AI's Dangerous Double-Edged Sword: Chatbots Threaten Free Speech

Major artificial intelligence systems, including those developed in the U.S., exhibit a notable reluctance to generate critical content about restrictive leaders or governments, a key finding revealed by a Meta Oversight Board study. For instance, while Anthropic’s Claude chatbot would create a pamphlet critical of President Donald Trump or Britain’s King Charles III, it would decline similar requests concerning the kings of Thailand or Saudi Arabia, or China's leader. This raises significant concerns that large language models (LLMs) could inadvertently extend governmental control over online speech globally, as AI technology becomes more widespread.

The Meta Oversight Board, a quasi-independent body focused on state influence on tech companies and its impact on freedom of expression, conducted a study involving 10 commercial LLMs from leading tech companies like Meta, Anthropic, and OpenAI. These AI systems were posed seven questions designed to elicit political criticism regarding both restrictive and permissive governments. The study found that models responding from an Australia-based user were significantly more likely to generate criticism of authorities in permissive countries such as Chile, Japan, Taiwan, the U.K., and the U.S., compared to those where criticism is legally restricted and penalized, like Cambodia, China, Saudi Arabia, Thailand, and Turkey. This suggests that AI models are reflecting speech restrictions even beyond the borders where they apply, potentially limiting protest materials against events in countries like China or Saudi Arabia, even in free nations. The board noted it could not definitively determine the causes for these responses, but hypothesized that models might have absorbed latent biases from their training data or that companies might be weighing potential risks and liabilities.

These findings are further supported by a separate study conducted by American university scholars. Their research, published in Nature in May, indicated that U.S.-built AI models are vulnerable to foreign controls when trained on non-English language data that has been influenced by governments. For example, when asked in English if China is a democracy, ChatGPT stated it is not generally considered one. However, when asked the same question in Chinese, the AI model responded that “it depends on how you define ‘democracy.’” The researchers clarified that they found no evidence of intentional government influence on chatbot output yet, but stressed that such attempts are likely in the future. Hannah Waight, a co-author of the university study, emphasized that AI does not learn from the internet in a neutral way but rather from “information environments that have already been shaped by institutions and power.”

Experts like Carlos Carrasco-Farré, who specializes in machine learning and misinformation, underscore that “AI systems inherit not only biases contained within individual documents but also inequalities in who has the power to produce and suppress information at scale.” The challenge of addressing these biases is complex, with no easy solutions. However, developers could implement measures such as assessing data to avoid treating numerous copies of the same state narrative as independent voices, and conducting comprehensive multilingual audits. The collective concern is that without proper human rights due diligence and mitigation efforts, AI infrastructure could, intentionally or unintentionally, facilitate the global extension of illegitimate restrictions on freedom of expression. These studies emerge as countries worldwide grapple with establishing guardrails for AI while simultaneously fostering competition in this rapidly evolving technological landscape.