AI safety – ETO Research Almanac

ETO Research Almanac

Between 2017 and 2022, researchers published roughly

30,000

English-language articles on AI safety

The global AI safety literature grew

313%

between 2017 and 2022

Articles on AI safety made up

of all articles on artificial intelligence

Authors from private companies contributed to

13%

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 December 13, 2024

Vital signs

31,270
AI safety articles were released between 2017 and 2022 (estd.)
313%
growth in research between 2017 and 2022
2%
of all artificial intelligence articles focused on AI safety

Citation impact

45
citations for the average AI safety article
21
citations for the average artificial intelligence article

Countries

36%
of AI safety articles have American authors
12%
have Chinese authors
20%
have European authors

Country	Articles (2017-2022)	Growth (2017-2022)	Global share	Citations per article
United States	11,170	181%	36%	73.2
China	3,649	528%	12%	54
United Kingdom	2,375	227%	8%	48.4

Top-cited articles

A Unified Approach to Interpreting Model Predictions
2017: arXiv. 22,749 citations
CBAM: Convolutional Block Attention Module
2018: arXiv. 21,432 citations
Towards Deep Learning Models Resistant to Adversarial Attacks
2017: arXiv. 11,270 citations
Training language models to follow instructions with human feedback
2022: arXiv. 9,807 citations
mixup: Beyond Empirical Risk Minimization
2017: arXiv. 9,351 citations
XLNet: Generalized Autoregressive Pretraining for Language Understanding
2019: arXiv. 8,702 citations
Cascade R-CNN: High Quality Object Detection and Instance Segmentation
2017: arXiv. 7,175 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

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 (2017-2022)	Citations per article	Country
Carnegie Mellon University	675	91.4	United States
Google	614	156.3	United States
Stanford University	564	141.2	United States

Patents and industry

10%
of AI safety articles are cited in patents
7%
of artificial intelligence articles are cited in patents
13%
of AI safety articles had authors from private companies (vs 7% of artificial intelligence articles in general)

Top industry organizations by publication count

Company	Articles (2017-2022)	Citations per article	Country
Google	614	156.3	United States
Microsoft	435	88.5	United States
Meta	214	180.9	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.