A Semantic Analysis of the Recent Evolution of AI Research
If we want to steer AI in societally desirable directions, first we need to understand where it is heading. This requires smarter data about smarter machines.
Fast-improving Artificial Intelligence (AI) systems stand to transform productivity and help us tackle big societal challenges around health, the environment and scientific discovery. However, AI technologies also come with risks for privacy, safety, competition, sustainability and societal stability. This is why policymakers all over the world are putting in place strategies to realise the benefits of AI, while managing its downsides.
This report presents the findings of our most in-depth analysis of AI research and development to date, using arXive (an open science site used by the AI research community). We argue that policies need to be informed by smarter data about smarter machines: relevant, inclusive, timely and open information that makes the most of novel data sources and analytics techniques.
- There has been rapid growth in AI research in recent years: 77 per cent of the AI papers we identified in the data were published in the last five years.
- AI techniques are not only being adopted in computer science. Other scientific disciplines, from particle physics to materials science are also adopting AI tools in their research.
- The composition of AI research has experienced a drastic change since the 2010s with the advent of powerful deep learning techniques that have transformed our ability to analyse big unstructured datasets of images, video and text. The share of deep learning papers in the data has multiplied four-fold since 2012.
- Technological disruption in AI has been accompanied by important changes in the geography of AI research. China has trebled its share of global of AI research since 2012, with a particular focus on cutting edge deep learning methods.
To build on this work we will:
- Consider other data sources in this work beyond arXiv, including research activity in traditional scientometric databases, as well as patenting, open-source software development and business activity.
- Develop our analysis of AI research trajectories, paying more attention to how multiple topics become part of, or splinter from, a trajectory over time.
- Explore mixed-methods research opportunities made possible by granular data like this.
Our analysis illustrates the potential of smarter data about smarter machines for understanding the recent evolution of AI research and for informing policies to steer it in societally-beneficial directions. In the coming months, we will be publishing follow-on analyses that consider the link between AI research trajectories and gender diversity in AI research teams, the implications of corporate participation in AI research, its regional dimension and the geography of AI R&D in techniques that could be used for mass surveillance.