AI capability is not plateauing — it is accelerating
Models now meet or beat human baselines on PhD science, multimodal reasoning, and competition math.
AI scaled faster than the systems around it could adapt
The 2026 AI Index — an independent Stanford HAI initiative — tracks AI across research, performance, responsibility, economy, science, medicine, education, policy, and public opinion. A few numbers that set the scene:
88% organizational AI adoption
53% GenAI population adoption in 3 yrs
286US private AI investment (US$ bn, 2025)
5,427US data centers — 10× any other country
362AI incidents logged in 2025 (was 233)
90% of notable frontier models from industry
The 15 Top Takeaways
The report's own headline findings. Tap any card for the full finding and its numbers.
The U.S.–China model performance gap has effectively closed
The two have traded the lead repeatedly; as of March 2026 the top U.S. model leads by just 2.7%.
The U.S. hosts the most data centers; one Taiwan foundry makes the chips
5,427 U.S. data centers — 10× any other country — and almost every leading AI chip comes from TSMC.
A gold medal at the IMO, but it cannot reliably tell time
The 'jagged frontier': Gemini Deep Think won IMO gold, yet the top model reads analog clocks just 50.1% of the time.
Robots still fail at most household tasks
Just 12% success on household tasks, even as software simulations reach 89.4% on RLBench.
Responsible AI is not keeping pace with capability
Safety benchmarks lag and incidents rose sharply — 362 documented in 2025, up from 233.
The U.S. leads investment, but its pull on global talent is fading
$285.9B private investment — 23× China — yet researchers moving to the U.S. fell 89% since 2017.
Adoption is spreading at historic speed — value often delivered for free
53% population adoption in 3 years (faster than PC or internet); U.S. consumer value hit $172B/yr.
Productivity gains appear where entry-level jobs are starting to shrink
14–26% gains in support and software dev; U.S. developers aged 22–25 saw employment fall nearly 20%.
AI's environmental footprint is expanding alongside its capabilities
Grok 4 training ≈72,816 tons CO₂e; data-center power hit 29.6 GW — comparable to New York at peak.
AI for science can outperform humans — but bigger isn't always better
Frontier models beat human chemists on ChemBench, yet score below 20% on astrophysics replication.
AI is transforming clinical care, but rigorous evidence stays thin
Note-writing tools cut documentation time up to 83%, yet only 5% of clinical AI studies use real patient data.
Formal education lags AI, but people learn AI at every stage of life
Over 80% of U.S. students use AI for school, yet just 6% of teachers say their school's AI policy is clear.
AI sovereignty is becoming a defining feature of national policy
National strategies are expanding fastest among developing economies; open source is redistributing who participates.
Experts and the public see very different AI futures
On jobs, 73% of experts expect a positive impact vs. just 23% of the public — a 50-point gap.
Nine chapters, in brief
Each chapter's headline highlights. Expand to read the key numbers.
1 · Research & Development
Industry now accounts for over 90% of notable models, and the most capable systems are the least transparent — training code, parameter counts, and dataset sizes are increasingly withheld. Compute has grown ~3.3× per year since 2022 to 17.1M H100-equivalents, with Nvidia supplying over 60%. China leads in publication volume, citations, and patent grants; the U.S. led notable model development (59 vs. 35). Gender gaps in AI talent remain entrenched, with no meaningful progress since 2010.Full deep dive: Research & Development →
2 · Technical Performance
Benchmark scores rose across language, reasoning, coding, and math — but evaluations are being outpaced by the progress they measure, and benchmarks are saturating. The gap between top models is shrinking, and the U.S.–China distance has closed almost completely. AI agents still fail roughly 1 in 3 attempts. Robots struggle in unstructured environments, while autonomous vehicles reached mass-scale deployment with promising safety records.Full deep dive: Technical Performance →
3 · Responsible AI
RAI benchmarking is increasing but not keeping up with deployment. Documented incidents rose to 362 in 2025 (from 233). Models struggle to separate knowledge from belief — hallucination rates span 22%–94%. Organizations are formalizing RAI: the share with no RAI policy fell from 24% to 11%, while top obstacles are knowledge gaps (59%), budget (48%), and regulatory uncertainty (41%). AI works best in English, and the gap is wider than global benchmarks suggest.Full deep dive: Responsible AI →
4 · Economy
Global corporate AI investment more than doubled in 2025; private investment grew 127.5%, with generative AI capturing nearly half of all private AI funding. The U.S. committed 23× more private investment than China. Organizational adoption rose to 88%, and consumer surplus from GenAI reached $172B annually. Productivity gains are largest in structured work (14–50%). China installed 54% of the world's industrial robots.Full deep dive: Economy →
5 · Science
AI scientific publications in the natural sciences reached ~80,150 in 2025 (+26%), now 5.8–8.8% of output (up from below 1% in 2010). Frontier models beat human chemists on ChemBench but score below 20% on astrophysics replication. Astronomy released its first foundation model (AION-1). An AI system (Aardvark Weather) ran a full weather-forecasting pipeline end-to-end for the first time. On end-to-end research tasks, the best agents score about half of PhD experts.Full deep dive: Science →
6 · Medicine
Clinical note-writing tools saw broad adoption — physicians reported up to 83% less time on notes and one system a 112% ROI. The FDA authorized 258 AI medical devices in 2025, mostly via modification pathways; only 2.4% of devices with clinical studies used randomized-trial data. A multi-agent system scored 85.5% on complex published cases vs. 20% for unaided physicians. AI-generated summaries now top 84%–92% of health-related Google searches.Full deep dive: AI in medicine trends →
7 · Education
CS enrollment fell 11% at U.S. four-year universities (2024–25), but AI master's graduates rose 17%. Four in five U.S. high-school and college students use AI for schoolwork, yet only half of schools have AI policies and just 6% of teachers call them clear. Over 90% of countries now offer CS in schools; China and the UAE mandated AI education from 2025–26. New AI PhDs in the U.S./Canada rose 22%, all of it going to academia.Full deep dive: Education →
8 · Policy & Governance
National AI strategies are expanding fastest among countries that had none five years ago; over half of 2024's new strategies came from emerging economies. AI sovereignty is becoming a central principle, but infrastructure is uneven — Europe and Central Asia grew state-backed supercomputing clusters from 3 to 44 (2018–2025). AI-related witnesses in U.S. congressional hearings grew twentyfold since 2017 (5 → 102). U.S. public AI investment (~$20.4B over 2013–24) is dwarfed by $285.9B private in 2025 alone.Full deep dive: Policy & Governance →
9 · Public Opinion
AI optimism is rising (benefits > drawbacks: 55% → 59%), but so is anxiety (52% feel nervous). Southeast Asia is most optimistic — over 80% in Malaysia, Thailand, Indonesia, and Singapore expect AI to change their lives. Experts and the U.S. public diverge sharply: on jobs, 73% vs. 23%. Nearly two-thirds of Americans (64%) expect fewer jobs over 20 years. The U.S. reported the lowest trust in its own government to regulate AI, at 31%.Full deep dive: Public Opinion →
In the report's own words
Lines that capture the year, from the co-chairs' message and the top takeaways.
The data does not point in a single direction. It reveals a field that is scaling faster than the systems around it can adapt.
At the technical frontier, leading models are now nearly indistinguishable from one another. Open-weight models are more competitive than ever.
AI can win a gold medal at the International Mathematical Olympiad but cannot reliably tell time — the jagged frontier of AI.
Generative AI reached 53% population adoption within three years — faster than the personal computer or the internet.
On how people do their jobs, 73% of experts expect a positive impact, compared with just 23% of the public — a 50-point gap.
About this page
What it is, where the numbers come from, and how to read them.
This is a non-official, educational reformatting of the Stanford HAI Artificial Intelligence Index Report 2026 — the ninth edition of an independent initiative at the Stanford Institute for Human-Centered AI. Every figure on this page is drawn directly from the report's Top Takeaways and per-chapter highlights.
How to read the numbers
- Benchmarks are saturating and frontier labs disclose less — the report cautions that independent testing does not always confirm what developers report.
- Many figures are estimates (e.g. compute, emissions, consumer surplus) and carry meaningful uncertainty.
- Hero counters round to whole numbers; exact figures (e.g. US$285.9B, 28.3%) appear in the cards and charts.
- 'Notable models', 'frontier', and similar terms follow the report's own definitions.
Design inspiration was drawn from editorial data-journalism (The Pudding) and clean chart aesthetics (Datawrapper); the layout and code are original. The report itself is licensed CC BY-ND 4.0 by Stanford University.
Read the full report
425 pages, nine chapters, hundreds of charts — all free from Stanford HAI, with raw data and an interactive Global AI Vibrancy tool.
Open the AI Index 2026 →