April 21, 2026 Artificial intelligence is advancing rapidly across capability, adoption and economic impact, according to new data from Stanford HAI’s AI Index. The report shows frontier AI systems now match or exceed human performance on complex benchmarks while adoption has reached 88 per cent of organizations and over half the global population within just a few years.
The AI Index, widely used by policymakers and industry leaders, highlights that progress is being driven primarily by industry, which produced more than 90 per cent of notable frontier models in 2025. Performance gains have been sharp: on the SWE-bench Verified coding benchmark, model accuracy improved from 60 per cent to near 100 per cent in a single year, while top systems now perform at or above PhD-level in science reasoning and competition mathematics.
At the same time, global competition is tightening. The performance gap between U.S. and Chinese models has effectively closed, with leadership shifting multiple times since early 2025. As of March 2026, Anthropic’s leading model holds a narrow 2.7 per cent edge. The U.S. still leads in high-impact models and private investment, reaching $285.9 billion in 2025, while China dominates in research output, patents and industrial robot deployment. South Korea leads globally in AI patents per capita.
Infrastructure remains concentrated, as the United States hosts 5,427 data centres – more than 10 times any other country – and consumes the most energy for AI workloads. However, the hardware supply chain is heavily dependent on TSMC, which fabricates the majority of advanced AI chips, creating a single point of dependency despite recent U.S. expansion efforts.
The report also underscores what researchers describe as the “jagged frontier” of AI capability. Systems can achieve gold-medal performance in competitions like the International Mathematical Olympiad yet fail basic tasks such as reading analog clocks, with accuracy around 50 per cent. AI agents have improved significantly in real-world task environments, rising from 12 per cent to roughly 66 per cent success rates, but still fail about one-third of structured tasks.
While adoption is accelerating – generative AI reached 53 per cent population usage faster than both the internet and personal computers – responsible AI development is not keeping pace. Documented AI incidents rose to 362 in 2025, up from 233 the previous year, and reporting on safety benchmarks remains inconsistent. Research also indicates trade-offs, where improving safety can reduce accuracy.
Workforce and education trends reflect similar imbalance. More than 80 per cent of students use AI tools, but only half of schools have formal policies, and just 6 per cent of educators describe those policies as clear. Meanwhile, the U.S. continues to lead in AI entrepreneurship with 1,953 new companies in 2025, but its ability to attract global talent has declined sharply, with an 89 per cent drop in inbound AI researchers since 2017.
AI adoption is also uneven globally. Countries like Singapore (61 per cent) and the United Arab Emirates (54 per cent) exceed expectations, while the U.S. ranks 24th at 28.3 per cent. Despite this, U.S. consumers derive significant value, estimated at $172 billion annually, with median user benefit tripling between 2025 and 2026.
Public perception continues to diverge from expert opinion. While 73 per cent of AI experts expect positive impacts on work, only 23 per cent of the public agree. Trust in governance is similarly fragmented, with the U.S. reporting the lowest confidence in its own regulatory ability at 31 per cent, while the European Union ranks higher globally.
