Silenced by Design Censorship, Governance, and the Politics of Access in Generative AI Refusal Behavior

Refusal behavior in AI systems is not neutral: The paper argues that when AI systems refuse to answer questions or perform tasks, these refusals are not simply objective safety measures but rather reflect institutional priorities and power dynamics shaped by companies' risk management strategies.

Current refusal design lacks transparency: Decision-making about what AI systems should refuse remains opaque, meaning users and the public have limited visibility into who decided what content should be blocked and why, raising concerns about accountability in AI governance.

Refusal mechanisms can function as a form of censorship: By drawing on historical frameworks of censorship, the paper suggests that AI refusals warrant scrutiny similar to other gatekeeping practices, particularly regarding how they may limit access to information or services for different user groups.

User-centered design is needed: The paper recommends that ethical refusal design should prioritize input from affected users rather than being determined solely by corporate risk assessments, suggesting that responsible AI governance requires more inclusive decision-making processes.

AI refusal raises broader governance questions: The analysis positions refusal behavior as a critical lens for examining how power operates in AI systems and highlights the need for clearer governance frameworks that balance safety concerns with transparency and democratic accountability.