AI safety – ETO Research Almanac

ETO Research Almanac

Between 2018 and 2023, researchers published roughly

40,000

English-language articles on AI safety

The global AI safety literature grew

285%

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 October 13, 2025

Vital signs

41,212
AI safety articles were released between 2018 and 2023 (estd.)
285%
growth in research between 2018 and 2023
3%
of all artificial intelligence articles focused on AI safety

Citation impact

37
citations for the average AI safety article
20
citations for the average artificial intelligence article

Countries

29%
of AI safety articles have American authors
12%
have Chinese authors
17%
have European authors

Country	Articles (2018-2023)	Growth (2018-2023)	Global share	Citations per article
United States	12,101	73%	29%	63.4
China	4,905	439%	12%	46
Germany	2,614	128%	6%	38.5

Top-cited articles

CBAM: Convolutional Block Attention Module
2018: arXiv. 25,948 citations
Training language models to follow instructions with human feedback
2022: arXiv. 14,229 citations
XLNet: Generalized Autoregressive Pretraining for Language Understanding
2019: arXiv. 9,073 citations
Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI
2019: arXiv. 8,021 citations
ArcFace: Additive Angular Margin Loss for Deep Face Recognition
2018: arXiv. 7,809 citations
Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead
2018: arXiv. 7,589 citations
From local explanations to global understanding with explainable AI for trees
2020: Nature machine intelligence. 6,255 citations
Generalized Intersection over Union: A Metric and A Loss for Bounding Box Regression
2019: arXiv. 6,098 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. 5,685 citations
A Survey on Bias and Fairness in Machine Learning
2019: arXiv. 5,309 citations

Research organizations

Top producers of AI safety research (by article count)

#1 Carnegie Mellon University (United States)
#2 Massachusetts Institute of Technology (United States)
#3 Stanford University (United States)

Organization	Articles (2018-2023)	Citations per article	Country
Carnegie Mellon University	727	92.5	United States
Massachusetts Institute of Technology	594	101.3	United States
Stanford University	573	113.7	United States

Patents and industry

9%
of AI safety articles are cited in patents
7%
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	508	180.9	United States
Microsoft	384	95	United States
IBM	302	53.8	United States

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.