Tracking artificial intelligence in climate inventions with patent data

The study investigates how artificial intelligence is intersecting with climate change mitigation and adaptation technologies, addressing the need for systematic, large-scale tracking of AI’s role beyond expert-based reviews. AI promises efficiency gains and innovation but also raises concerns about increased energy use, carbon footprints, and resource impacts, risking techno-optimism that could delay effective climate policy. To provide empirical evidence at scale, the paper uses national patent data to track AI within climate-relevant inventions, aiming to quantify AI’s prevalence across climate technology domains, its association with subsequent invention activity, and how resulting knowledge spillovers distribute between climate and non-climate fields.

Prior work on AI and climate has largely consisted of expert-based assessments framed by the UN Sustainable Development Goals, identifying both positive and negative impacts of AI on climate mitigation and adaptation. Expert reviews can synthesize cross-domain insights despite limited data but face challenges in transparency of judgment and transferability of expertise across domains. Separately, patent data have been used to study trends in AI and in climate technologies. This study builds on and complements these literatures by combining AI and climate classifications within a single, large-scale patent dataset to move beyond qualitative expert assessments.

Data comprise over six million granted US patents from 1976–2019 sourced from USPTO full-text and classification files. Climate-related inventions were identified using the Cooperative Patent Classification (CPC) Y02 scheme, which covers technologies for greenhouse gas mitigation and climate adaptation. AI-related patents were identified using the World Intellectual Property Organization (WIPO) method that combines patent classification codes and keyword matching in key patent text sections (including terms such as machine learning, deep learning, and natural language processing). Patents labeled as both AI and climate were obtained by intersecting the two classification approaches. To evaluate technological impact, forward citations within three years after grant were counted for patents granted between 2010 and 2017. Breakthroughs were defined as the top 1% of patents by three-year forward citations within a technical domain and year. Count regression models (for forward citation counts) estimated the predictive difference associated with AI, controlling for grant year, technology area, and other factors consistent with prior work. As broader context, similar models were run on control groups of related but more general technological domains (e.g., buildings, electricity, smart grids, transport) not restricted to climate relevance. For knowledge spillovers, citing patents were categorized as climate or non-climate, and statistical tests compared the share of spillovers using a null model based on the hypergeometric distribution while accounting for domain size and age. For statistical power, carbon capture/storage and waste management were excluded from some analyses due to very small numbers of AI patents. Descriptive analyses included time trends in counts and shares, and disaggregation across Y02 technology areas (adaptation, buildings/housing, energy-efficient ICT, energy technologies, production/processing, transportation).

• Growth patterns: Counts of AI, climate, and climate–AI patents rose steeply, exhibiting exponential growth in the past decade. However, within climate patents the share of AI grew approximately linearly, and climate–AI counts were lower than expected under statistical independence of AI and climate labels.
• Distribution across technologies: Since 1976, more than half of AI-in-climate patents appeared in transportation, energy, and industrial production. Energy-efficient ICT and adaptation had smaller absolute counts but higher recent shares of AI. There were 4,390 AI–climate patents between 1976 and 2019, about 1.5% of all climate patents and 2.7% of all AI patents.
• Forward citations: Climate AI patents received more forward citations on average than climate non-AI patents. Regression estimates showed AI was associated with roughly 30–100% higher three-year forward citation counts across climate technology groups, with statistically significant effects. The largest increases were observed in buildings and energy-efficient ICT; adaptation and energy technologies showed weaker increases. Similar effect sizes and rankings appeared in control groups (buildings, electricity, smart grids, transport) not restricted to climate.
• Knowledge spillovers: Within climate patents, AI was associated with a smaller share of spillovers to climate patents and a larger share to non-climate domains compared with non-AI climate patents. The mosaic plot tests were highly significant (e.g., p < 2 × 10^−16), and similar patterns held both in aggregate and when disaggregated by Y02 areas.
• Breakthroughs: Although AI patents are fewer in number and thus had fewer total breakthroughs, the share of AI breakthroughs (top 1% by three-year citations) was higher in climate adaptation and transport than in other climate mitigation groups. Transport, which also had the most AI climate patents, showed the most observed AI breakthroughs. For most other mitigation areas, uncertainty remains due to wide confidence intervals.
• Additional context: AI climate patents exhibited increasing absolute counts with notable accumulation in transport, energy, and production; energy-efficient ICT and adaptation showed high AI shares. The analysis suggests that AI’s influence manifests as more subsequent inventions overall but with knowledge more frequently spilling into non-climate fields.

The findings indicate that AI within climate-related patents is linked to greater subsequent inventive activity, as reflected by higher forward citation counts, suggesting enhanced technological impact. However, the knowledge generated by AI–climate patents more often spills over to non-climate domains than comparable non-AI climate patents. This dual pattern implies that while AI may accelerate innovation, much of the benefit diffuses into broader technological areas rather than remaining within climate-focused innovation ecosystems. Notably, adaptation and transport stand out with higher shares of AI breakthroughs, pointing to domains where AI may be particularly transformative. These results underscore the importance of tracking both the magnitude of invention activity and the destination of spillovers to better assess AI’s net effect on climate-related innovation and to inform policy and investment strategies aimed at maximizing climate-relevant benefits.

This paper demonstrates a scalable approach for tracking AI’s role within climate mitigation and adaptation technologies using large-scale patent data and combined classification methods. Empirically, AI–climate patents are associated with higher forward citation impact and elevated breakthrough shares in adaptation and transport, but they also exhibit a greater tendency to spill knowledge into non-climate areas. Policymakers and stakeholders should therefore consider not only increases in inventive activity but also the distribution of spillovers across technological domains when evaluating AI’s contribution to climate objectives. The methodology can be extended to other countries and regions; future research should incorporate international patent systems, refine AI classification methods as technologies evolve, and complement patent-based indicators with non-patented innovation and deployment data to assess real-world climate impacts.

• Patent coverage and generalizability: Not all inventions are patentable or patented. Software and algorithmic innovations may require links to practical applications for patentability (US) or technical character (EU), and some AI innovations are disseminated via open-source or kept as trade secrets. The unknown share of non-patented AI inventions limits generalizability beyond patents.
• Jurisdictional scope: The study focuses on US patents, reflecting one legal and economic context; patentability criteria and innovation patterns may differ across countries.
• Field-specific incentives: Propensity to patent varies by industry (e.g., pharmaceuticals vs software), potentially biasing observed trends in AI across technological domains.
• Measurement via citations: Forward citations are proxies for technological impact and spillovers and may not perfectly reflect real-world usefulness or deployment.
• Data and classification limits: AI identification follows the WIPO method and CPC classes available up to 2019; classification errors or omissions may occur. Small sample sizes in some areas (e.g., carbon capture/storage, waste) limit statistical power and led to exclusions from some analyses.
• Time window: Breakthroughs and impacts are measured using a three-year citation window, which may not capture longer-term influence.