What did the US government-backed study say about China’s AI?

The National Institute of Standards and Technology (NIST) found that China’s DeepSeek V4 Pro is about eight months behind leading US AI models. The evaluation was conducted by its CAISI unit. This suggests a measurable but not massive gap in capability.

Why are experts questioning the study’s conclusions?

Critics argue the methodology relied on private, unverifiable datasets and selective comparisons. Only one US model was included in cost comparisons. This raises concerns about transparency and fairness in the evaluation.

What is DeepSeek V4 Pro and how does it compare to US models?

DeepSeek V4 Pro is one of China’s most advanced AI models. On public benchmarks, it performs close to leading US systems in areas like coding, math and scientific reasoning. This suggests the gap may be smaller than claimed.

What does the Stanford AI Index say about the US-China AI gap?

Stanford University’s 2026 AI Index estimates the gap at just 2.7% on key leaderboards. This indicates rapid convergence between US and Chinese models. It contradicts the idea of a widening gap.

What is Item Response Theory (IRT) and why does it matter here?

IRT is a statistical method used to estimate overall model capability based on which problems it solves. It weights difficulty rather than using simple accuracy scores. However, results depend heavily on the chosen benchmarks.

Why do private benchmarks create controversy?

Private benchmarks cannot be independently verified or reproduced by other researchers. This makes it difficult to validate results. Transparency is a key issue in AI evaluation.

How big is the actual performance gap between US and Chinese AI?

Estimates vary depending on methodology. Some analyses suggest a meaningful gap, while others show a narrow difference. The true gap is uncertain and depends on how performance is measured.

What does this mean for global AI competition?

The debate shows that the US-China AI race is tightening. China appears to be catching up faster than previously expected. This increases competition in both technology and policy.

Image for illustrative purposes only. Not a real photo.

US AI gap with China questioned after study

Written by Mahathir BayenaPublished May 5, 2026, 4:02 AMLast Updated May 7, 2026, 2:49 AM

A US government-backed evaluation claims China’s leading AI model lags the frontier by eight months, but experts are challenging the methodology and conclusions.

The study by National Institute of Standards and Technology’s CAISI unit ranked China’s DeepSeek V4 Pro behind top US models using a statistical scoring system across multiple benchmarks.

However, critics argue the analysis relied partly on private, unverifiable datasets and filtered comparisons to just one US model, raising questions about fairness and transparency.

Public benchmark data paints a different picture, with DeepSeek performing close to leading US models in areas like scientific reasoning, coding and mathematics.

According to Stanford University’s 2026 AI Index, the performance gap between US and Chinese models has narrowed to just 2.7% on key leaderboards.

Some analysts say the gap is shrinking rather than widening, with competing indices showing Chinese models approaching US capabilities more quickly than official assessments suggest.

The debate highlights growing uncertainty over how to measure AI progress, as governments and researchers compete to define leadership in a rapidly evolving field.