Quantifying the Effect of AI Invention on Productivity

Measuring AI-related invention: the PaLLaFi framework

The research project “Quantifying the Effect of AI Invention on Productivity” constructs a novel dataset linking U.S. patents granted between 1990 and 2024 to firm-level microdata, with the aim of providing new and more comprehensive measures of AI-related innovation that complement existing studies based on firm surveys and labor market data. Our earlier efforts to identify AI patents relied on traditional machine learning classification methods, but these proved inadequate once large language models (LLMs) and generative AI emerged as a major and rapidly growing category of invention. The project therefore started afresh using a new methodology--the Patent Labeling via Language Models and Fine-Tuned Inference (PaLLaFi) framework. The approach begins with a hand-labeled subset of patents identified by domain experts as AI-related and uses these to train a series of classification models, again evaluating their performance and accuracy. The best-performing models are then used to expand the training data, with humans kept in the loop to verify and correct machine-generated labels. The refined models are applied to the entire corpus of U.S. patents to produce a comprehensive classification across ten AI subdomains, including machine learning, natural language processing, robotics, and computer vision. These AI patents are subsequently linked to firm-level data from two sources: Compustat, which covers publicly traded U.S. companies, and confidential Census Bureau microdata covering essentially all U.S. firms. The Census data are connected to patent assignees via a carefully maintained crosswalk that tracks mergers, acquisitions, and divestitures. This linkage enables rigorous analysis of the productivity and employment effects of AI invention at two margins: the extensive margin, capturing what happens when firms begin patenting AI for the first time, and the intensive margin, capturing what happens as their stock of AI patents accumulates over time.

The geography and distribution of AI invention

AI patenting has grown very rapidly over the sample period. In the most recent years of the data, more than 20,000 new AI-related patents are generated annually by inventors in the U.S. and other countries. Two particularly sharp upticks are visible in the data: one following Geoffrey Hinton’s breakthrough developments in deep learning, and a more recent one coinciding with the introduction and rapid spread of transformer-based architectures. The pipeline of new AI-related invention shows no sign of slowing, and the slight dip at the very end of the data simply reflects the fact that data collection ended in the fall of 2024 before the full calendar year was complete. Importantly, AI patents are not concentrated in a narrow set of technology classes. They appear broadly across industries and patent classes, for example in medical devices, land vehicles, aviation and avionics, and many other fields, reflecting the character of AI as a genuine general-purpose technology whose applications span virtually every corner of the economy.

Within the United States, AI patenting is concentrated not only in the expected clusters of the Northeast Corridor and California, but also around Detroit, driven by self-driving vehicle research among automakers; Texas, where energy companies use AI to process seismic imaging data in oil and gas exploration; and Florida, reflecting AI applications in the space and aviation sector. This geographic diversity reinforces the general-purpose nature of the technology. Internationally, U.S. firms lead in absolute numbers of AI patents. This result partly reflects the tendency of firms to patent most intensively in their home market, but Japanese firms consistently rank second, ahead of both Chinese and South Korean inventors. The comparison is striking in specific cases: Sony and Toyota each hold more U.S. AI patents than Facebook, NEC holds more than Apple, and Canon holds more than Qualcomm. Japanese firms also conduct significant AI-related research and development activities outside Japan. Therefore, this count of patents filed by Japan-based inventors actually understates the full extent of Japanese AI inventive activity. In the early 1990s, when the absolute number of AI patents was much smaller, Japan’s share of the total was actually quite large, and it has declined only in relative terms as the global volume of AI patenting has exploded. In terms of the types of AI being patented, recent years show a clear convergence toward deep learning and transformer-based approaches assuming a dominant position, suggesting that a single family of upstream techniques is increasingly underpinning AI applications across diverse fields such as speech recognition, computer vision, and natural language processing.

AI invention and firm productivity: strong and broad-based effects

The core empirical results establish a strong and robust positive relationship between AI patenting and firm productivity. In panel data regressions using Census microdata, firms transitioning into AI-related patenting for the first time experience a 23% to 27% boost in labor productivity and slightly above 8% in total factor productivity (TFP). Along the intensive margin, a doubling of a firm’s stock of AI patents is associated with a 6% to 15% increase in productivity. These are statistically significant and economically meaningful results, especially given that many firms in the sample progressed from near-zero AI patents to hundreds or even thousands over the observation period, implying very large cumulative productivity effects. Similar positive and significant effects on both employment and labor productivity are obtained when the analysis is restricted to the extensive margin using the event-study difference-in-differences approach.

The event-study design matches each firm transitioning into AI invention for the first time to the closest counterpart in the entire Census database, which consists of millions of enterprises, that is similar in observable characteristics but that is not making a simultaneous transition to AI use. This matching approach represents the closest approximation to causal inference available outside of a randomized controlled trial. Productivity effects emerge with a one- to two-year lag after the initial patent filing, which is consistent with the time typically required for a firm to take a new technology from the research stage and work it into the products or services it delivers to customers. Once the effects materialize, they are substantial, reaching nearly 30% within five years of a firm’s transition into AI invention. Equivalent results obtained from Compustat data on publicly traded firms using standard Cobb-Douglas production function estimation confirm the Census-based findings. Across all specifications and data sources, the productivity effects appear broad-based rather than concentrated in a small number of high-profile firms or a particular sector.

AI invention and income inequality: no measured effect

One of the most prominent concerns in the academic and policy debate surrounding AI is whether it will widen income inequality, as earlier generations of information technology innovation are widely documented to have done by raising relative demand for high-skill workers and depressing wages at the lower end of the distribution. To investigate this, the research draws on the Longitudinal Employer-Household Dynamics (LEHD) database, a matched employer-employee dataset that exists for most U.S. firms across most states and tracks the full earnings distribution within firms over time. The impact of AI invention on within-firm income inequality is measured using three standard ratios from the labor economics literature: the 90th-to-10th percentile ratio, the 90th-to-50th percentile ratio, and the 50th-to-10th percentile ratio of employee incomes.

The findings are striking in their absence of effect. Neither the transition into AI-related patenting at the extensive margin nor the growth of a firm’s AI patent stock at the intensive margin is associated with any statistically significant change in within-firm earnings inequality across any of the three measures. This result holds even as the same specifications yield large and significant positive effects on productivity. It is important to note the limitations of this analysis: it examines only the income inequality within inventing firms and cannot speak directly to what happens to the firms that adopt AI-enabled technologies developed elsewhere, nor to economy-wide distributional dynamics. Nevertheless, the finding that strong productivity gains and unchanged within-firm income inequality coexist in the same data, for the same firms, at the same time, is a significant and perhaps somewhat reassuring result, and one that merits further investigation as AI adoption becomes more widespread.

Tracking elite AI scientists to measure uncodified innovation

Not all AI innovation is captured in patents, and some of the most economically consequential AI-related activity may leave no patent trail at all. Quantitative hedge funds, for example, hire large numbers of computer science PhD graduates from leading research universities at salaries of several hundred thousand U.S. dollars per year, suggesting they believe this investment in AI talent yields significant returns, but due to the nature of their industry, they produce no patents. To extend the measurement of AI innovation beyond what patents can capture, a closely related project attempts to track the movement of elite AI scientists, who are trained at the very frontier of AI research, across firms and geographies. The approach uses partnership data from Elsevier to identify the world’s top AI scientists based on their publication records, regardless of where they are located, and then uses faculty websites and other sources to map the doctoral and postdoctoral advisees of those scientists. The career histories of these advisees, once they enter the labor market, are tracked using the Revelio database, a U.S.-based service that compiles and structures scraped LinkedIn profiles covering more than half a billion individuals worldwide. By linking these elite scientist trajectories to Compustat firm data, the research asks whether the formation of a critical mass of frontier AI talent within a firm is associated with subsequent increases in AI-related patenting (in appropriate industries) and firm productivity. Preliminary results confirm that it is, and that the effects are large.

Extending this approach to Japan presents structural challenges. LinkedIn is not widely used in Japan, making the Revelio-based career history data far less comprehensive there than in the United States. Labor mobility in Japan is also substantially lower in general, which means that the movement of elite researchers across organizational boundaries is a less frequent phenomenon and therefore harder to detect and measure. These structural differences are not merely obstacles to comparative research but are themselves of substantive interest. The higher degree of labor mobility in the United States may shape the speed at which frontier AI science is translated into commercial products and services, and the much more stable employment patterns in Japan may produce a fundamentally different pattern of AI innovation, potentially one that is more incremental and more closely tied to the capabilities of established firms. Rigorous empirical investigation of these questions, making use of appropriate Japanese data sources and conducted in collaboration with Japanese economists, represents a promising and important direction for future research.

Comment

NAGAOKA Sadao:
AI exemplifies a general-purpose technology, and economic theory predicts that such technologies are typically supported by vertical disintegration—a division of innovative labor between upstream and downstream firms. Upstream firms, such as OpenAI and Google, develop general-purpose AI models exploiting enormous economies of scale in their creation, while providing services to a competitive and heterogeneous downstream market at relatively low marginal cost. The performance of downstream firms is therefore heavily shaped by the quality of the AI services they access, and broad user spillovers arise from the low-cost provision of those services. This vertical structure has important implications for how AI-driven innovation should be assessed.

On the question of how upstream firms protect their innovations, patents appear to play a less exclusionary role than in downstream technology sectors. Upstream firms tend to use patents defensively to preserve freedom to operate rather than to block competitors. OpenAI has pledged to use its patents only defensively, and a broader industry-wide effort, involving Anthropic, IBM, Meta, and Microsoft, is underway to establish a shared AI License Foundation. This suggests that the competitive advantage of upstream firms rests more on undisclosed know-how embedded in specific AI models than on formal intellectual property rights. Upstream and downstream firms may therefore require different analytical frameworks.

Regarding the methodology of using AI to identify AI patents, which constitutes a valuable contribution, the data clearly captures the impact of two landmark breakthroughs: deep learning in the early 2010s and transformers in 2017. It also reveals that AI inventions are distributed across a wide range of technology classes, not only in computing or physics. The Japan Patent Office (JPO) has conducted similar work identifying AI-core and AI-related inventions through patent classification codes and keyword searches on patent abstracts, with results that are broadly consistent in shape. The open question is how we can validate AI-assisted identification methods relative to these conventional approaches and to assess their reliability.

On the econometric modeling, firm performance depends on both AI and non-AI patents, and both should be controlled for in order to avoid confounding effects. Similarly, total R&D spending encompasses AI and non-AI components, a distinction that matters more for downstream firms. A further technical consideration is that R&D spending precedes patenting, and patents in turn precede observed productivity increases, which may introduce a timing artifact into estimates of productivity effects following AI adoption.

Lee BRANSTETTER:
Robustness checks not shown during the presentation do control for R&D spending, and prior work with an earlier patent database comparing AI versus non-AI patenting yields results consistent with those shown. The upstream-downstream distinction is an especially stimulating analytical framing. The dataset captures both segments, though numerically the majority of patents are downstream, meaning estimated average effects likely reflect downstream innovation more than upstream.

The observation that downstream patents cite both earlier patents and academic papers, including papers generated by major upstream entities, points to an underexplored body of data that could more formally link downstream invention to its upstream origins, an avenue now worth pursuing seriously.

On market concentration at the upstream level, many computer scientists anticipate a significant outflow of researchers from large AI companies into more focused ventures targeting specific industries or sectors. A leaked internal memo attributed to Google, titled “We have no moat,” captured a widely held view that the conceptual foundations of large language models are now broadly understood, lowering barriers to entry for specialists who wish to build focused models optimized for particular applications. If such a model proves superior to a general-purpose alternative within its target domain, the boundary between upstream and downstream innovation could begin to blur over time, which is a dynamic that should eventually become visible in patent and publication data.

Q&A

Q:
Japan has certain strengths in both patents and AI talent, yet these do not appear to be fully translating into productivity gains. Where do the main bottlenecks lie, and what policies or firm-level actions are needed?

Lee BRANSTETTER:
The entities in the United States that have been most successful at translating AI science into new products and services tend to be relatively young. IBM holds the largest single count of AI patents among the firms studied, having generated them since the earliest days of the field, yet has struggled to convert that scientific leadership into commercial products. Meanwhile, younger enterprises have been far more effective at taking foundational science and turning it into value. OpenAI itself was founded because researchers inside an established company could not get sufficient organizational backing for their ideas and chose to pursue them in a new venture instead. The critical enabling factor in the U.S. context is the openness of the system. This leads to a capacity to allow new enterprises, whether their ideas ultimately prove sound or not, to attract the human and financial capital needed to experiment. Japan’s corporate culture has many admirable qualities, but the relative difficulty for individuals with strong ideas to secure the resources they need outside established organizations represents a significant impediment to innovation, not only in AI but across many domains.

Q:
AI inventions span a wide spectrum, from large frontier models such as ChatGPT and Claude to small-scale prompt-based applications. Does the research account for such differences?

Lee BRANSTETTER:
This question goes to the heart of what downstream innovation looks like for a general-purpose technology of this kind. With earlier general-purpose technologies such as electricity, upstream invention, which often came from universities and large corporations, fed into downstream innovation inside separate firms. With AI, however, meaningful downstream innovation can happen at the level of the individual worker. An employee or small group can experiment with large language models and generate a new internal application or procedure that drives real productivity gains for the enterprise without producing a patent or a research paper. At present, there is no way to capture this activity through conventional data sources. The large technology companies are presumably collecting data on how users interact with their models, but obtaining access to it for research purposes poses significant challenges. There is likely an enormous amount of value being created through exactly this kind of low-level but highly practical prompt engineering, and it remains largely invisible to researchers.

Q:
Why has AI invention not yet had a measurable impact on income inequality?

Lee BRANSTETTER:
The analysis to date has focused on what happens inside the inventing firms themselves, which makes it difficult to track how the effects of AI diffuse outward to adopting firms and the broader economy. On the question of whether AI will ultimately widen inequality, the historical record of IT diffusion over the past 30 years provides some grounds for expecting a modest worsening, but the outcome is genuinely uncertain. One reason for caution is that AI, in important respects, makes very high-level analytical capability accessible to people across the full range of skill levels, potentially creating new competitive pressures on highly skilled workers while empowering those with lower formal credentials. At the same time, much high-level cognitive work is beginning to be automated, while tasks requiring physical dexterity in unstructured environments remain difficult for machines. Given these competing forces, the distributional consequences of AI diffusion remain an open question that the current data cannot yet resolve.

Q:
Does the income inequality analysis include board members, or only employees?

Lee BRANSTETTER:
Board members were not included, as they are not classified as employees under U.S. law and their compensation data is not captured in the underlying dataset. In principle, compensation data for board members of publicly traded companies exists and could be assembled, but it would represent a relatively small number of individuals who, in most cases, already have high incomes. The likely impact on the overall distributional findings would therefore be limited.