econ.GNApr 12, 2026
SaveAI Patents in the United States and China: Measurement, Organization, and Knowledge Flows
Organizations: Department of Economics, University of Pennsylvania & the NBER, USA · Department of Finance, Durham University, UK · School of Economics and Management, Tsinghua University, China
Abstract
We develop a high-precision classifier to measure artificial intelligence (AI) patents by fine-tuning PatentSBERTa on manually labeled data from the USPTO's AI Patent Dataset. Our classifier substantially improves the existing USPTO approach, achieving 97.0% precision, 91.3% recall, and a 94.0% F1 score, and it generalizes well to Chinese patents based on citation and lexical validation. Applying it to granted U.S. patents (1976-2023) and Chinese patents (2010-2023), we document rapid growth in AI patenting in both countries and broad convergence in AI patenting intensity and subfield composition, even as China surpasses the United States in recent annual patent counts. The organization of AI innovation nevertheless differs sharply: U.S. AI patenting is concentrated among large private incumbents and established hubs, whereas Chinese AI patenting is more geographically diffuse and institutionally diverse, with larger roles for universities and state-owned enterprises. For listed firms, AI patents command a robust market-value premium in both countries. Cross-border citations show continued technological interdependence rather than decoupling, with Chinese AI inventors relying more heavily on U.S. frontier knowledge than vice versa.
Figures & tables
Figure 1: Model Performance by AI subcategories
Figure 2: Citation-Based Connectivity of AIPD and FGYZ Patents to Dual-Positive (Negative) Patents
Figure 3: Lexical Similarity of AIPD and FGYZ Patents to Dual-Positive (Negative) Patents
Figure 4: Model Performance on Chinese AI Patent Classification
Figure 5: Citation-Based Connectivity of Chinese and US Patents
Figure 6: Lexical Similarity of Chinese and US Patents
Figure 7: Comparing AI Patents: USPTO vs. CNIPA
Figure 8: Number of AI patents by Subcategory: USPTO versus CNIPA
Figure 9: Percentage of AI patents by Subcategory: USPTO vs. CNIPA
Figure 10: Geographic Distribution of AI Patents
(a) Chinese AI Patents
Figure 12: Spatial Diffusion of AI and Non-AI Patents
Figure 13: Innovation Concentration of AI and Non-AI Patents
Figure 14: Economic Value of AI and Non-AI Patents
Figure 15: Citation Propensity of AI and Non-AI Patents between the US and China
Table 1: Model Performance of Fine-tuned LLM Classifiers and Sample Composition
Table 2: Top Assignees with Most AI Patents
Table 3: Citation Propensity of AI and NonAI Patents between Different Assignee Types
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Type | Illustration | Example(abstract) |
| Knowledge processing (kr) | The field of knowledge processing contains methods to represent facts about the world and to derive new facts (or knowledge) from a knowledge base. | The performance of a given computer task is optimized by utilizing a plurality of intelligent agents suited to perform the computer task but having varied degrees of domain knowledge. Based upon an objective criteria that may be determined for a given situation, one of the plurality of intelligent agents may be selected and dispatched to perform the task, thereby optimizing the performance of the computer task for a wide variety of situations. (US6192354B1) |
| Speech (speech) | Speech recognition includes methods to understand a sequence of words given an acoustic signal. It often integrates signal processing techniques to reduce interference from factors like background noise. | A computer-implemented method and an apparatus are provided. The method includes obtaining, by a processor, a frequency spectrum of an audio signal data. The method further includes extracting, by the processor, periodic indications from the frequency spectrum. The method also includes inputting, by the processor, the periodic indications and components of the frequency spectrum into a neural network. The method additionally includes estimating, by the processor, sound identification information from the neural network. (US10062378B1) |
| AI hardware (hardware) | The field of AI hardware includes physical hardware designed to implement artificial intelligence software. It may include logic circuitry, memory, video, processors, and solid-state technologies. | A number of consecutive samples in unit distance code (here Gray code) are effectively stacked and supplied to respective sum and threshold devices 20 corresponding to each bit position, to determine the bulk property or “generic result” of the consecutive samples. This filtered output is converted back to binary and analogue if required. The digital filter may be used in any of the above applications to “clean” data. (US6519577B1) |
| Evolutionary computation (evo) | Evolutionary computation contains a set of computational methods utilizing aspects of nature and, specifically, evolution. | A multi-objective optimization method. The method comprises a population of objects for each objective utilizing an optimization process and determining a measure for the variation in values of each gene within each population. A crossbreed between objects from different ones of the populations is performed, wherein a selection of gene values for the child object is weighted based on the measures for the variations in the values of each gene within the respective populations. (US9047569B2) |
| Type | Illustration | Example(abstract) |
| Natural language processing (nlp) | Natural language processing contains methods for understanding and using data encoded in human natural language. | A system, method and computer readable medium for providing translated web content is disclosed. The method on an information processing system includes retrieving a first content in a first language and parsing the first content into a plurality of translatable components. The method further includes generating a unique identifier for each of the plurality of translatable components of the first content and queuing the plurality of translatable components and corresponding unique identifiers for translation into a second language. The method further includes, for each of the plurality of translatable components, storing a translated component and an associated unique identifier corresponding to the translatable component, thereby storing a plurality of translated components and corresponding unique identifiers. (US7627479B2) |
| Machine learning (ml) | The field of machine learning contains a broad class of computational learning models, including supervised learning classification models, neural networks, fuzzy logic, adaptive systems, probabilistic networks, regression and intelligent searching. | A neural network, which may be implemented either in hardware or software, is constructed of neurons or neuron circuits each having only one significant processing element in the form of an adder. Each neural network further includes circuits for applying a logarithmic function to its inputs and for applying an inverse-logarithmic function to the outputs of its neurons. The neural network utilizes a training algorithm which does not require repetitive training and which yields a global minimum to each given set of input vectors. (US5778153A) |
| Computer vision (vision) | The field of computer vision contains methods to extract and understand information from visual input, including images and videos. Areas of computer vision may include object recognition, manipulation (e.g., transformation, enhancement, or restoration), color processing, and conversion. | The present invention relates to a method and system for detecting biologically relevant structures in a hierarchical fashion, beginning at a low-resolution and proceeding to higher levels of resolution. The present invention also provides probabilistic pairwise Markov models (PPMMs) to classify these relevant structures. The invention is directed to a novel classification approach which weighs the importance of these structures. The present invention also provides a fast, efficient computer-aided detection/diagnosis (CAD) system capable of rapidly processing medical images (i.e. high throughput). The computer-aided detection/diagnosis (CAD) system of the present invention allows for rapid analysis of medical images the improving the ability to effectively detect, diagnose, and treat certain diseases. (US8718340B2) |
| Planning/ control (planning) | The field of planning and control contains methods to identify and execute plans to achieve specified goals. Key aspects of planning include representing actions and states of the world, reasoning about the effects of actions, and efficiently searching over potential plans. | In an embodiment, a processor includes a fuzzy thermoelectric cooling (TEC) controller to: obtain a current TEC level associated with the processor; obtain a current fan power level associated with the processor; fuzzify the current TEC level to obtain a first fuzzy fan level; fuzzify the current fan power level to obtain a second fuzzy fan level; determine a new TEC power level based at least in part on the first fuzzy fan level, the second fuzzy fan level, and a plurality of fuzzy rules; and provide the new TEC power level to a TEC device associated with the processor, where the TEC device is to transfer heat from the processor to a heat sink. Other embodiments are described and claimed. (US9857809B2) |
This figure presents the distribution of AI patents by the number of AI subfields to which they belong. We train eight classifiers, each corresponding to a specific AI subcategory, and apply them to patent texts to determine whether a patent belongs to each subcategory. Because these classifications are not mutually exclusive, a single patent may be assigned to multiple AI subfields. The histogram reports the percentage of AI patents that fall into one, two, three, or more subfields. The evolutionary computation category ( evo ) is excluded, so the number of subfields ranges from 1 to 7. “US-AIPD” refers to classifications provided by the USPTO AI Patent Dataset, while “US-FGYZ” and “China-FGYZ” denote classifications generated by our FGYZ classifier for U.S. and Chinese patents, respectively.
More in this topic
Classification
Classification maps observations to discrete labels, including familiar, unknown, hierarchical, or multiple categories.
See all 408 papers