LLM Content Policies: How AI Companies Decide What You Can Ask
Every major AI language model has extensive content policies governing what it will and won't discuss. These policies are made by private companies, applied globally, and have enormous influence on what information and perspectives billions of users can access through AI systems.
Large language models — AI chatbots like ChatGPT, Claude, and Gemini — are trained not only on vast datasets but also on human feedback that shapes what they will and won't say. The companies that build these systems make complex decisions about speech policy that affect billions of users.
Current content policy approaches:
Hard restrictions: All major LLMs refuse to produce certain categories of content: instructions for creating weapons of mass destruction, child sexual abuse material, and a handful of other clearly harmful categories. These restrictions have broad support.
Political and controversial content: Most major LLMs attempt to avoid political positions, refuse to express opinions on contested political questions, and try to present 'both sides' of controversial issues. The implementation of these policies is inconsistent and arguably non-neutral — critics on both left and right claim their perspectives are treated differently.
Medical and legal information: LLMs often add extensive disclaimers or refuse to provide certain medical and legal information out of liability concern, even when the information is publicly available. This creates information inequality — users who can afford professional advice can get it; others are deflected.
Sensitive historical topics: Policies on discussions of violence, genocide, and sensitive historical events vary in ways that critics argue reflect cultural biases.
The policy process is non-transparent. Companies do not publish the full criteria used to train their models' values. Users cannot know what perspectives have been systematically downweighted in training.
From a First Amendment perspective, AI companies have the right to set their content policies as private actors. The question is whether the companies making these decisions understand the speech consequences, exercise their discretion consistently, and provide sufficient transparency to allow meaningful public scrutiny of decisions with such enormous influence on public discourse.