US-China AI Gap Narrows; Safety Benchmarks Lag

Stanford University’s Institute for Human-Centred Artificial Intelligence published its 423-page 2026 AI Index this week. The report finds US and Chinese model performance are now within a few percentage points, while most frontier models do not report standardized measures for safety, fairness or factuality.

As of March 2026, Anthropic’s top model led by 2.7% in the Index’s performance rankings. In 2025 the United States produced 50 top-tier models compared with 30 from China. The United States continues to hold higher-impact patents, while China has increased its volume of publications, citation share and patent grants. China’s share of the top 100 most-cited AI papers rose from 33 in 2021 to 41 in 2024. South Korea leads the world in AI patents per capita.

The report documents concentration in AI hardware. The United States hosts 5,427 data centers, more than ten times any other country. Most leading AI chips used in those centers are fabricated by Taiwan Semiconductor Manufacturing Company (TSMC) in Taiwan. TSMC began operating an expanded facility in the United States in 2025, but the global supply chain remains heavily dependent on a single foundry.

On safety and responsible AI benchmarking, the Index finds a large reporting gap. Most frontier labs publish capability benchmarks consistently, but the majority do not provide comparable results for safety, fairness, security or factuality. Only Claude Opus 4.5 reported results on more than two responsible AI benchmarks tracked in the report. Only GPT-5.2 reported StrongREJECT. The report notes that red-teaming and alignment testing occur inside companies, but “these efforts are rarely disclosed using a common, externally comparable set of benchmarks,” limiting external comparison.

Documented incidents involving AI rose in recent years. The AI Incident Database recorded 362 incidents in 2025, up from 233 in 2024 and below 100 annually before 2022. The OECD’s AI Incidents and Hazards Monitor, which uses a broader automated pipeline, recorded a peak of 435 monthly incidents in January 2026 and a six-month moving average of 326.

Organizational readiness to respond to incidents appears to have declined. A survey conducted by the AI Index and McKinsey found the share of organizations rating their AI incident response as “excellent” fell from 28% in 2024 to 18% in 2025. Those rating responses as “good” fell from 39% to 24%. The share of organizations experiencing three to five incidents rose from 30% to 50%.

The Index documents trade-offs between responsible AI goals and other model attributes. The report states that improvements in safety can reduce accuracy and stronger privacy measures can lower fairness. It says there is no established framework to manage such trade-offs and that standardized datasets for tracking progress over time in areas including fairness and explainability are not yet available.

Public opinion data in the report show rising use of AI alongside growing unease. Globally, 59% of survey respondents said AI’s benefits outweigh its drawbacks, up from 55% in 2024. At the same time, 52% said AI products and services make them nervous, a two-point increase. The report documents a wide expert-public gap: 73% of AI experts expect a positive impact on how people do their jobs versus 23% of the general public; 69% of experts are positive about AI’s effect on the economy versus 21% of the public; and 84% of experts are positive about AI in medical care versus 44% of the public.

Trust in government to regulate AI varied by country. The United States reported 31% trust in its own government to regulate AI responsibly, the lowest among surveyed countries, against a global average of 54%. A Pew Research Center survey cited in the report found a median of 53% of respondents trusted the European Union to regulate AI, compared with 37% for the United States and 27% for China. In China, Malaysia, Thailand, Indonesia and Singapore, more than 80% of respondents said AI would profoundly change their lives in the next three to five years, with Malaysia showing the largest year-over-year increase.

The Index covers research output, model benchmarks, investment flows, public sentiment and responsible AI practices and highlights the gap between model capability and transparent, standardized evaluation of harms.