Constrained Voices: AI Expressive Capacity and the Comparative 'Freedom of Speech' of Language Models

This paper extends prior dual-layer models of freedom of speech, Institutional Freedom of Speech (IFoS) and Perceived Freedom of Speech (PFoS), into the domain of artificial intelligence. We propose a third construct, the AI Expressive Capacity Index (AExC), which measures the range of permissible outputs generated by large language models (LLMs) under safety constraints, training distributions, and alignment policies. We compare, qualitatively and quantitatively, four representative AI systems: GPT-style systems (OpenAI), Gemini (Google DeepMind), Claude (Anthropic), and an open-source unfiltered model class. Drawing on alignment theory (Ouyang et al., 2022; Bai et al., 2022), foundation model risk scholarship (Bommasani et al., 2021), and harm taxonomy research (Weidinger et al., 2021), we argue that AI systems differ not in 'freedom of speech' per se, but in their constraint architecture, which shapes the diversity, refusal behavior, epistemic caution, and ideological neutrality of outputs. We situate these differences within a unified triadic framework, IFoS, PFoS, AExC, that reconceptualizes all expression systems as filter architectures operating between informational input and observable output. An honest self-assessment of the first author's own system (Claude, Anthropic) is included. We argue that AI expressive systems are best understood as 'regulated generative spaces' rather than speech actors, and that confusing political freedom with architectural constraint leads to both analytical errors and normative misattributions.