1 · Industry races ahead of academia
In 2022 industry released 32 significant machine learning models against academia's three. Until 2014 most such models came from universities.
2022: the public met generative AI, and the money walked out
The 2023 AI Index covers calendar year 2022 across eight chapters — research, technical performance, technical AI ethics, the economy, education, policy, diversity and public opinion. It is the edition that recorded two opposite things at once: DALL·E 2, Stable Diffusion and ChatGPT put AI in front of everyone, while private investment fell year over year for the first time in a decade. A few numbers that set the scene:
32significant ML models from industry in 2022 (academia: 3)
91.9global private AI investment (US$ bn), down 26.7%
26fold rise in AI incidents and controversies since 2012
37AI bills passed into law across 127 countries (one in 2016)
65.4% of new AI PhDs going to industry (28.2% to academia)
35% of Americans saying AI benefits outweigh drawbacks (China: 78%)
The Top Ten Takeaways
The report's own summary of the year, in its own order. Click any card for the full text.
2 · Performance saturation on traditional benchmarks
AI kept posting state-of-the-art results, but year-over-year improvement on many benchmarks stayed marginal — and the speed at which benchmarks saturate is increasing.
3 · AI is both helping and harming the environment
BLOOM's training run emitted 25 times more carbon than one air passenger flying New York to San Francisco — while models like BCOOLER show AI optimizing energy use.
4 · The world's best new scientist … AI?
In 2022 AI models were used to aid hydrogen fusion, improve the efficiency of matrix manipulation and generate new antibodies.
5 · The number of AI misuse incidents is rapidly rising
The AIAAIC database records a 26-fold increase in AI incidents and controversies since 2012, including a deepfake video of President Zelenskyy surrendering.
6 · Demand for AI skills is rising across nearly every US sector
Across every US sector with data except agriculture, forestry, fishing and hunting, AI-related job postings rose on average from 1.7% in 2021 to 1.9% in 2022.
7 · For the first time in a decade, private AI investment decreased
Global private AI investment was $91.9 billion in 2022, a 26.7% decrease from 2021. Funding events and newly funded companies fell too.
8 · Adoption has plateaued, but adopters keep pulling ahead
The share of companies adopting AI has more than doubled since 2017 but has sat between 50% and 60% in recent years — and those that adopted report real cost cuts and revenue gains.
9 · Policymaker interest in AI is on the rise
Across 127 countries, bills containing 'artificial intelligence' passed into law grew from one in 2016 to 37 in 2022; mentions in parliamentary proceedings rose nearly 6.5-fold since 2016.
10 · Chinese citizens feel most positively about AI; Americans do not
In a 2022 IPSOS survey, 78% of Chinese respondents agreed AI products have more benefits than drawbacks — against just 35% of Americans, among the lowest of all surveyed countries.
Eight chapters, in brief
Each chapter's headline highlights. Expand to read the key numbers.
1 · Research & Development
Industry released 32 significant machine learning models in 2022 against academia's three. Large models kept growing and getting costlier: GPT-2 had 1.5 billion parameters and cost an estimated $50,000 to train in 2019; PaLM had 540 billion and cost an estimated $8 million — 360 times larger and 160 times more expensive. China continues to lead total AI journal, conference and repository publications, while the US still leads on conference and repository citations, and produced 54% of the world's large language and multimodal models.Full deep dive: Research & Development →
2 · Technical Performance
Year-over-year gains on many benchmarks stayed marginal while saturation arrived faster, prompting broader suites like BIG-bench and HELM. Generative AI broke into public consciousness with DALL·E 2, Stable Diffusion, Make-A-Video and ChatGPT — all prone to hallucination. Systems became more flexible (BEiT-3, PaLI, Gato) but still struggled with complex planning. Nvidia used reinforcement learning to improve chip design, and Google used PaLM to suggest improvements to PaLM.Full deep dive: Technical Performance →
3 · Technical AI Ethics
AI incidents and controversies logged by the AIAAIC database rose 26-fold since 2012. Generative models arrived with their ethical problems attached — text-to-image generators are routinely biased along gender lines and chatbots can be tricked into harmful uses. Fairer is not always less biased: models scoring better on some fairness benchmarks tend to show worse gender bias. Accepted submissions to FAccT more than doubled since 2021 and grew tenfold since 2018.Full deep dive: Technical AI Ethics →
4 · The Economy
Global private AI investment fell 26.7% to $91.9 billion — the first decline in a decade, though still 18 times the 2013 figure. The US led with $47.4 billion, about 3.5 times China's $13.4 billion. Medical and healthcare drew the most investment ($6.1B), then data and cloud ($5.9B) and fintech ($5.5B). Adoption plateaued between 50% and 60%. In a GitHub survey, 88% of Copilot users felt more productive. China installed more industrial robots in 2021 than the rest of the world combined.Full deep dive: The Economy →
5 · Education
The share of new US computer science PhDs specializing in AI jumped to 19.1% in 2021, from 14.9% in 2020 and 10.2% in 2010. 65.4% of new AI PhDs took industry jobs against 28.2% entering academia — in 2011 the split was nearly even. North American CS, CE and information faculty hires stayed flat (710 in 2021 against 733 in 2012). The funding gap widened: median external research expenditure was $9.7 million at private US CS departments against $5.7 million at public ones. 181,040 AP CS exams were taken in 2021, ninefold the 2007 figure.Full deep dive: Education →
6 · Policy & Governance
Across 127 countries, AI bills passed into law grew from one in 2016 to 37 in 2022, and mentions of AI in the parliamentary records of 81 countries rose nearly 6.5-fold since 2016. The US moved from talk to enactment: 2% of federal AI bills became law in 2021 against 10% in 2022, with 35% of state-level bills passing. US government AI-related contract spending has grown roughly 2.5 times since 2017. There were 110 AI-related legal cases in US courts in 2022, 6.5 times the 2016 figure.Full deep dive: Policy & Governance →
7 · Diversity
North American CS students grew more ethnically diverse: white students were 71.9% of new resident CS bachelor's graduates in 2011 and 46.7% in 2021. New AI PhDs remained overwhelmingly male — 78.7% in 2021, with the female share up just 3.2 percentage points since 2011. The share of new female CS, CE and information faculty hires rose from 24.9% in 2017 to 30.2%, though 75.9% of such faculty are still male. The female share of AP CS exams rose from 16.8% in 2007 to 30.6% in 2021.Full deep dive: Diversity →
8 · Public Opinion
In a 2022 IPSOS survey, 78% of Chinese respondents said AI products have more benefits than drawbacks, followed by Saudi Arabia (76%) and India (71%); only 35% of Americans agreed. Men were consistently more positive than women. Self-driving cars remain unconvincing — only 27% globally would feel safe in one, and only 26% of Americans think driverless passenger vehicles are good for society. Among NLP researchers surveyed, 77% agreed private AI firms have too much influence and 73% expected AI to bring revolutionary societal change.Full deep dive: Public Opinion →
In the report's own words
Lines that capture the year, from the Top Ten Takeaways and the chapter highlights.
Until 2014, most significant machine learning models were released by academia. Since then, industry has taken over.
2022 saw the release of text-to-image models like DALL-E 2 and Stable Diffusion, text-to-video systems like Make-A-Video, and chatbots like ChatGPT.
Global AI private investment was $91.9 billion in 2022, which represented a 26.7% decrease since 2021.
Language models which perform better on certain fairness benchmarks tend to have worse gender bias.
Only 27% of respondents reported feeling safe in a self-driving car.
Read the full report
386 pages, eight chapters, hundreds of charts — free from Stanford HAI, with the underlying public data.
Open the AI Index 2023 →