1 · Private investment soared while concentration intensified
Private AI investment reached about $93.5 billion in 2021, more than double 2020 — while the number of newly funded AI companies kept falling, from 1,051 in 2019 to 746.
2021: investment doubled while it flowed to fewer companies
The 2022 AI Index covers calendar year 2021 across five chapters — research, technical performance, technical AI ethics, the economy and education together, and AI policy and governance. It is the leanest edition on this site, and the one that first records the shape of the decade to come: money concentrating, models getting cheaper to run and more biased to use, and legislatures beginning to write AI into law. A few numbers that set the scene:
93.5global private AI investment in 2021 (US$ bn), over 2× 2020
15funding rounds worth $500m or more in 2021 (four in 2020)
746newly funded AI companies in 2021, down from 1,051 in 2019
29% more elicited toxicity from a 280B model than a 2018 117M one
63.6% fall in the cost to train an image classifier since 2018
18AI bills passed into law across 25 countries (one in 2016)
The Top Takeaways
The report's own summary of the year, in its own order. Click any card for the full text.
2 · The US and China dominated cross-country AI collaboration
Despite rising geopolitical tension, US–China collaborations on AI publications grew fivefold since 2010 and produced 2.7 times more papers than the next pair, the UK and China.
3 · Language models are more capable than ever, but also more biased
A 280-billion-parameter model built in 2021 showed a 29% increase in elicited toxicity over the 117-million-parameter model considered state of the art in 2018.
4 · The rise of AI ethics everywhere
Research on fairness and transparency has exploded since 2014, with a fivefold increase in publications at ethics-related conferences; industry-affiliated researchers contributed 71% more year over year.
5 · AI becomes more affordable and higher performing
Since 2018 the cost to train an image classification system fell 63.6% while training times improved 94.4% — and the same pattern holds across other MLPerf categories.
6 · Data, data, data
Nine of the ten benchmarks in this report had state-of-the-art systems trained with extra data — a trend that implicitly favors private-sector actors with vast datasets.
7 · More global legislation on AI than ever
Across 25 countries, bills containing 'artificial intelligence' passed into law grew from one in 2016 to 18 in 2021. Spain, the UK and the US each passed three.
8 · Robotic arms are becoming cheaper
An AI Index survey found the median price of robotic arms fell 46.2% in five years — from $42,000 per arm in 2017 to $22,600 in 2021.
Five chapters, in brief
Each chapter's headline highlights. Expand to read the key numbers.
1 · Research & Development
US–China cross-country collaboration on AI publications grew fivefold since 2010 and produced 2.7 times more papers than the UK–China pair. China led the world in journal, conference and repository publications combined — 63.2% higher than the United States — while the US held a dominant lead in conference and repository citations. Collaboration between educational and nonprofit organizations produced the most publications. AI patents filed in 2021 were more than 30 times the 2015 figure, a compound annual growth rate of 76.9%.Full deep dive: Research & Development →
2 · Technical Performance
Nine of the ten benchmarks in the report had top results from systems trained with extra data. AI already exceeded human performance on SuperGLUE and SQuAD by 1%–5%, while on abductive natural language inference the human lead shrank from nine percentage points in 2019 to one. General reinforcement learning improved 129% on Procgen in two years, and the top chess engine exceeded Magnus Carlsen's peak Elo by 24%. Training cost fell 63.6% since 2018 and training time improved 94.4%; median robotic arm prices fell 46.2% in five years.Full deep dive: Technical Performance →
3 · Technical AI Ethics
A 280-billion-parameter model built in 2021 showed 29% more elicited toxicity than the 117-million-parameter 2018 state of the art — capability and bias growing together. Research on fairness and transparency rose fivefold at ethics conferences since 2014, with industry-affiliated researchers contributing 71% more year over year. Multimodal models learned multimodal biases: experiments on CLIP showed images of Black people misclassified as nonhuman at over twice the rate of any other race.Full deep dive: Technical AI Ethics →
4 · The Economy & Education
Private AI investment reached about $93.5 billion in 2021, more than double 2020, while newly funded companies fell to 746 from 1,051 in 2019 and rounds worth $500 million or more rose from four to fifteen. Data management, processing and cloud drew the most investment, 2.6 times the 2020 figure. New Zealand, Hong Kong, Ireland, Luxembourg and Sweden had the highest growth in AI hiring since 2016. A McKinsey survey found 29% and 41% of respondents recognize equity and explainability as risks, but only 19% and 27% act on them. In 2020, one in five CS PhD graduates specialized in AI or ML.Full deep dive: The Economy & Education →
5 · AI Policy & Governance
Across 25 countries, bills containing 'artificial intelligence' passed into law grew from one in 2016 to 18 in 2021, with Spain, the UK and the US each passing three. In the US the federal record shows a sharp rise in proposed AI bills from 2015 to 2021 while only 2% became law. State legislators passed one in every 50 proposed AI bills in 2021, from a proposal count that grew from two in 2012 to 131. The 117th Congress was on track for the most AI mentions since 2001 — 295 by the end of 2021, halfway through the session, against 506 for the whole previous session.Full deep dive: AI Policy & Governance →
In the report's own words
Lines that capture the year, from the Top Takeaways and the chapter highlights.
The private investment in AI in 2021 totaled around $93.5 billion — more than double the total private investment in 2020, while the number of newly funded AI companies continues to drop.
Large language models are setting new records on technical benchmarks, but new data shows that larger models are also more capable of reflecting biases from their training data.
Top results across technical benchmarks have increasingly relied on the use of extra training data. This trend implicitly favors private sector actors with access to vast datasets.
Algorithmic fairness and bias has shifted from being primarily an academic pursuit to becoming firmly embedded as a mainstream research topic.
The federal legislative record in the United States shows a sharp increase in proposed bills that relate to AI from 2015 to 2021, while the number passed remains low, with only 2% ultimately becoming law.
Read the full report
230 pages, five chapters, hundreds of charts — free from Stanford HAI, with the underlying public data.
Open the AI Index 2022 →