Five editions, five years, one trend line
The AI Index is an independent Stanford HAI initiative that has tracked artificial intelligence since 2017 — across research, technical performance, responsible AI, the economy, science and medicine, policy, education and public opinion. This site holds five editions, covering calendar years 2021 through 2025. Reading them together shows two things no single edition can: the rate of change, and the places where the Index has since revised what it said. A few numbers that set the scene:
Pick an edition
Each edition is a complete site of its own — an overview page plus one deep dive per chapter, in English and Traditional Chinese.
The same chapter has five different numbers
Where each subject sits in each edition. Only the first four rows hold still; everything below them moves, appears or disappears. This is why a citation like 'Chapter 6' means nothing without the edition attached.
| Subject | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|
| Research & Development | 1 | 1 | 1 | 1 | 1 |
| Technical Performance | 2 | 2 | 2 | 2 | 2 |
| AI Ethics → Responsible AI | 3 | 3 | 3 | 3 | 3 |
| Economy | 4 | 4 | 4 | 4 | 4 |
| Education | 4 | 5 | 6 | 7 | 7 |
| Science & Medicine | no such chapter in the 2022 edition | no such chapter in the 2023 edition | 5 | 5 | 5 + 6 |
| Policy & Governance | 5 | 6 | 7 | 6 | 8 |
| Diversity | no such chapter in the 2022 edition | 7 | 8 | dropped from the 2025 edition | dropped from the 2026 edition |
| Public Opinion | no such chapter in the 2022 edition | 8 | 9 | 8 | 9 |
How the chapters split, merged and vanished
The table above shows where each subject landed. This shows how it got there — one chapter became two, two became one, one was renamed, and one was quietly dropped.
A peak, a two-year trough, then a doubling
Global AI investment in billions of US dollars. Corporate investment is the broadest measure — private investment, mergers and acquisitions, minority stakes and public offerings together; private investment is the venture and private-equity slice inside it.
- Corporate (all four kinds)
- Private investment only
The same year, reported twice, years apart
Global corporate AI investment, in billions of US dollars. Switch between what the edition of the day printed and what the 2026 edition now says about those same years. Both views share one scale, so the bars are directly comparable.
Each bar is the figure printed by the earliest edition on this site that covered that year: 2019–2021 from the 2022 edition, 2022 from the 2023 edition, 2023 from the 2024 edition.
Every edition's figure for every year
Global corporate AI investment, US$ billions. Read a row across and you are watching one year get rewritten. The data provider says why: it keeps widening the set of companies it tracks, so each edition looks further back with a bigger net.
| Year measured | 2022 ed. | 2023 ed. | 2024 ed. | 2025 ed. | 2026 ed. |
|---|---|---|---|---|---|
| 2014 | 29.17 | 19.04 | 19.04 | 19.04 | 19.04 |
| 2019 | 66.98 | 95.57 | 103.27 | 103.27 | 103.27 |
| 2020 | 119.54 | 146.74 | 202.49 | 221.87 | 221.87 |
| 2021 | 176.47 | 276.14 | 337.40 | 360.73 | 360.73 |
| 2022 | the 2022 edition predates this year | 189.59 | 234.95 | 253.25 | 253.25 |
| 2023 | not yet covered | not yet covered | 189.16 | 201 | 201 |
| 2024 | not yet covered | not yet covered | not yet covered | 252.33 | 253.02 |
| 2025 | not yet covered | not yet covered | not yet covered | not yet covered | 581.69 |
Organizations moved first; the public warmed slowly
Percentages. Organizational adoption sat flat for years and then stepped up 23 points in a single year, while the share of people saying AI's benefits outweigh its drawbacks crept up a few points at a time — and the American figure has never caught the global one.
- Organizations using AI
- Benefits outweigh drawbacks — global
- Benefits outweigh drawbacks — US
Incidents kept climbing; US rulemaking levelled off
Counts. Both lines measure things that are counted rather than estimated, which makes them unusually comparable — but they are not the same kind of count, and the gap between them is the point.
- AI incidents logged
- New US federal AI regulations
What it cost to train the model of the year
AI Index estimates of the compute used to train notable models, in US dollars. Six years separate the Transformer from Gemini Ultra, the most expensive model here, and across them the bill grows by a factor of about 200,000. The last two rows are where the curve bends back.
| Model | Released | Estimated compute cost | Estimated by |
|---|---|---|---|
| Transformer | 2017 | $930 | 2024 edition |
| RoBERTa Large | 2019 | $160,018 | 2024 edition |
| Chinchilla | 2022 | $2.11m | 2023 edition |
| GPT-4 | 2023 | $78.4m | 2024 edition |
| Gemini Ultra | 2023 | $191.4m | 2024 edition |
| Llama 3.1-405B | 2024 | $170m | 2025 edition |
| DeepSeek-V3 | 2024 | $6m | 2025 edition, developer figure |
One line per edition
What each edition found when it looked at its year.
- 2021 · 2022 edition
The money doubled; the recipients did not
Private AI investment reached about $93.5 billion, more than double 2020, while the count of newly funded AI companies fell for a third straight year and rounds of $500 million or more went from four to fifteen.
- 2022 · 2023 edition
Generative AI arrived and investment fell
DALL·E 2, Stable Diffusion and ChatGPT put AI in front of everyone, and in the same year private investment fell year over year for the first time in a decade. Traditional benchmarks began saturating.
- 2023 · 2024 edition
The frontier passed $100 million
Overall private investment fell for a second year while generative AI funding nearly octupled to $25.2 billion. Gemini Ultra's training compute was estimated at $191 million. Responsible-AI evaluation was found to have no shared standard at all.
- 2024 · 2025 edition
Benchmarks stopped holding
Benchmarks introduced barely a year earlier were largely solved, inference cost for GPT-3.5-level performance fell more than 280-fold in eighteen months, and organizational adoption jumped from 55% to 78%.
- 2025 · 2026 edition
Scaling outran everything around it
Corporate investment more than doubled to $581.7 billion and adoption reached 88%, while documented incidents rose to 362 and the US–China gap at the technical frontier effectively closed.
How to read this site
Four things worth knowing before you compare anything across editions.
An edition is named for the year it is published, not the year it measures
The 2022 edition reports on calendar year 2021; the 2026 edition reports on 2025. This trips people up constantly when comparing figures across sources, so every chart on this site labels the measured year rather than the edition.
The chapter structure changes almost every year
The five editions here have five, eight, nine, eight and nine chapters respectively. The chapter map above shows where each subject sits in each edition, and the diagram below it shows how one chapter became two, two became one, one was renamed and one was dropped. Chapter numbers do not line up across editions — follow the names, not the numbers.
The Index revises its own history
This is the one that catches people out, because nothing in a single edition hints at it. Global corporate AI investment for 2021 was printed as $176.5 billion in the 2022 edition and is now stated as $360.7 billion — 104% higher for the same year. The data provider explains why: it keeps widening the set of companies it tracks, so each edition looks back at the same years with a bigger net. Revisions are not all upward — 2014 was cut from $29.2 billion to $19.0 billion — and they settle after an edition or two.
So every multi-year line on this page is drawn from one edition's own back-series, or from years where the overlapping editions agree. None of them is assembled by taking each edition's figure for its own year, which is the natural thing to do and produces a shape the data does not have.
A metric can quietly change what it counts
Incident counts up to the 2023 edition come from the AIAAIC database; from the 2024 edition onward they come from the AI Incident Database. Chaining the two would draw a line falling from 260 incidents in 2021 to 123 in 2023 — an artifact of the switch, not something that happened. Two more in the same family:
- Global AI laws passed: the country coverage changed in every single edition — 25 countries, then 127, then 128, then down to 114, and finally to the G20 alone on a different data source. The 2026 edition states it outright: its numbers “may differ from what was reported in the 2025 AI Index report, as the methodology has changed.” The raw counts are not a time series at all.
- Legislative mentions of AI: coverage shrank from 81 countries to about 75, and the 2026 edition dropped the metric entirely, replacing it with a US-only proxy.
- US federal AI regulations: the one metric on this page whose definition never moved, which is why it is the one drawn beside incident counts above.
This is a reformatting, not the report
Everything here is drawn from the reports' own top takeaways, per-chapter highlights and figures, restructured for reading on screen. Nothing is inferred, extrapolated or supplemented from other sources. Where a figure matters, go to the original — each edition links straight to it.
Read the original reports
The AI Index is published open access by the Stanford Institute for Human-Centered AI under CC BY-ND 4.0. Full PDFs, the underlying raw data, and the Global AI Vibrancy tool are all available there.
Stanford HAI · AI Index