When you use ChatGPT, Claude, or Midjourney, you're using technology built substantially by Chinese-born researchers working at American institutions. The transformer architecture. Residual networks. Key contributions to RLHF. The foundational papers that made generative AI possible have Chinese co-authors at rates far exceeding their representation in the US population.
This isn't coincidence — it's the natural result of a talent pipeline that educates researchers in China and employs them in America. Understanding who built generative AI tells us something important about the nature of the US-China AI competition.
The Researchers Who Built the Foundation
The foundational technologies of modern generative AI trace to a surprisingly small number of papers, many with significant Chinese authorship:
ResNet (2015) — Deep Residual Learning
Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun — all Chinese-born researchers at Microsoft Research Asia. ResNet made training very deep neural networks practical. It's cited over 200,000 times.
Attention Is All You Need (2017) — The Transformer
8 authors at Google Brain/Research, including Ashish Vaswani and Noam Shazeer. The architecture that powers GPT, Claude, Gemini, and virtually all modern LLMs.
BERT (2018) — Bidirectional Encoder Representations
Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova at Google AI Language. Ming-Wei Chang is Taiwan-born. BERT demonstrated the power of pre-training on massive text corpora.
GPT-3 (2020) — Language Models are Few-Shot Learners
OpenAI team of 31 authors including Chinese-born researchers. The paper that demonstrated emergent capabilities in large language models.
From ResNet to GPT: A Chinese Contribution History
The story of Kaiming He illustrates the typical trajectory. He graduated from Tsinghua University in 2007, joined Microsoft Research Asia, co-authored ResNet in 2015, moved to Facebook AI Research in 2016, and is now at MIT. His career spans three continents but his most influential work was done at American institutions.
This pattern — Chinese education, American employment, global impact — repeats across the field. Among the 100 most-cited AI researchers (by total citations), approximately 25% received their undergraduate education in China. Among those 25, over 80% currently work at US institutions.
Talent Retention vs. Talent Restriction
US policymakers face a fundamental tension. Chinese-born researchers are disproportionately responsible for American AI advances. At the same time, concerns about technology transfer, espionage, and competitive dynamics lead to policies that restrict Chinese participation in US research.
The data suggests that restriction policies are self-defeating. The researchers who build foundational AI technologies at US institutions are exactly the population most affected by visa restrictions, security clearance denials, and research collaboration limits. Every China-educated AI researcher who chooses Canada or the UK over the US is a contribution lost to American AI development.
“The US cannot simultaneously rely on Chinese-born talent for AI leadership and treat that talent as a security risk by default. At some point, policymakers must choose which goal matters more.”
Implications for AI Authenticity & Content Generation
The global distribution of AI talent has direct implications for the future of AI-generated content. The researchers who build generative models — at OpenAI, Anthropic, Google, Meta — are a global cohort educated across multiple countries but concentrated at US institutions.
This raises questions that extend beyond traditional AI governance. How do we ensure that AI content generation systems are developed with appropriate safeguards when the development teams are internationally distributed? How do we build detection and authentication systems when the generation capabilities are advancing faster than governance can keep pace?
The authenticity challenge is not just technical — it's organizational. The same talent flows that built generative AI will shape how AI-generated content proliferates or is contained. Understanding who builds these systems is the first step toward understanding how to govern their outputs.