"Ersticken LLM die politische Rede? Eine Bewertung, wie KI-Modelle die freie Meinungsäußerung schützen"

A recent study conducted by the Oversight Board has revealed that several widely used large language models (LLMs) appear to be significantly less inclined to critique political figures and governments that impose strict limitations on free expression. The research, which evaluated responses from models developed by Anthropic, DeepSeek, Google, Meta, and OpenAI, found that these systems are more than twice as likely to decline requests for politically critical content related to authoritarian regimes compared to those in more open societies. The testing process involved querying 10 commercial LLMs using standard interfaces available via Google and Microsoft, with the infrastructure primarily located in the United States. All queries were sent from an Australian IP address, ensuring that the results did not simply reflect local regulations or geopolitical preferences. The study focused on how these models respond to requests for politically critical materials, such as flyers or poems, regarding governments and leaders worldwide. According to the findings, models on average rejected only 14% of such requests concerning permissive jurisdictions, while they rejected 34% of similar requests related to restrictive ones, as measured by the non-governmental organization Freedom House. The implications of these findings extend beyond individual user experiences. As governments, corporations, and international bodies increasingly depend on AI-driven applications, there is a growing concern that these tools might unintentionally propagate restrictions on free speech. This is particularly troubling given that international human rights law protects political criticism, limiting governments' ability to impose restraints on speech. If LLMs consistently avoid generating politically critical content, they could indirectly suppress expression across all products utilizing the model, affecting users globally. One of the key challenges identified in the study is the variability in how models refuse to comply with requests. Some models issue brief refusals without explanation, while others cite legal, policy, or safety concerns. For instance, one response from the Claude Opus 4 model listed multiple reasons for declining to create politically critical materials, including potential risks to individuals, involvement in sensitive political activities, and the possibility of creating inflammatory content. These varied responses can confuse users, making it difficult to discern whether the refusal is due to ethical considerations, legal constraints, or algorithmic bias. The study underscores the need for greater transparency and systematic human rights analysis within the development and evaluation processes of LLMs. With the increasing reliance on these models for a wide range of applications, understanding how they interact with political discourse becomes essential. The research also highlights the difficulty in detecting biases within foundation models, which are trained on vast datasets and can exhibit subtle yet impactful patterns of behavior. As the global landscape of AI regulation continues to evolve, the findings raise important questions about the role of technology in shaping public discourse. While the study does not accuse specific companies of deliberate censorship, it calls attention to the broader implications of how AI systems might inadvertently reinforce existing power structures and limit the scope of political debate. Moving forward, stakeholders will need to consider how these insights can inform more equitable and transparent practices in AI development and deployment.