How to Train Your Algorithm

For years, governments worried about the chips that powered machines. Now they worry about what the machines think. The same transformer architecture that fuels ChatGPT in the United States also drives China’s DeepSeek R1 and Doubao AI, but each model learns under a different rulebook. Where one optimizes for openness and creativity, the others are trained for caution and compliance. The difference is no longer in the hardware but in their curriculum.

Lessons in Obedience

When OpenAI released ChatGPT in 2022, it became the fastest-growing consumer application in history, reshaping classrooms and programming labs around the world. In mainland China, the chatbot remains officially unavailable, and access through VPNs is restricted. Chinese companies instead built their own conversational AIs. DeepSeek R1, approved by regulators in early 2025, and ByteDance’s Doubao, integrated into Douyin and Toutiao, operate under national rules that require generative models to “adhere to socialist values” and undergo security review. These rules come from the Interim Measures for the Management of Generative Artificial Intelligence Services, enacted in August 2023. They require companies to submit datasets for inspection and to filter outputs containing “false, harmful, or politically sensitive” content. In practice, engineers spend as much time refining keyword filters and response templates as they do improving reasoning accuracy. The model’s intelligence becomes a balance between probability and permission.

Different Teachers, Different Lessons

The United States has no equivalent approval system. Developers face voluntary standards on transparency and bias but no mandatory pre-screening before release. OpenAI, Anthropic, and Google DeepMind train their models to avoid hate speech and misinformation but not to conform to political ideology. This difference produces two kinds of engineers. In Beijing, teams focus on “secure alignment”—ensuring compliance with regulators before deployment. In California, teams test for factuality and fairness while keeping debate and critique within the model’s range of expression.

For students entering AI research, this divide feels like two versions of the same discipline. A university lab in Hangzhou might study how to tune a model to prevent sensitive keywords, while one at Stanford explores ethical transparency through open-source audits. Both call it alignment, but they mean different things.

The New Form of Restriction

Tech restrictions once meant chips, antennas, and app stores. Now they exist in the data and the logic that interprets it. When a government defines what an algorithm can say, control moves from hardware to language. DeepSeek and Doubao are technically impressive, capable of coding, summarizing, and generating essays, yet they avoid whole categories of conversation—from current politics to history—by design. The absence is not a flaw; it is a carefully designed feature. This is what the next phase of localization looks like: intelligence tailored to national context. The same neural architecture that learns fluidly across cultures can also be trained to stay within borders.

The Future of Training

For young engineers, the question is no longer just how to build smarter models but how to train them responsibly within systems that define responsibility differently. They will inherit not only the technology but also the values encoded within it. In the quiet race for alignment, each country is teaching its algorithms how to behave and, in doing so, teaching a generation of creators what innovation means.