Tracking AI in climate innovation


 Innovation in artificial intelligence (AI) is spreading rapidly in many areas of technology, and AI technologies may be of help to mitigate and adapt to climate change. However, previous studies of AI in the climate context mainly rely on expert judgement of the research literature, not large-scale data. Here, we present a new approach to analyzing the relation between AI and climate innovation on the economy-wide scale. We analyze over six million patents from the past 45 years from the United States, and find that the greatest amount of climate AI innovation has occured in transportation, energy, and manufacturing technologies. Green ICT and climate adaptation technologies is where AI innovations have higher shares, and breakthrough innovations have made up a larger share in adaptation technologies compared to technologies for climate mitigation. We estimate the difference that AI makes with statistical analysis: AI in mitigation and adaptation technologies is associated with 30-100% more subsequent innovations. Our approach provides new capabilities to track the exponential growth of AI in climate innovation.

nologies. 4-6 However, an increasing capability to automate and transform produc-23 tion, equip industries with new tools, and draw increasing government support, also 24 means that technological breakthroughs could lead to a higher demand for comput- 25 ing power, larger carbon footprints, shifts in patterns of electricity demand, and an 26 accelerated depletion of natural resources. [7][8][9] Whether the net effect of AI on the 27 climate system will be ameliorative or detrimental is currently an open question: 28 Concerns about the effects of AI have recently been followed by calls for new regula-29 tions and increased international oversight. 10-14 This suggests a need for increased 30 research and analysis capabilities to track, clarify, examine, and understand these 31 new technologies. Here, we present a new approach based on large-scale data to 32 track AI innovation in technologies that can contribute to climate adaptation and 33 mitigation. 34 The initial research into the connection between AI and climate change has often 35 rested on the framework of UN Sustainable Development Goals. Here, experts find 36 both positive and negative effects of AI. 4, 15-17 The UN goals provide a broad social 37 and political context, but covering them all might currently be too broad a task for 38 a single project or data source. In this study, we restrict ourselves to climate change 39 and use data sources for this specific issue to investigate the effect of AI innovation. 40 For climate change, expert analysis has previously suggested that machine learning 41 could have broad potential in both mitigation and adaptation strategies, with a 42 mixed message regarding the potential net effect on the climate system. 18 To examine if AI makes a difference in climate innovations, we choose to analyze 97 the forward citations subsequent patents make to a given patent after its publication.  Fig. 1a shows a steeper rise starting around 2010. Fig. 1b  in on the highly cited breakthroughs (the highest counts in Fig. 2), we see that most 124 highly cited patents appear to be non-AI innovations. In total counts, AI is related 125 to more innovations on average but has fewer of the highly cited breakthoughs.   with large-scale quantitative analysis. We see the potential to increase our use of 183 large-scale data sources to study whether AI innovations lead to the ends we seek. Estimated difference and 95% confidence intervals Figure 5: Estimated differences between the 99th percentile for AI and non-AI patents in adaptation amd mitigation technologies. For adaptation, AI technologies is clearly associated with more breakthroughs. For green ICT, production, and transport, the direction of estimates suggest similar results. However, wide confidence bands reflect a large statistical uncertainty because a small absolute number of AI innovations. For mitigation technologies, the jury is still out with respect to AI breakthroughs: More data are needed.

185
Our raw data are the texts of six million US patents from 1976 to 2019, from 186 which we extract and generate the most covariates. We label patents as climate

208
Our statistical investigation shows that the target count variable of forward cita-209 tions is zero-inflated, heavy-tailed, and related to grouping by different technologies.

210
The Poisson model served as a natural starting point for count data, but a direct 211 comparison of mean and variance showed that the target variable is over-dispersed.

212
To adjust for this, we also fit negative binomial models. The regression model for the conditional mean of forward citations y is specified 231 in short-hand notation which leaves out the term coefficients in β: