Stanford HAI Report: US-China AI Gap Nearly Closed, Safety Lags Behind

The Stanford HAI AI Index 2026 report demolishes the convenient myth of unconditional U.S. leadership in artificial intelligence: by the quality of cutting-edge models, China has nearly caught up with American companies, and in terms of public safety verification, the market still lags far behind its own ambitions.

In the 2026 AI Index report published by Stanford HAI on April 13, 2026, data on model performance, investments, scientific publications, patents, public sentiment, and responsible AI practices have been compiled. The main conclusion sounds unpleasant for the American industry: a sustainable technological lead for the U.S.

is no longer visible. According to Stanford, American and Chinese models have repeatedly swapped positions in the top rankings since the beginning of 2025. In February 2025, DeepSeek-R1 briefly matched the best American model, and by March 2026, the lead of Anthropic's top model had shrunk to 2.

7%. This does not mean the U.S.

has lost all advantages. American companies still release more top-tier models: 50 versus 30 for China in 2025. The U.

S. also has higher patent quality and incomparably greater private investment. But China already leads in the volume of AI publications, citation share, and number of issued patents.

In other words, the market is ceasing to be a story about one undisputed leader. Now it's a tight race where the lead can change after every major release. Stanford separately points to a structural vulnerability in American advantage.

The U.S. is home to 5,427 data centers—more than ten times as many as any other country.

But almost all cutting-edge AI chips for this infrastructure are manufactured by a single company, TSMC, on Taiwan. Expansion of TSMC production in the U.S.

has already begun in 2025, but the supply chain logic itself remains fragile: dominance in computing does not equal control over the entire technology stack. Even more importantly, the gap between model capabilities and how the industry measures risks is not shrinking but growing. Almost all frontier model developers publish results on capability benchmarks like MMLU or SWE-bench, but responsible AI data comes out fragmentarily.

Stanford's tables on safety, fairness, factuality, and human agency contain many empty cells, so external comparison of models by risks is often simply impossible. This does not mean laboratories are not engaged in internal red teaming or alignment testing. The problem is different: these checks are rarely disclosed in a unified and comparable format.

Against this backdrop, the number of documented AI incidents continues to rise. The AI Incident Database recorded 362 incidents in 2025 versus 233 a year earlier, and before 2022 such cases numbered less than a hundred per year. Organizations also do not appear ready for the scale of the problem.

In a joint survey by AI Index and McKinsey, the share of companies calling their response to AI incidents excellent fell from 28% to 18% over a year, and the share of those reporting three to five incidents rose from 30% to 50%. Additional complexity arises from the fact that improving one responsible AI metric can worsen another: for example, increasing safety can reduce accuracy, and strengthening privacy can harm fairness. Public perception is also becoming more contradictory.

Globally, 59% of respondents believe AI products bring more benefit than harm, but simultaneously 52% say such services cause them anxiety. Experts are noticeably more optimistic than ordinary users: 73% of specialists expect positive AI impact on work versus 23% of the U.S.

population, on the economy—69% versus 21%, on medicine—84% versus 44%. Meanwhile, the U.S.

has recorded the lowest level of trust in its own government on responsible AI regulation—31% compared to the global average of 54%. The conclusion from the report is simple: the debate over who is ahead today in model quality becomes less important than the question of who knows how to make AI verifiable, safe, and resilient to failures. The U.

S. maintains scale, capital, and infrastructure; China maintains pace and scientific-patent mass. But the real unclosed gap runs not between the two countries, but between the speed of model development and the maturity of control mechanisms around them.