In 2024, the United States led responsible AI research with 788 papers accepted at the major AI conferences. China followed with 322. In 2025, China led with 812 papers. The United States dropped to 394. The flip happened in twelve months.

RAI papers accepted at select AI conferences by geographic area, 2025:

The data point alone is striking. The structural pattern it reveals is more so.

The methodology: responsible AI papers accepted at six leading AI conferences (AAAI, AIES, FAccT, ICML, ICLR, NeurIPS) are counted using RAI-related keywords for identification. The 2025 totals show China at 812, United States at 394, Singapore at 112, United Kingdom at 103, Hong Kong at 98, Australia at 84, Germany at 68, South Korea at 57, Canada at 54, Italy at 29.

The reversal from 2024 to 2025 is not gradual. The US RAI paper count fell from 788 to 394, roughly halved. The Chinese count rose from 322 to 812, more than doubled. Both moves happened in twelve months. The pattern does not look like organic growth in the Chinese RAI research community; it looks like a step change.

What explains the flip? The available data does not directly answer the question, but several factors are visible in the broader field.

First, the cumulative volume picture shows the United States still holds the largest cumulative total of RAI papers accepted over the full 2019–2025 period. The 2025 reversal is a single-year inversion against a longer-term US lead. Whether the inversion persists into 2026–2027 will determine whether this is a one-year anomaly or the start of a sustained pattern.

Cumulative RAI papers accepted by geographic area, 2019–25:

Second, the broader Chinese AI publication trajectory has been climbing fast across most measures: total publications, citation share, frontier-model releases. RAI research as a sub-category is climbing alongside the wider trajectory, partly because Chinese AI policy explicitly emphasises responsible development as a strategic priority.

Third, the US RAI research environment in 2025 was disrupted by several factors visible in the broader field. The Biden-era US AI Executive Order was revoked in early 2025 and US federal AI policy shifted toward a more deregulatory, "innovation-first" approach. The US declined to sign the 2025 AI Action Summit declaration in France. These policy shifts did not necessarily reduce US RAI research output, but they coincided with the drop in the data, and the timing suggests possible mechanisms: reduced federal research funding, shifted academic research priorities, or migration of researchers to other jurisdictions.

The data does not separate these drivers cleanly. The observation is that the reversal happened, that it is large, and that it coincides with structural shifts in the policy environment. Whether it persists is the open question.

For strategic AI planning, three observations follow.

The first: the geography of responsible AI research is rebalancing. The 2024 picture (US dominance with European participation and growing Asian presence) has been replaced by a 2025 picture where the leading country is China, the next four are Singapore, UK, Hong Kong, and Australia, and the US is in second place. For organisations building responsible AI capacity that depends on tracking the research literature, the talent pool, or the institutional partnerships, the geographic shift matters. The networks of researchers, institutions, and conferences that produced the 2024 baseline are not the same networks that will produce the 2026–2028 outputs.

The second: this is a leading-edge indicator for the broader RAI ecosystem. Academic publication volume often leads policy adoption, regulatory frameworks, and commercial RAI infrastructure by 2–3 years. If the 2025 reversal persists, the implication is that Chinese responsible AI practices, frameworks, and standards will gain influence relative to Western ones over the next 2–5 years. For multinational organisations, the RAI regulatory landscape may shift in directions shaped by research output that is no longer concentrated in US institutions.

The third: the relative position of the next-tier countries (Singapore 112, UK 103, Hong Kong 98, Australia 84) is now substantial in absolute terms relative to the US (394). The "second tier" in 2025 looks more diversified than the 2024 picture suggested. For talent and partnership strategy, the implication is that responsible AI research capacity is genuinely distributed across more geographies than the previous narrative suggested.

The data observation: the responsible AI research community geography rebalanced sharply in 2025. China leads in single-year output for the first time. The US retains the cumulative lead but its position is now contested in a way it was not 12 months ago. The next four countries (Singapore, UK, Hong Kong, Australia) are more substantial relative to the leaders than the 2024 picture suggested. Whether this represents a one-year anomaly or a sustained pattern is open. The next 12–24 months of conference acceptance data will resolve the question. Until then, strategic plans built on the assumption of US dominance in responsible AI research are operating against current data that contradicts the assumption.
Sources
  • Primary: Stanford AI Index 2026, Chapter 3 (Responsible AI) 3.4 — hai.stanford.edu/ai-index/2026
  • Underlying data: AI Index 2026 own analysis — paper counts from AAAI, AIES, FAccT, ICML, ICLR, NeurIPS
  • Policy context: US Executive Order on AI revocation (Jan 2025); 2025 AI Action Summit declaration (Paris)