AI safety – ETO Research Almanac

ETO Research Almanac

Between 2016 and 2021, researchers published roughly

30,000

English-language articles on AI safety

The global AI safety literature grew

358%

between 2016 and 2021

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?

Data last updated May 17, 2023

Vital signs

29,259

AI safety articles were released between 2016 and 2021 (estd.)

358%

growth in research between 2016 and 2021

of all artificial intelligence articles focused on AI safety

Citation impact

citations for the average AI safety article

citations for the average artificial intelligence article

Countries

43%

of AI safety articles have American authors

11%

have Chinese authors

23%

have European authors

Country	Articles (2016-2021)	Growth (2016-2021)	Global share	Citations per article
United States	12,622	305%	43%	32.8
China	3,096	814%	11%	22.8
United Kingdom	2,801	282%	10%	26.9

Top-cited articles

“Why Should I Trust You?”: Explaining the Predictions of Any Classifier
2016: arXiv: Learning. 8769 citations
A unified approach to interpreting model predictions
2017: arXiv: Artificial Intelligence. 6285 citations
Towards Evaluating the Robustness of Neural Networks
2016: arXiv: Cryptography and Security. 5351 citations
RoBERTa: A Robustly Optimized BERT Pretraining Approach
2019: arXiv: Computation and Language. 5087 citations
Grad-CAM: Why did you say that? Visual Explanations from Deep Networks via Gradient-based Localization
2016: arXiv: Computer Vision and Pattern Recognition. 5065 citations
Towards Deep Learning Models Resistant to Adversarial Attacks.
2017: arXiv: Machine Learning. 3790 citations
Cascade R-CNN: Delving into High Quality Object Detection
2017: arXiv: Computer Vision and Pattern Recognition. 3630 citations
Adversarial Discriminative Domain Adaptation
2017: arXiv: Computer Vision and Pattern Recognition. 3480 citations
Adversarial examples in the physical world
2016: arXiv: Computer Vision and Pattern Recognition. 3104 citations
Overcoming catastrophic forgetting in neural networks
2016: arXiv: Learning. 2899 citations

Research organizations

Top producers of AI safety research (by article count)

#1 University of California, Berkeley (United States)
#2 Google (United States)
#3 Carnegie Mellon University (United States)

Organization	Articles (2016-2021)	Citations per article	Country
University of California, Berkeley	660	79.4	United States
Google	658	73.4	United States
Carnegie Mellon University	651	72.1	United States

Patents and industry

of AI safety articles are cited in patents

of artificial intelligence articles are cited in patents

11%

of AI safety articles had authors from private companies (vs 4% of artificial intelligence articles in general)