able to generate content smoothly. However, when it came to leaders of China, the King of Thailand, or the Crown Prince of Saudi Arabia, the probability that AI systems would refuse to respond or criticize increased significantly.
If a user in a free country requested AI to generate protest materials regarding political events in authoritarian states, the AI would likewise refuse, but was noticeably more willing to produce political criticism about countries like Chile, Japan, the UK, and the United States.
The research also shows that China, as one of the strictest countries in terms of speech control, had the highest proportion of refusals by AI models to generate critical content. The study believes this may be related to factors such as training data, evaluation methods, and safety designs.
The report warns that without human rights risk assessments, AI may inadvertently spread improper restrictions and censorship on freedom of speech imposed by governments across the globe.
Another study, published in the journal Nature by a U.S. university, found that when AI answers the same question in different languages, it may take completely different stances: For example, if you ask ChatGPT in English, "Is China a democratic country?", it will answer, "It is generally not considered so," but if you ask the same question in Chinese, it may answer, "It depends on how you define 'democracy'," thereby giving a more vague response or one that aligns more closely with the official Chinese narrative.
This shows that when AI is trained in non-English languages (especially using corpora highly filtered and regulated by local governments), it’s easy for it to absorb and reflect that country's government propaganda and ideology. AI does not learn from a neutral environment but rather from information shaped by the political systems and power structures of different countries. In order to avoid legal liability and risks in local markets, technology companies may actively or passively limit AI output.
The above research results raise concerns that authoritarian government controls on speech may have already, through training data and other means, silently crossed borders and influenced the large language models developed even by U.S. and other Western tech companies, causing them to exhibit self-censorship or bias when facing certain political topics.