The Number of AI Research Publications is Accelerating Faster in China than the U.S.
Note: this post was completed as part of a written work trial for Epoch.
An extremely important input to transformative artificial intelligence timelines is research activity. In order to prioritize interventions, AI strategists need an idea of how quickly the AI research effort is accelerating. It is especially important to understand differential research growth rates between major actors.
One particularly decision-relevant differential is the relative amount of AI-related research activity occurring in the US versus China. These two countries drive the bulk of AI progress, so from a governance standpoint it’s quite salient which is more likely to achieve AI dominance over the next few decades. For example, if the AI safety community makes progress in the US landscape, but most research is happening (in a risky way) in China, then global risk will remain high.
Forecasters remain highly uncertain about the relative growth of AI in the US versus China. While the US leads China in private investment and new startups founded, China leads the US significantly in the number of AI-related publications per year.
To investigate this latter trend, I gathered data from CSET’s AI Country Activity Tracker (CAT). CSET’s corpus “combines (and deduplicates) scholarship from six datasets — arXiv, China National Knowledge Infrastructure (CNKI), Digital Science, Microsoft Academic Graph (MAG), Papers With Code, and Web of Science,” all sourced from Dimensions. More about this corpus, including how AI-relevant papers were identified, is available here; for our purposes it represents a good (if imperfect) proxy for total AI publications.
The following plot presents the total number of AI publications for each country during the 11-year period from 2010–2020, and extrapolates this trend to 2030 using a quadratic regression model. 90% confidence intervals are also depicted.
China has been publishing significantly more AI-related papers than the US since 2010. This probably doesn’t indicate more actual AI research productivity in China since 2010, since the most major advances during that time occurred in the US.
The noteworthy trend, however, is that AI-related publication is accelerating faster in China than in the US, which could indicate that China is pulling ahead in underlying research activity, despite lower levels of private investment. Presumably, this would eventually translate to China outstripping the US in actual research productivity, and Chinese actors developing the most powerful models in the world.
AI timeline forecasters should monitor not only metrics which measure bottom-line ML performance and complexity, but also the factors that drive research progress. This is especially true for longer-range forecasts. In addition, a range of metrics should be considered, since — as in the case of private investment vs. paper publication — different metrics can tell different stories. It would be worthwhile to forecast more comprehensively which research communities strategists should expect to pull ahead.
It’s also worth mentioning that in order for this kind of analysis to be decision-relevant, strategists also need to model the relationship between research activity and advances in the field, for example between research publication rates and actual performance advances.
Overall, however, the story is that lower amounts of private AI investment in China and fewer new companies being founded does not necessarily suggest less R&D growth. To the contrary, the trend above suggests a tentative update towards increasing levels of Chinese AI dominance over the next decade.
Note: The code and data to produce the displayed plot are available here.
