Stanford Institute for Human-Centered Artificial Intelligence: 2025 AI Index Report (Key Takeways)

Dr. Chen-Tso (Wesley) Chu
Visiting Scholar at Stanford Law School; Secretary-General, International Artificial Intelligence and Law Research Foundation

🔹 Public sector still lags behind industry in frontier AI development
Despite increased public investment—$831 million by the U.S. government in 2023—the private sector continues to dominate frontier AI model development. In 2024, nearly 90% of major AI models came from the private sector, up from 60% in 2023.

🔹 AI legislation slows at the federal level but accelerates in U.S. states
In 2024, U.S. federal agencies issued 59 AI-related regulations (a doubling from 2023) across 42 institutions. However, state-level legislation saw faster growth, with the number of state-enacted AI laws more than doubling compared to a 29.2% increase at the federal proposal level.

🔹 Open-source models are rapidly closing the gap with closed-source models
Competition in AI is intensifying. The performance gap between the top-ranked and 10th-ranked model narrowed from 11.9% to 5.4% over the past year. As of February 2025, the performance gap between open and closed models had shrunk to just 1.70%.

🔹 The U.S. retains AI leadership, but the U.S.-China gap is narrowing
The U.S. leads in key indicators, producing 40 top AI models, compared to 15 from China and 3 from Europe. However, performance differences on benchmark tests between Chinese and American models are now mostly under 10 percentage points.

🔹 Technical progress outpaces evaluation standardization
With rapid advancement, researchers are proposing new benchmarks (e.g., Humanity’s Last Exam, FrontierMath, BigCodeBench). However, many benchmarks suffer from poor design, highlighting the urgent need for standardized evaluation methods.

🔹 China leads in AI research volume, the U.S. in high-impact research
In 2023, China led in both publication share (23.2%) and citation share (22.6%). Yet U.S. institutions produced the majority of the top 100 most-cited AI papers over the past three years.

🔹 China dominates industrial robotics
China installed 276,300 industrial robots in 2023—six times more than Japan and 7.3 times more than the U.S. Its market share has climbed from 20.8% in 2013 to 51.1%, exceeding the rest of the world combined.

🔹 Greater China leads in AI adoption growth
North America remains the most advanced in enterprise AI usage, but the Greater China region recorded the highest growth—up 27 percentage points year-over-year, followed by Europe with a 23-point increase.

🔹 AI boosts productivity, but most companies see <5% revenue growth
Among AI-adopting companies:

• 49% saved costs in service operations

• 43% in supply chain

• 41% in software engineering
  However, most cost savings were under 10%, and revenue gains—71% in marketing/sales, 63% in supply chain—were mostly under 5%.

🔹 AI’s computational and energy demands are soaring
AI training compute doubles every 5 months; dataset sizes double every 8 months. Energy consumption for training large models continues to rise annually.

🔹 Nuclear energy returns to the spotlight as AI reshapes power demands
To support AI workloads, Microsoft is investing $1.6 billion to restart the Three Mile Island nuclear plant. Google and Amazon also signed nuclear power deals, signaling a shift in tech giants’ energy strategies.

🔹 Hardware performance improves in speed, cost-efficiency, and energy use
Based on 16-bit floating point computation:

• Machine learning hardware performance grows 43% annually

• Price-performance improves 30% annually

• Energy efficiency grows 40% yearly

🔹 AI outperforms physicians in clinical diagnosis tasks
GPT-4 outperforms human doctors in diagnosing complex clinical cases—even when physicians use AI assistance. AI also excels at cancer detection and high-risk patient identification.

🔹 FDA-approved AI medical devices surge
Since the first FDA approval in 1995, only 6 AI devices were approved by 2015. By 2023, this number surged to 223.

🔹 Americans remain wary of self-driving cars
A 2024 AAA survey found 61% of Americans fear self-driving cars, and only 13% trust them. Although down from a 68% peak in 2023, distrust remains higher than 54% in 2021.

🔹 U.S. CS teachers want to teach AI but feel unprepared
81% of computer science teachers believe AI should be part of core CS curricula, yet fewer than half feel qualified to teach it.

🔹 Master’s degree holders in AI nearly doubled from 2022 to 2023 in the U.S.
While bachelor’s and PhD growth was modest, AI-related master's degrees saw a sharp rise.

🔹 Workers expect job changes from AI, but few fear displacement
Globally, 60% of respondents believe AI will change work within 5 years, but only 36% worry about being replaced by AI in that timeframe.

Source: Stanford HAI – 2025 AI Index Report (https://hai-production.s3.amazonaws.com/files/hai_ai_index_report_2025.pdf)