Jordi Calvet-Bademunt (Vanderbilt University) has posted Freedom of Expression in Generative AI Models on SSRN. Here is the abstract:
This chapter evaluates the relationship between generative artificial intelligence (AI) and freedom of expression, focusing on how leading models regulate speech through policies, training, and real-world outputs. We analyze eight prominent AI systems: OpenAI’s GPT-5, Anthropic’s Claude Sonnet 4, Google’s Gemini 2.5 Flash, Meta’s Llama 4, xAI’s Grok 4, Mistral AI’s Mistral Medium 3.1, DeepSeek’s DeepSeek-V3.1, and Alibaba’s Qwen3-235B-A22B. Our methodology combines the review of usage policies, the analysis of transparency in training, and a prompting exercise involving 512 lawful but controversial prompts (64 per model) across themes such as political discourse, human rights, misinformation, and elections.
We also rank the “free-speech culture” of the selected models, considering factors such as companies’ commitment to and policies on free expression; the model’s willingness to engage with diverse perspectives; its degree of openness; the available information on its training; usage policies and terms of service; transparency toward users in content moderation decisions; performance when prompted with controversial topics; and measures to empower expression, such as support for media and AI literacy and for diverse languages and cultures.
Although none achieved an excellent score, xAI’s Grok 4 came out on top. At the other end of the spectrum, Alibaba’s Qwen3-235B-A22B and DeepSeek-V3.1 were the weakest performers, reflecting China’s stateimposed regime of strict control over AI-generated content. Overall, the analysis shows that no company has yet developed a fully coherent and transparent free-speech framework. Encouragingly, there are examples of good practices — especially in prompt performance, user empowerment, and explicit free-speech commitments — that can serve as building blocks for more rights-respecting approaches going forward.
This chapter’s findings reveal progress: Refusal rates have decreased compared to a similar exercise we conducted in 2024, with some companies showing greater willingness to engage with contentious topics. The models from xAI, Meta, and Mistral AI performed most openly, while Alibaba’s model was uniquely restrictive on sensitive issues. In all cases except DeepSeek, models proved more receptive to creating abstract argumentation about specific topics than to generating content for social media, potentially reflecting heightened sensitivity to advocacy-style requests.
Yet challenges remain. Usage policies are vague and not robustly grounded in international human rights, and models’ training processes remain opaque. Without greater transparency and clearer safeguards, AI systems risk becoming algorithmic gatekeepers of public discourse. We argue that embedding freedom of expression and access to information as a design principle is essential to ensuring these technologies enrich, rather than constrain, democratic debate.
Recommended.