Machine learning statistics now span model research, computing scale, organizational adoption, operating costs, measured outcomes and governance. The figures below distinguish observed counts, estimates and survey results, while stating the relevant period, geography and population.
Key Machine Learning Statistics
The most broadly useful figures include:
- In 2024, 78% of respondents globally reported organizational AI use, compared with 55% in 2023.
- In 2024, 71% of global respondents reported generative-AI use in at least one business function, compared with 33% in 2023.
- In 2023, 149 foundation models were released globally, and 65.7% were open-source.
- From 2013 to 2023, annual AI publications rose from approximately 102,000 to more than 242,000 globally.
- In 2024, nearly 90% of notable AI models came from industry globally, compared with 60% in 2023.
- In 2024, U.S. institutions produced 40 notable AI models, versus 15 from China and three from Europe combined.
- From 2023 to 2024, SWE-bench success rose from 4.4% to 71.7% of coding problems for the evaluated global AI systems.
- From 2023 to 2024, the U.S.-China MMLU model gap narrowed from 17.5 points to 0.3 points among the report’s selected leading models.
- In 2024, 13.5% of EU enterprises with at least 10 employees used at least one AI technology, up from 8.0% in 2023.
- In 2024, AI use reached 27.6% of enterprises in Denmark, 25.1% in Sweden and 24.7% in Belgium.
- In 2024, corporate AI investment was estimated at $252.3 billion globally.
- From November 2022 to October 2024, the cost of querying a GPT-3.5-equivalent MMLU model fell from $20.00 to $0.07 per million tokens.
- In 2024 surveys, 49% of AI-using organizations reported service-operations cost savings.
- Estimated training emissions were 5,184 metric tons for GPT-4 in 2023.
- From 2016 to 2024, U.S. state-level AI-related laws increased from one to 131.
- In a global survey reported in 2025, 60% expected AI to change how they do their job within five years, while 36% expected AI to replace their job.
Contents
- Key Machine Learning Statistics
- Machine Learning Research and Infrastructure Scale
- Machine Learning Benchmark and Efficiency Trends
- Machine Learning Adoption by Organizations and Regions
- Machine Learning Investment, Training Costs and Prices
- Measured Business, Labor and Environmental Outcomes
- Machine Learning Governance and Public Expectations
Machine Learning Research and Infrastructure Scale
Research output and model development expanded materially during the measured periods. The AI Index counts AI publications in its classified venues, so these figures are not a count of every publication worldwide.
| Measure | Earlier period | Later period |
|---|---|---|
| Annual AI publications | Approximately 102,000 in 2013 | More than 242,000 in 2023 |
| AI share of computer-science publications | 21.6% in 2013 | 41.8% in 2023 |
| AI patents | 3,833 in 2010 | 122,511 in 2023 |
| Open-source share of new foundation models | 33.3% in 2021; 44.4% in 2022 | 65.7% in 2023 |
Source: Research and Development | The 2025 AI Index Report. Foundation-model counts and notable-model selection follow the AI Index taxonomies.
Nearly 90% of notable AI models released in 2024 came from industry, compared with 60% in 2023. The sample is defined by the AI Index, rather than by all machine-learning releases.
U.S. institutions produced 40 notable AI models in 2024, compared with 15 from China and three from Europe combined. China accounted for 69.7% of global AI patent grants in 2023, but patent grants do not equal deployed systems.
Training compute for notable AI models has been estimated to double approximately every five months, while training-dataset sizes for large language models have been estimated to double approximately every eight months. These are trend estimates for notable models and LLM datasets, not measurements of all machine-learning training.
Machine Learning Benchmark and Efficiency Trends
Benchmark results show rapid gains, but each result depends on its test set, model selection and protocol. Benchmark performance should therefore be read as evaluated capability, not as a direct measure of production reliability.
From 2023 to 2024, scores gained 18.8 percentage points on MMMU and 48.9 points on GPQA. SWE-bench success increased from 4.4% of coding problems in 2023 to 71.7% in 2024, although SWE-bench remains a benchmark rather than production output.
| Comparison | Earlier result | Later result |
|---|---|---|
| U.S.-China MMLU gap | 17.5 points at end-2023 | 0.3 points at end-2024 |
| U.S.-China MMMU gap | 13.5 points in 2023 | 8.1 points in 2024 |
| U.S.-China MATH gap | 24.3 points in 2023 | 1.6 points in 2024 |
| U.S.-China HumanEval gap | 31.6 points in 2023 | 3.7 points in 2024 |
Source: Technical Performance | The 2025 AI Index Report. The comparisons cover only the report’s selected leading models.
The leading closed-weight model’s Chatbot Arena advantage narrowed from 8.04% in January 2024 to 1.70% in February 2025. By early 2025, the gap between the top and tenth Chatbot Arena model had narrowed from 11.9% to 5.4%; these leaderboard gaps vary with voting periods.
OpenAI o1 scored 74.4% on an IMO qualifying exam, compared with 9.3% for GPT-4o, in the cited 2024 comparison. The same comparison found o1 nearly six times more expensive and 30 times slower than GPT-4o, with results dependent on pricing, hardware and task setup.
Machine Learning Adoption by Organizations and Regions
Adoption figures use different populations, so a global respondent survey should not be compared directly with an enterprise census. Eurostat explicitly includes machine learning among AI technologies and covers enterprises with at least 10 employees in specified sectors.
The share of global survey respondents reporting organizational AI use rose from 55% in 2023 to 78% in 2024. Reported generative-AI use in at least one business function rose from 33% to 71% over the same period; generative AI is a subset of broader AI.
| Geography or population | 2023 | 2024 |
|---|---|---|
| EU enterprises with at least 10 employees using at least one AI technology | 8.0% | 13.5% |
| Denmark enterprises | — | 27.6% |
| Sweden enterprises | — | 25.1% |
| Belgium enterprises | — | 24.7% |
| Romania enterprises | — | 3.1% |
| Poland enterprises | — | 5.9% |
| Bulgaria enterprises | — | 6.5% |
Source: Usage of AI technologies increasing in EU enterprises. The country figures are for enterprises with at least 10 employees.
In the cited global survey, organizational AI-use rates increased by 27 percentage points year over year in Greater China and by 23 percentage points in Europe. These are changes within the survey sample, not population-wide estimates.
Within the EU in 2024, text-mining AI was used by 6.9% of enterprises, natural-language generation by 5.4%, and speech recognition by 4.8%. These technology-specific categories are reported separately from Eurostat’s machine-learning definition.
Machine Learning Investment, Training Costs and Prices
Investment and cost figures are broader than machine learning in several cases. The AI Index investment series includes wider AI activity, while training costs are estimates derived from cloud compute rental prices and affected by disclosure limits.
Corporate AI investment reached an estimated $252.3 billion globally in 2024; private investment grew 44.5% and mergers and acquisitions grew 12.1% year over year. Total AI investment grew more than 13-fold from 2014 to 2024.
Private generative-AI investment reached an estimated $33.9 billion globally in 2024, up 18.7% from 2023 and more than 8.5 times 2022. In 2024, estimated U.S. private AI investment was $109.1 billion, compared with $9.3 billion in China and $4.5 billion in the United Kingdom.
| Selected training run | Estimated inflation-adjusted cost |
|---|---|
| Transformer, 2017 | About $670 |
| RoBERTa Large, 2019 | $160,000 |
| GPT-4, 2023 | $79 million |
| Gemini 1.0 Ultra, 2024 | $192 million |
| Mistral Large, 2024 | $41 million |
| Llama 3.1-405B, 2024 | $170 million |
Source: Economy | The 2025 AI Index Report and Research and Development | The 2025 AI Index Report. These are estimated costs per training run, not audited budgets.
The listed price for querying a GPT-3.5-equivalent MMLU model fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024. The AI Index also reports estimated LLM inference-price declines ranging from ninefold to 900-fold per year, depending on the task.
Measured Business, Labor and Environmental Outcomes
Surveyed outcomes indicate reported benefits, not established causal effects. Among AI-using organizations in 2024, 49% reported cost savings in service operations, 43% in supply-chain management and 41% in software engineering; most reported savings were below 10%.
Among AI-using respondents, 71% in marketing and sales, 63% in supply-chain management and 57% in service operations reported revenue gains. Most reported revenue increases were below 5%.
A UK government survey of 350 businesses in selected sectors that were using or considering AI found that 27% of current AI users used machine learning; the machine-learning base was 94 current users. Among those surveyed machine-learning users, 52% had used it for more than three years, 27% for one to three years and 14% were piloting it, while 7% answered “don’t know.”
Source: AI cyber security survey – main report. The sample was screened for AI use or planned use and was not representative of all UK businesses.
Reported AI-related incidents reached 233 in 2024, a 56.4% increase over 2023. The incident database is not comprehensive and counts reported incidents.
| Selected model and year | Estimated training emissions |
|---|---|
| GPT-3, 2020 | 588 metric tons of carbon |
| GPT-4, 2023 | 5,184 metric tons |
| Llama 3.1 405B, 2024 | 8,930 metric tons |
Source: Responsible AI | The 2025 AI Index Report. Emissions depend on the energy mix, hardware and system boundaries.
Restricted tokens in actively maintained C4 domains rose from 5–7% in 2023 to 20–33% in 2024. This is a sampled-corpus measure, not a count of all available training data.
Machine Learning Governance and Public Expectations
Policy activity expanded alongside model development and adoption. Counts below use the AI Index classifications and measure legislative or administrative attention, not the effectiveness of resulting rules.
U.S. state-level AI-related laws increased from one in 2016 to 49 in 2023 and 131 in the following year. AI mentions in proceedings across 75 major countries rose from 1,557 in 2023 to 1,889 in 2024, a 21.3% increase.
The United States recorded 59 AI-related federal regulations in 2024, up from 25 in 2023, while the number of issuing agencies rose from 21 to 42.
Source: Policy and Governance | The 2025 AI Index Report.
The average Foundation Model Transparency Index score rose from 37% in October 2023 to 58% in May 2024. The index measures disclosure against its own framework, rather than every dimension of model safety or accountability.
In a global survey reported in 2025, 83% of respondents in China, 80% in Indonesia and 77% in Thailand said AI products and services offer more benefits than drawbacks. The corresponding shares were 40% in Canada, 39% in the United States and 36% in the Netherlands.
Globally, 60% expected AI to change how they do their job within five years, while 36% expected AI to replace their job. These are expectations, not observed employment changes.
Source: Public Opinion | The 2025 AI Index Report.