AI safety – ETO Research Almanac

ETO Research Almanac

Between 2018 and 2023, researchers published roughly

45,000

English-language articles on AI safety

The global AI safety literature grew

312%

between 2018 and 2023

Articles on AI safety made up

of all articles on artificial intelligence

Authors from private companies contributed to

11%

of AI safety articles

What counts as "AI safety" research?•How precise are these numbers?•What are the other caveats?

Explore relevant research in the Map of Science

Data last updated January 6, 2025

Vital signs

45,595
AI safety articles were released between 2018 and 2023 (estd.)
312%
growth in research between 2018 and 2023
2%
of all artificial intelligence articles focused on AI safety

Citation impact

28
citations for the average AI safety article
16
citations for the average artificial intelligence article

Countries

30%
of AI safety articles have American authors
12%
have Chinese authors
18%
have European authors

Country	Articles (2018-2023)	Growth (2018-2023)	Global share	Citations per article
United States	13,787	112%	30%	50.5
China	5,675	514%	12%	34.4
United Kingdom	3,063	176%	7%	34.7

Top-cited articles

CBAM: Convolutional Block Attention Module
2018: arXiv. 21,432 citations
Training language models to follow instructions with human feedback
2022: arXiv. 9,807 citations
XLNet: Generalized Autoregressive Pretraining for Language Understanding
2019: arXiv. 8,702 citations
Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels
2018: arXiv. 7,163 citations
Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI
2019: arXiv. 6,381 citations
Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead
2018: arXiv. 6,043 citations
Generalized Intersection over Union: A Metric and A Loss for Bounding Box Regression
2019: arXiv. 5,123 citations
From local explanations to global understanding with explainable AI for trees
2020: Nature machine intelligence. 4,711 citations
A Survey on Bias and Fairness in Machine Learning
2019: arXiv. 4,035 citations
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? ��
2021: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. 3,987 citations

Research organizations

Top producers of AI safety research (by article count)

#1 Carnegie Mellon University (United States)
#2 Google (United States)
#3 Stanford University (United States)

Organization	Articles (2018-2023)	Citations per article	Country
Carnegie Mellon University	807	69.8	United States
Google	696	129	United States
Stanford University	646	86	United States

Patents and industry

6%
of AI safety articles are cited in patents
5%
of artificial intelligence articles are cited in patents
11%
of AI safety articles had authors from private companies (vs 6% of artificial intelligence articles in general)

Top industry organizations by publication count

Company	Articles (2018-2023)	Citations per article	Country
Google	696	129	United States
Microsoft	498	74.8	United States
Amazon	242	44.5	United States

The Research Almanac runs on bibliometric data from Clarivate Web of Science and other sources. Research subject classifications are based on work supported in part by the Alfred P. Sloan Foundation under Grant No. G-2023-22358. For details on our sourcing, methodology, and limitations, see the tool documentation.