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

303%

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 October 15, 2024

Vital signs

31,122
AI safety articles were released between 2017 and 2022 (estd.)
303%
growth in research between 2017 and 2022
2%
of all artificial intelligence articles focused on AI safety

Citation impact

42
citations for the average AI safety article
20
citations for the average artificial intelligence article

Countries

35%
of AI safety articles have American authors
11%
have Chinese authors
20%
have European authors

Country	Articles (2017-2022)	Growth (2017-2022)	Global share	Citations per article
United States	11,036	221%	35%	68.4
China	3,549	476%	11%	50.4
United Kingdom	2,322	286%	7%	49.9

Top-cited articles

A Unified Approach to Interpreting Model Predictions
2017: arXiv. 20,338 citations
CBAM: Convolutional Block Attention Module
2018: arXiv. 19,286 citations
Towards Deep Learning Models Resistant to Adversarial Attacks
2017: arXiv. 10,668 citations
mixup: Beyond Empirical Risk Minimization
2017: arXiv. 8,814 citations
XLNet: Generalized Autoregressive Pretraining for Language Understanding
2019: arXiv. 8,299 citations
Training language models to follow instructions with human feedback
2022: arXiv. 8,181 citations
Co-teaching: Robust Training of Deep Neural Networks with Extremely Noisy Labels
2018: arXiv. 7,027 citations
Cascade R-CNN: High Quality Object Detection and Instance Segmentation
2017: arXiv. 6,654 citations
ArcFace: Additive Angular Margin Loss for Deep Face Recognition
2018: IEEE Transactions on Pattern Analysis and Machine Intelligence. 6,274 citations
Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI
2019: arXiv. 5,816 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	665	86.4	United States
Google	613	150.4	United States
Stanford University	551	137.5	United States

Patents and industry

8%
of AI safety articles are cited in patents
6%
of artificial intelligence articles are cited in patents
13%
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 (2017-2022)	Citations per article	Country
Google	613	150.4	United States
Microsoft	430	84	United States
Meta	207	179.7	United States

The Research Almanac runs on bibliometric data from Clarivate Web of Science and other sources. For details on our sourcing, methodology, and limitations, see the tool documentation.