Western and Chinese AI: Who decides what’s forbidden?


Jorge Costa Oliveira
Ask an artificial intelligence chatbot a difficult question, and its refusal may tell you as much about its creators as any answer could. Large language models do not operate outside politics. Their boundaries reflect the laws, commercial interests and political values of the societies in which they are developed and deployed.
The comparison between Western and Chinese systems is particularly revealing. Both restrict harmful content. But the purposes and scope of those restrictions differ, with consequences for freedom of expression and the future of public debate.
Western models generally emphasize preventing harm and managing legal and reputational risks. They restrict assistance with child sexual exploitation, violence, fraud, dangerous weapons and other criminal activities. Privacy protections limit disclosures of personal information, while safeguards address harassment, hateful abuse and certain forms of manipulation.
Other boundaries concern copyrighted material, sexually explicit content and professional advice. These systems may provide medical, legal or financial information while qualifying their answers and discouraging reliance on them as substitutes for qualified professionals. Policies vary across providers, products and jurisdictions; there is no single Western rulebook.
These restrictions deserve scrutiny. Corporate definitions of safety can be inconsistent, overly broad or politically contentious. Models can reproduce biases, misrepresent disputed issues and refuse legitimate requests. Transparency remains incomplete. Still, users can generally question governments, examine historical controversies and criticize political leaders without those subjects automatically becoming forbidden territory.
Chinese services share many safeguards against criminal or harmful content. However, systems operating under China’s regulatory framework face additional requirements linked to state security, political stability and ideological conformity.
The resulting restrictions can reach far beyond preventing immediate harm. Critical discussion of Communist Party leadership, contested historical events, human rights and territorial disputes may prompt refusals, evasions or answers that reproduce official positions.
The 1989 Tiananmen crackdown, policies in Xinjiang, Hong Kong protests and Taiwan’s status illustrate the sensitivity of this terrain. Religion, political satire and criticism of senior leaders can also encounter restrictions. Their application varies, however, and a model’s behavior may differ between a domestic service and a version deployed elsewhere.
That distinction matters. A model’s national origin does not, by itself, explain every refusal. Training, later adjustments, service policies and deployment conditions all shape what users receive. Sweeping claims that every Chinese model blocks every sensitive reference obscure how these controls actually operate.
Nevertheless, the central political difference remains. In Western systems, restrictions are usually presented as protections for individuals and society. In Chinese domestic services, those protections coexist with obligations to preserve the authority and narratives of the party-state.
This gives censorship a new interface. Instead of a blocked website or a missing newspaper article, users encounter a conversational assistant that sounds helpful, authoritative and complete. What it leaves unsaid may be difficult to recognize, especially when an official narrative arrives dressed as a neutral explanation.
As chatbots become everyday tools for research, education and journalism, their omissions will shape public understanding alongside their errors. Digital literacy must therefore include asking who defines a model’s boundaries, which topics trigger restrictions and whether alternative accounts remain accessible. These are questions of public accountability, not merely matters of product design.
No chatbot should be mistaken for an independent arbiter of truth. Users need transparent policies, meaningful scrutiny and access to competing sources. The crucial question is not simply whether artificial intelligence can answer us. It is who decides which questions deserve an answer, and whose interests that decision ultimately serves.
linkedin.com/in/jorgecostaoliveira
Leave a reply
You must be logged in to post a comment.





















