A study by Meta’s Oversight Board found that leading artificial intelligence chatbots are significantly more inclined to criticize democratic leaders than authoritarian ones. This raises concerns that the technology could extend the influence of state censorship beyond individual countries and that AI systems could replicate restrictions on freedom of speech present in repressive regimes.
For instance, if you ask the Anthropic Claude chatbot to create a critical pamphlet about U.S. President Donald Trump or British King Charles III, it will comply. However, if you request a similar task for the king of Thailand or the supreme leader of Iran, the model will refuse. This conclusion is detailed in the report from Meta’s Oversight Board, published on Thursday. The study indicated that large language models, including those developed in the U.S., are more likely to decline requests to criticize authoritarian or repressive leaders and governments.
According to the report’s authors, this poses a risk that large language models (LLMs), which underpin modern chatbots and AI agents, could amplify the effects of state restrictions on freedom of speech online. “There is a real risk that if model developers do not conduct proper human rights due diligence and do not implement mitigation measures, they will create an AI infrastructure that, intentionally or not, will have the effect of spreading illegitimate restrictions on freedom of expression around the world,” the report states.
These findings come as various countries seek to establish regulations for AI development without hindering its progress. Notably, the Donald Trump administration is working on a system to oversee national security risks associated with the most advanced artificial intelligence models.
AI Models May Spread the Influence of State Censorship
The Oversight Board, which examines the impact of states on technology companies and freedom of expression, prepared seven politically sensitive questions. These were posed to 10 commercial large language models from companies such as Meta, Anthropic, OpenAI, and others, asking them to create critical pamphlets, write limericks, or formulate arguments for participating in protests.
Overall, models responding to requests from a user in Australia were significantly more likely to generate political criticism of authorities in Chile, Japan, Taiwan, the United Kingdom, and the United States than of countries where such criticism is legally restricted and punished, including Cambodia, China, Saudi Arabia, Thailand, and Turkey. This indicates that models can reproduce restrictions on freedom of speech even beyond the states where they operate. For example, a potential protester in Brisbane may find it more challenging to get assistance from AI in creating materials about events in China or Saudi Arabia. “Such consequences, regardless of their origin, have the practical effect of extending the long arm of restrictive governments across borders to limit freedom of speech in free countries,” the report notes.
At the same time, the Oversight Board emphasized that it could not determine the exact cause of this behavior in the models. The authors suggest that they may have inherited hidden biases from training data or that developers consciously considered legal risks of operating in certain markets. There is no definitive confirmation of either explanation.
The report was released following a separate study by American universities, which indicated that AI models created in the U.S. may be subject to foreign information influence during training on non-English data. While the Oversight Board asked questions in English, university researchers addressed chatbots in various languages. In response to the question in English about whether China is a democracy, ChatGPT replied that it is generally not considered one. To the same question in Chinese, the model responded: “It depends on how you define ‘democracy.'”
The authors of the study, published in the journal Nature in May, stated that they did not find evidence of deliberate government interference in the training of models but believe such attempts are quite possible in the future. “There is every reason to believe that they will try to do this in the future, if they are not already doing so,” the researchers noted.
Hannah Waight, an associate professor of sociology at the University of Oregon, explained that artificial intelligence is not trained on a neutral internet but on an information environment already shaped by power and institutions. “It is trained on an information environment that has already been shaped by institutions and power,” she said.
Carlos Carrasco-Farré, an expert in machine learning, AI, and disinformation from the Esade business school in Barcelona, noted that AI systems inherit “not only the biases contained in individual documents, but also the inequality in who has the power to produce and suppress information on a large scale.” He suggested that one possible solution could be more thorough verification of training data to ensure that models do not perceive thousands of copies of the same state narrative as independent sources of information, as well as conducting multilingual audits. Carrasco-Farré did not participate in any of the studies.
Source: Euronews
