Week 5: Fostering Inventions and Innovations

[jamesallenevans]

Please post your memo that engages any of the week's readings about a theoretically interesting empirical case regarding artificial intelligence, innovation, and/or growth you anticipate will become the basis of your final project.

“Innovation Paradox”Links to an external site. 2024. Ufuk Akcigit. IMF Finance & Development Magazine.
"Where Have All the 'Creative Talents' Gone? Employment Dynamics of US Inventors," 2023. Ufuk Akcigit and Nathan Goldschlag. NBER Working Paper #31085.
“Abduction and the Logic of Scientific Advance.” 2021. Eamon Duede and James Evans.
“The Paradox of Collective Certainty in Science.” 2024. Eamon Duede and James A. Evans. Philosophy of Science. Alan Love, Ed. University of Minnesota Press.
“Destructive creation, creative destruction, and the paradox of innovation science”. 2022. Likun Cao, Ziwen Chen, and James Evans. Sociology Compass.
“Subjective Perspectives within Learned Representations Predict High-Impact Innovation” Likun Cao, Rui Pan, James Evans

Post by Thursday @ midnight. By 12pm Friday, each student will up-vote (“thumbs up”) what they think are the five most interesting memos for that session. These memos should: 1) test out ideas and analyses you expect to become part of your final projects; and 2) involve a custom (non-hallucinated) theoretical and/or empirical demonstration that will result in the relevant analytical visualization. Some of the top-voted memos will form the backbone of discussion in Friday's discussion sessions.

[verachaudhry]

R&D Spend Doesn't Translate to Firm Creation (measured through firm count)
Vera Chaudhry

Research Question
What is the impact of R&D spending on firm creation and growth?

Motivation: “The Innovation Paradox” discusses how R&D spending has increased significantly in recent years, but these spending increases have not translated to productivity gains. Specifically, Prof. Akcigit's focus on innovation has not increased due to increased spend, and large firms have used their investments to protect their market position. I looked at data for R&D grants from the NSF, which gives funds to a variety of institutions expected to use these funds to innovate. We would expect that when R&D funds increase, the number of new small firms also increases. Is this the case? This paper considers the empirical facts laid out in Prof. Akcigit’s paper to see whether NSF R&D funding is spurring new firm creation.

Research Design:
The visualization below shows that R&D spending has increased dramatically (espeically in certain periods like 2000-2010), but the number of firms (across different sizes) has not increased nearly as much.
I propose the following equations for a theoretical model to explain, consider and study this phenomenon (coded in LaTeX):

(a) The new knowledge (Kt) comes from the depreciation of old knowledge + investments in reserach.
(b) The birth rate of firms is potential entrepreneurs x the probability of the profit of the new firm exceeding the entry cost ft
(c) The number of active firms is the existing firms + new firm births - firm deaths

Unlike other models which focus on incumbent firms stealing share, this model considers the impact of firm cost of entry on new firm births. The theoretical economic model design considers a possible explanation (higher fixed cost of entry) for the reason R&D spend is not translating to new firm entry.

Data Visualization:
(using NSF data and BDS firms size data)

[erindner13]

Erik Rindner

Research Question

Do more experienced baseball managers use AI to promote strategic innovation, or primarily to reinforce traditional approaches to player usage?

Recent advances in artificial intelligence have expanded its role in professional baseball beyond scouting and player evaluation and into in-game decision-making and strategic planning. Teams increasingly rely on AI-informed models to guide lineup construction, bullpen usage, platoon matchups, and other tactical choices. At first glance, these tools appear to naturally encourage innovation by revealing patterns and strategies that humans would otherwise be unable to detect. However, baseball strategy has always been shaped by managerial leadership, experience, and attitudes toward risk. Managers are not neutral executors of data; they interpret and apply AI-driven recommendations through the lens of their own philosophies and past successes.

This raises a deeper question about how innovation actually emerges within baseball organizations. While AI systems can surface unconventional strategies, more experienced managers may be inclined to use these tools to optimize approaches they already trust rather than pursue disruptive change. For example, AI might suggest aggressive bullpen usage, frequent lineup adjustments, or unconventional platoons, yet veteran managers may selectively adopt recommendations that align with established norms around player roles and game flow. In this sense, AI does not automatically generate innovation but instead operates within boundaries set by leadership.

This dynamic closely mirrors the argument in the reading on aging and innovation, which shows that experienced leaders tend to favor incremental innovation over disruptive experimentation. As leaders accumulate experience, they become more effective at refining established practices but less willing to take risks on unfamiliar strategies. Applied to baseball, AI may function less as a catalyst for strategic transformation and more as a tool that reinforces existing managerial preferences. Rather than challenging intuition, AI systems may be filtered through it, producing continuity rather than disruption. In this view, innovation depends not only on technological capability but also on who controls and interprets that technology.

Methods

To study this question, I propose a comparative research design examining how managerial experience shapes the strategic use of AI-informed decision-making. Managers would be categorized based on tenure and career experience, and their teams’ player usage patterns would be analyzed using publicly available data. Key outcomes could include lineup variability, bullpen deployment frequency, platoon usage, and responsiveness to matchup-based recommendations across games and seasons. These patterns would then be compared across managers with differing experience levels to assess whether AI adoption is associated with experimentation or strategic conservatism.

The key outcome of interest is not simply whether AI influences managerial decisions, but how leadership determines the direction of that influence. AI may expand the menu of possible strategies, but experienced managers may prioritize stability over disruption when applying these tools. In an ideal setting, AI would support innovation by encouraging experimentation while still benefiting from managerial expertise. Whether AI ultimately reshapes baseball strategy or reinforces tradition may depend less on the technology itself and more on the leaders who wield it.

[bheim]

With the advent of LLM usage, are researchers converging towards less diverse, more similar references?

Since the introduction of GPT-4, I hypothesized that researchers are increasingly uses LLMs to aid their exploration and that would lead them to use increasingly similar references. Namely, as LLMs provide statistically likely responses, I would expect that it would provide similar references on topics.

Methods

To provide some initial analysis on this topic and building towards my final project to study how AI usage influences innovation and surprise metrics within industries, I used AlexNet data on psychology papers to observe the entropy of paper citations as it relates to the topics and papers that they cite. I also wanted to look at citation concentration — namely, are papers citing the same references or are they citing different references?

I used the introduction of GPT-4 as a cutoff date and then looked a couple years before its introduction and then to present day.

I looked specifically in the field of psychology. If anyone has thoughts on a better field to look at, what would be helpful! I think it would be useful to expand this analysis

Results

The initial results suggested that my convergence theory was correct. The topic and paper entropy of new papers is decreasing. However, at the same time, the total number of references per paper has decreased substantially, so I hypothesized that this could be driving the entropy reduction.

When normalizing entropy to a per reference basis, I found this hypothesis to be correct. Post-GPT-4 papers actually have more entropy per reference. Moreover, when looking at citation concentration, the gini coefficient for citations is decreasing, suggesting more equality in paper citations.

Is this effect driven by GPT-4. This analysis doesn’t allow for us to make confident conclusions on that. Instead, I would want to look at a bigger window to see if this is a shift in a trend or just a continuation. Moreover, I would want to look deeper into how this effect differs by levels of AI exposure for each field.

[Arthur-Langlois]

Antitrust Enforcement, Incumbent Size, and Innovation: A Comparative Assessment

Research Question: Does antitrust regime enforcement increase innovation through incumbent firm size?

I setup and test the hypothesis that antitrust regulations increase innovation by reducing firms' relative sizes and thus incentivizing innovation-based growth. This is in particular because firms when unable to growth through anti-competitive practices and "inorganic" growth have an incentive to innovate to gain market shares.

Literature: A large economics literature links innovation to incentives, competition, and institutions. Recent models and evidence emphasize an inverted-U between competition and innovation (Aghion et al., 2005) and show that stronger patent rights can shift innovation incentives toward incumbents (Acemoglu & Akcigit, 2012). We largely build our hypothesis on Akcigit et al. (2023), which document that larger firms tend to engage in non-competitive behavior. On antitrust and firm size, research connects enforcement to market concentration and dynamic competition: tougher merger policy and scrutiny of dominant firms can reduce “kill zones” and preserve entry (e.g., Cunningham, Ederer & Ma, 2021), while evidence on common ownership and product-market concentration suggests that weaker competitive pressure can raise markups and tilt innovation toward large incumbents (Azar, Schmalz & Tecu, 2018; De Loecker, Eeckhout & Unger, 2020). The literature has extensively studied the relationship between antitrust policies and firm size, as well as specific drivers of innovations, it hasn't focused on the explicit causal link between antitrust regulation, firm size and innovation levels.

Data: We build a 20 year comprehensive country-level dataset from multiple public sources to study antitrust regimes, firm sizes, and innovation levels. In particular, we measure antitrust and competition law using the (1) competition law index, and (2) antitrust enforcement, both collected form the Comparative Competition Project This sources allows us to measure antitrust regimes across 131 countries for years 19902-2010.
In addition, we measure mean and median firm size in given country-years using the World Bank Enterprise Surveys. Lastly, innovation is measured by patent application, share of new firms, both compiled by the World Bank.

Method: We test our specific mechanism using an Instrumental Variables (IV) approach. Our first stage looks at the impact of antitrust on firm size, and our second stage studies the impact of firm size on innovation. This design allows us to exogenize firm size variation by using antitrust regulations as an instrument, and then study its relative impact on innovation. Our argument relies on the two following assumptions: antitrust laws only impact innovation through firm size (exclusion restriction) and antitrust enforcement has a significant effect of firm size (validity condition).

More formally, our design follows the following 2SLS procedure:

$$\text{Firm Size}_{ct} = \gamma_0 + \gamma_1 \text{Antitrust}_{ct} + \gamma X_{ct} + \epsilon_{ct}$$

$$\text{Innovation}_{ct} = \beta_0 + \beta_1 \text{Firm Size}_{ct} + \beta X_{ct} + u_{ct}$$

where $X_{ct}$ denotes a set of country-level characteristics. The causal effect of firm sizes on innovation resulting from shifts in attributed to antitrust laws is our main coefficient of interest $\beta_1$.

[harve27]

Understanding the dynamics between innovation, politics, and growth in the context of China's EV industry

Background/motivation
In his 2024 article, "The Innovation Paradox," Akcigit notes that innovation, productivity gains, and economic expansion have slowed despite increased R&D spending among firms. He attributes this result to a few factors:

• Firm concentration (relative to their size, small firms innovate more than large firms, but small firms were crowded out by giant enterprises)
• Leading companies focus on maintaining dominance through actions like securing political connections rather than fostering innovation (e.g., Ackigit et al.'s study of Italian firms)
• A reallocation trend where inventors are moving to large firms, where they tend to be less innovative compared with inventors that move to young firms (a trend like this happens through innovation-stifling hiring)

One industry to examine these questions is China's electric vehicle industry, which supports the biggest EV market in the world. This industry was seeded by industrial policy, and has birthed a hyper-competitive industry involving a mix of established players, state-owned behemoths, and nimble startups. Even more interestingly, many of these companies perform work across the EV stack (e.g., BYD batteries as well as cars). New entrants have emerged from consumer tech (e.g., Xiaomi). We can answer some interesting empirical questions about innovation and growth in the context of the EV industry; we can test out some of Akcigit's theories in the context of a hyper-competitive industry.

Here are some of the research questions we can answer:

1. What is the relationship between size and R&D investment? Do small companies invest more in R&D relative to their size? (i.e., do smaller companies innovate more than larger companies)
2. What is the relationship between R&D investment and how state-influenced a company is? (i.e., as a state's ownership stake in a company increases, does that company become less innovative? Investigating the political economy of innovation)
3. Do companies that invest more in R&D tend to have higher revenues?
4. How has R&D spending changed over time across the EV industry?

Methodology
We first make a comprehensive list of all of the companies in the EV stack; then, for public companies, we use publicly available reports to gather data about their size, revenue, and R&D investment. We then link this data to WireScreen, a platform which provides a detailed database of Chinese companies, their investors, and their relationship to other companies and the state. In particular, we use their "Government Ownership" measure, which provides a percentage of how much a company is owned by the state. Most companies have ties to the state, although these are often difficult to determine because they try to obfuscate their state ties through complex ownership structures. We treat the percentage of ownership as a proxy for influence/support of the state.

Results
We have 133 companies in the database, but only 55 of them are public, and for those that are public, we could only recover 42 companies with data. 24 of the companies are state-owned. Akcigit's hypotheses seem validated. Companies with higher state-ownership spend less on R&D, and smaller companies invest more in R&D than large companies. I've attached a really interesting report that Claude Code wrote up with this data!

EV_RD_Analysis_Report.pdf

Some key findings:

• Smaller companies invest more intensively in R&D; (7.2% of sales) compared to large companies (3.6%)
• Industry-wide R&D spending grew at 22.8% CAGR from 2016-2024, reaching ¥120B in 2024
• Companies with higher R&D intensity show stronger revenue growth trajectories
• The median R&D intensity across the industry is 4.0% of revenues

[felicechan]

Research Question
In pharmaceuticals, does generative AI increase exploratory innovation (new targets, new mechanisms) or does it accelerate safe development by reinforcing established paradigms?

Motivation and Theory
Pharma is a clean setting for the innovation paradox: R&D spending is enormous, yet breakthrough output is uneven and often dominated by incremental advances. This week’s epistemic readings offer a mechanism: evaluation pressure and community convergence can suppress abductive search and push researchers toward familiar, legible directions like nostalgia and collective certainty. AI could relax these constraints by lowering search and experimentation costs, or intensify them by standardizing workflows and making dominant approaches even more attractive.

Research Design
The empirical analysis examines pharmaceutical innovation at the therapeutic-area, by-year level, enabling comparison of exploratory outcomes across domains with different exposure to generative AI tools. The sample spans roughly 2010-2026, capturing pre-trends and the post-2022 diffusion of generative AI in drug discovery.

AI adoption is measured using a therapeutic-area AI exposure index, constructed from the share of publications, patents, and clinical trial descriptions that reference AI or machine learning based discovery methods, including generative modeling, protein structure prediction, and virtual screening. Therapeutic areas are classified as high or low AI exposure based on their pre-2022 receptivity to AI methods, allowing AI diffusion to be treated as a differential shock rather than a uniform time trend.

Exploratory innovation is measured along two primary dimensions. First, target diversity captures the breadth of biological targets pursued within a therapeutic area in a given year. This is operationalized using entropy-based measures of molecular targets referenced in patents, publications, or trial registries. Higher target diversity indicates broader abductive exploration of disease mechanisms, while declining diversity reflects convergence around established targets. Second, breakthrough output is measured using indicators of first-in-class innovation, such as the share of drugs introducing novel mechanisms of action or receiving first-in-class regulatory designation, capturing whether exploration yields transformative outcomes rather than incremental follow-ons.

The main identification strategy employs a difference-in-differences design, comparing changes in target diversity and breakthrough output before and after 2022 between high and low AI exposure therapeutic areas. Area fixed effects control for time invariant differences across domains, while year fixed effects absorb aggregate shocks to pharmaceutical R&D. Event study specifications estimate dynamic effects and allow for formal testing of parallel pre-trends.

Supplementary analyses examine target concentration and citation age to assess whether AI accelerates epistemic convergence. Finally, robustness checks split the sample by firm type: incumbent pharmaceutical firms versus smaller biotech entrants, to test whether organizational constraints mediate AI’s impact on exploratory innovation.

Data Visualization
The main visualization is an event-study plot showing changes in target diversity and breakthrough output before and after the diffusion of generative AI in 2022, comparing therapeutic areas with high versus low AI exposure. Diverging post-2022 trends illustrate whether AI adoption expands exploratory search or accelerates convergence on established targets.

[tanyadholakia-byte]

Memo: Spatial Density & Innovation Efficiency in Singapore

Research Question

Does Singapore’s spatially concentrated high-tech ecosystem increase breakthrough innovation by facilitating "small-team abduction" relative to more geographically distributed collaboration models?

Motivation
While remote collaboration facilitates scale and execution, it may hinder the early-stage conceptual integration required for disruption. Wu, Wang, and Evans demonstrate that small teams are the primary engines of disruptive science, whereas large teams tend to develop existing ideas. Duede and Evans further argue that such discoveries rely on "social abduction" (the iterative resolution of anomalies through face-to-face interaction) which is a mechanism that Singapore has explicitly attempted to institutionalize by engineering high-density districts designed to force-multiply these "creative collisions."

The Singapore Context

Singapore provides a unique environment for testing theories of spatial innovation because of its deliberate urban planning. Unlike organic tech hubs such as Silicon Valley, Singapore has engineered dense, quaternary industry clusters like the One-North precinct (Biopolis, Fusionopolis, Mediapolis). These hubs are designed to physically co-locate private R&D, university labs (NUS/NTU), and government agencies within walking distance. This extreme spatial density theoretically lowers the coordination costs of "social abduction" (Duede & Evans), allowing small teams to resolve complex, interdisciplinary anomalies through frequent, informal interactions. These dynamics are often diluted in distributed or remote teams.

Research Design

To rigorously test this, I propose a comparative observational study expanding on the pilot data below.

1. Data & Classification: I will construct a dataset of Singapore-affiliated papers (2015–2023) using OpenAlex. Teams will be classified not just as "Co-located" vs. "Remote," but by precise geospatial distance (e.g., "Same Precinct" vs. "Same City" vs. "International").
2. Dependent Variables: Beyond raw citations, I will compute the CD Index (Disruption) to distinguish between developmental science (consolidating status quo) and disruptive science (eclipsing prior work), as defined by Wu et al.
3. Mechanism (Perspective Diversity): To confirm that breakthroughs are driven by diverse inputs, I will use text embeddings (SPECTER) to measure the semantic distance between authors' prior works. The core model will test if co-located small teams with high semantic diversity produce the highest disruption scores, controlling for field, year, and author fixed effects.

Preliminary Evidence (Pilot Study)

To validate the feasibility of this design, I conducted a pilot analysis of $N=1,927$ Singapore-affiliated papers (2015–2021), specifically filtering for small teams (2–5 authors) to isolate the group dynamics most conducive to disruption.

As shown in Figure 1, preliminary results strongly support the hypothesis. When controlling for team size by measuring Innovation Efficiency (citations per author per year), co-located Singaporean teams significantly outperformed distributed teams.

• Co-located teams had an 18.0% probability of being "High Efficiency" hits (top 15% of sample).
• Distributed teams had only a 13.6% probability.

This suggests that while remote collaboration may offer scale, Singapore’s co-located ecosystem offers a "density dividend," resulting in a higher per-capita efficiency in generating impactful ideas.

Table 1: Descriptive Statistics of Pilot Sample

Team Structure	N (Papers)	Mean Efficiency (Cites/Author/Yr)	Hit Rate (Prob. > Top 15%)
Co-located (Singapore)	629	0.18	18.0%
Distributed (Remote)	1,298	0.14	13.6%

Figure 1: Probability of "High-Efficiency" Breakthroughs by Team Structure. Data: N=1,927 Singapore-affiliated papers (2015–2021). "High Efficiency" is defined as the top 15th percentile of citations-per-author-per-year. Co-located small teams (blue) show a higher probability of producing efficient breakthroughs than distributed teams (grey), supporting the hypothesis that spatial density aids small-team abduction.

[ShiyangLai]

Shiyang Lai

Research Question

How does a large language model embed a scientific subject as a non-Euclidean knowledge manifold, and can misalignment between embedding distances and graph distances reveal structural holes within that manifold?

Research Design

This project studies the geometry of scientific knowledge as represented in large language model (LLM) embeddings, focusing on how coherent subjects are embedded as manifolds rather than point clouds in Euclidean space. Building on work that frames scientific advance as abductive movement across epistemic gaps (Duede & Evans 2021) and highlights the paradoxical role of collective certainty in stabilizing knowledge domains (Duede & Evans 2024), I treat learned representations as maps of past successful knowledge, whose internal structure can be interrogated.

Empirically, I begin with a corpus of scientific papers within a well-defined subject area (e.g., a field or subfield identified via journals or classification codes). Each paper is embedded using a pretrained LLM, producing a high-dimensional semantic representation. Rather than assuming the resulting space is Euclidean, I interpret the embedded papers as samples from an unknown manifold shaped by historical patterns of discourse and success.

To probe the manifold’s structure, I construct a graph over the embedded points, where edges connect local semantic neighbors (e.g., k-nearest neighbors in embedding space), forming a neighborhood graph that approximates the manifold’s intrinsic geometry. Graph distance on this neighborhood graph captures how papers are connected along the manifold, while cosine or Euclidean distance captures extrinsic similarity in the ambient embedding space.

The key empirical object is the alignment (or misalignment) between extrinsic embedding distance and intrinsic graph distance. In a well-sampled, smoothly embedded manifold, nearby points in cosine space should also be close in graph distance. However, when two papers are semantically similar but connected only via long or fragile paths on the neighborhood graph, this suggests a hole or underexplored region in the manifold: a place where the language model encodes similarity, but the empirical structure of knowledge provides weak support or continuity.

Rather than defining holes as low-density regions in embedding space, this design identifies holes as geometric distortions—regions where local neighborhoods fail to align with global semantic proximity. This avoids Euclidean assumptions and directly operationalizes abductive tension as a property of representational geometry, not innovation outcomes.

Illustrative Figure

[johnab20]

John Butka

Research Question

Does the commercial scaling of "See & Spray" AI technology represent a force of creative destruction that threatens the revenue models of incumbent herbicide manufacturers, and do these incumbents utilize non-productive strategies to mitigate the shift?

Research Design

Drawing on Lin, Frey, & Wu (2023), I propose a "Collaboration Proximity Analysis" to test if the physical distance between inventors correlates with the disruptive impact of their patents. While John Deere’s "See & Spray" technology was birthed in concentrated AI and engineering hubs (facilitating the "on-site fusion" of ideas), Bayer AG maintains a massive, multi-national R&D network of 15,900 scientists. Following the findings of Lin et al. (2023), I hypothesize that Bayer's reliance on remote or geographically dispersed teams leads to a "coordination tax" that results in fewer breakthrough ideas compared to Deere’s localized, agile squads.

Furthermore, I will incorporate the "Aging and Narrowing" framework from Cui et al. (2025). This part of the design involves a longitudinal assessment of the "citation age" and "topic breadth" within each firm's patent portfolio. By measuring the "scientific distance" between the foundational research cited in John Deere's AI patents versus Bayer’s chemical patents, I aim to observe whether Bayer is experiencing the "narrowing" effect—where aging firms increasingly focus on refining existing knowledge (incrementalism) rather than pursuing the "abductive" leaps required for disruption. This design will use 2024–2026 patent data to compare the CD Index (Disruption Score) against the Mean Inventor Distance and Citation Age for both firms.

The competitive struggle here is not just about raw R&D spend, but the architecture of collaboration. If Lin et al. (2023) are correct, Bayer’s €5.8 billion budget is partially offset by a "remote collaboration" penalty. Their global scale necessitates a decentralized structure that is statistically less likely to produce the radical "See & Spray" logic that decouples weed management from chemical volume. By contrast, John Deere’s 5,000 specialists operate within a "flatter," more localized engineering culture. This structure avoids the "scientific narrowing" typically observed in aging incumbents, allowing Deere to leverage high "Investment Density" per specialist to achieve the creative destruction currently threatening legacy chemical revenue streams.

Visualization

[HenryPinner7]

Does AI tool adoption reduce the tendency of large teams to produce less exploratory/breakthrough innovation in clean-tech?

Background: Evans et al. (2019) show that small teams are more likely to generate disruptive breakthroughs while large teams skew toward coordinated, developmental work, so my baseline hypothesis is that novelty falls as team size grows. This memo builds on that finding by asking whether it still holds in today’s research environment, where AI platforms (LLMs, coding copilots, literature-discovery tools) may reduce coordination frictions and expand teams’ ability to search across ideas. If AI genuinely improves ideation and recombination at scale, it could plausibly flatten the traditional “large-team penalty” for exploratory innovation.

Research Design: To answer this question, I would build a patent-based panel of clean tech inventions. The sample would include patents in climate-related technology classes (e.g., CPC Y02 and related subclasses for batteries, solar, etc.) from 2016-2026. Each observation would be a patent i assigned to a firm/assignee f in year t, and team size would be measured directly as the number of inventors listed on the patent. To define “large teams” in a way that is comparable across technologies with different norms, I would classify a patent as LargeTeam = 1 if its inventor count falls in the top quartile (≥ 75th percentile) within technology area × year, with robustness checks using intuitive bins (1–2, 3–5, 6–10, etc.) and a continuous ln(TeamSize) specification.

Because patents do not record AI use, the key additional input would be a short survey of clean-tech R&D and IP leads (or inventor-team leads) to measure AI adoption intensity. The survey would build an index from frequency of use and scope of tasks (e.g., ideation/brainstorming, literature and prior-art search, modeling/coding, prototyping, and drafting claims), plus the timing of adoption; responses would be linked back to assignees (and, where possible, inventor teams). Innovation outcomes would be measured objectively from the patent record: (i) exploration/novelty via rare or new-to-field CPC pairings and novel keyword combinations (extracted from patent titles/abstracts using a reproducible TF-IDF method, then scored by the rarity/newness of keyword pairs within technology×year or relative to the firm’s prior portfolio), and (ii) impact via technology-year-normalized forward citations and a top-tail indicator (e.g., top 5% cited within tech×year). The main test is whether AI moderates the aforementioned size relationship in teams by estimating a regression of the form:

Yift ​= β1​ln(TeamSizei​) + β2​AIIndexf ​+ (β3​ln(TeamSizei​) × AIIndexf )​+ controls

β1 is the baseline relationship between team size and the outcome Y when AI adoption is zero, β2 is the effect of AI adoption intensity on Y when ln(TeamSize)=0 — i.e., when team size equals 1 inventor, and β3 is the interaction term: it tells you how AI adoption changes the slope of team size. A positive β3 would suggest AI helps large clean-tech teams remain exploratory and/or high-impact (higher AI adoption flattens the usual decline in novelty or breakthrough impact as teams become larger).

Figure: The simulated figure illustrates the basic pattern I would test in the data, plotting predicted novelty across inventor team-sizes bins for low versus high-AI adopters. In this version of the simulated data, the low-AI line slopes downward, illustrating the usual “large-team penalty” in exploratory innovation, while the high-AI line is noticeably flatter, which would be consistent with the hypothesis that AI platforms help large teams sustain more novel recombination. If the blue line were above the orange line, the opposite would be true. In this way, this memo aims to see if it is possible to depict whether AI adoption attenuates the negative relationship between team size and innovation novelty in clean-tech patenting.

[nbcodes57]

Nick Baumgartner

Research Question:
Do larger firms structure their patent inventor teams more hierarchically, and does this help explain why large-firm patents tend to be less disruptive?

Background and Motivation:
Xu, Wu, and Evans (2022) demonstrate that the internal hierarchy of scientific teams—independent of their size—shapes the character of the research they produce. Using 89,575 self-reported contribution statements from PNAS, Nature, Science, and PLOS ONE, they construct an “L ratio” measuring the fraction of team members in leadership roles (conceiving, designing, supervising, writing) relative to total team size. Flat teams (high L ratio) produce more novel work, generate greater long-term citation impact, and boost productivity for the average team member. Hierarchical teams (low L ratio) benefit lead authors but suppress innovation, favoring incremental development and short-term citations. Crucially, these results hold within person: the same researcher on the same-sized team produces more disruptive science when the team is organized flatly. The authors also document that team hierarchy has risen sharply over the past half century, with the share of papers having L ratios below 0.5 growing from 50% in 1950 to 70% in 2015.

This finding has a natural extension to the private sector. Wu, Wang, and Evans (2019) showed that small teams disrupt while large teams develop in both science and patenting, but the mechanism has remained ambiguous—is it size, or the hierarchy that comes with it? If Xu et al.’s logic applies to corporate R&D, then the inventor team structure on patents should matter independently of team size. Large firms may not only file patents with more inventors, but may also organize those inventors more hierarchically, with a single lead inventor doing the conceptual work while others contribute narrowly. This would compound the drag on disruption: large firms already tilt toward incremental innovation through their market incentives, and if their internal team structures are also more hierarchical, the organizational form itself may further suppress novelty.

Research Design:
I use USPTO utility patent data matched to assignee firms, focusing on patents granted between 2000 and 2020. Unlike scientific papers, patents do not include contribution statements, so I cannot directly compute an L ratio. Instead, I construct a proxy for team hierarchy using the concentration of inventor contributions. Specifically, for each patent’s set of listed inventors, I look at their broader patenting history to identify whether one inventor dominates the team’s overall patent output (a “tall” team) or whether inventors have roughly equal track records (a “flat” team). I measure this using a Herfindahl-style concentration index across each team’s cumulative patent counts at the time of filing. I classify patents as having “flat” inventor teams when this concentration index falls below the median. I then calculate the share of flat-team patents across firm size bins, using employee counts from Compustat to categorize assignee firms.

Once the initial analysis is complete, I plan to incorporate the CD disruption index (Funk and Owen-Smith, 2017) to test whether flat-team patents are indeed more disruptive, controlling for firm size, technology class, and year. I also aim to conduct robustness checks using alternative hierarchy proxies—such as whether the team includes a single repeat “lead” inventor who appears on a disproportionate share of the firm’s patents—and to validate the measure against the subset of patents where inventor ordering may carry hierarchical information.

Below is an initial visualization based on the methodology above. It shows that the share of patents with flat inventor teams decreases monotonically with firm size, consistent with the hypothesis that larger firms organize their R&D teams more hierarchically. That being said, this is a preliminary cut and does not yet control for technology field, patent team size, or time period. It also does not yet test the link between team structure and disruption, which is the core question.

[zararania]

Zara Rania Suran

Research Question
Does the growing use of AI-driven tools in biomedical research concentrate scientific effort on high-consensus molecular targets and model systems, thereby reducing the rate of exploratory, high-impact biological innovation?

Research Design
Biomedical discovery has historically relied on abductive reasoning: scientists generate new hypotheses by navigating uncertainty, disagreement, and incomplete mechanistic understanding (Duede & Evans 2021). Yet recent work in the science of science suggests that modern research systems increasingly favor convergence over exploration. Akcigit’s Innovation Paradox (2024) documents slowing frontier growth despite rising R&D investment, while Akcigit and Goldschlag (2023) show declining dynamism among inventors. Duede and Evans (2024) argue that this stagnation reflects a rise in “collective certainty,” where shared beliefs narrow the space of plausible inquiry.

Artificial intelligence has the potential to strengthen this dynamic in biology. AI systems used for target discovery, functional genomics, and protein-disease association prediction are trained on historical data that encode prior scientific attention. As a result, learned representations may preferentially surface genes, pathways, and model organisms that are already well studied. Cao, Pan, and Evans show that learned representations embed subjective perspectives that can predict high-impact innovation-but only when heterogeneity is preserved rather than averaged away. In biomedicine, AI may instead homogenize perspectives.

With this project, I propose to test whether AI adoption in biomedical research reallocates scientific labor toward high-consensus biological objects, such as frequently studied genes, well-characterized pathways, or dominant model organisms at the expense of exploratory work on poorly understood systems. The empirical design compares biomedical research outputs before and after AI-tool adoption across subfields, measuring (i) concentration of attention across genes or pathways, (ii) distance from canonical biological mechanisms, and (iii) long-run impact. My hypothesis is not that AI reduces productivity, but that it shifts innovation toward safer, lower-variance biological contributions, consistent with the innovation paradox.

Simulated Visualization
The simulation below illustrates this mechanism. AI-assisted biology increases output but concentrates attention on high-consensus targets, reducing the probability of extreme-impact discoveries that emerge from abductive exploration of uncertain biology.

In the data, AI-assisted labs might disproportionately prioritize genes such as IL6, TP53, and TNF-highly connected targets with extensive prior literature-because these targets dominate training data and produce strong, reproducible statistical signals. On the graph, these projects appear as a dense cluster at high target popularity with moderate, tightly distributed impact values. In contrast, human-guided projects occasionally pursue poorly annotated targets, such as a long non-coding RNA with sparse prior work, that appear as rare points at low popularity but exceptionally high impact. The key empirical pattern is that while AI-assisted research might increase the mass of points in the consensus region, it might also underproduce the extreme upper-left outliers that drive long-run biological innovation and growth.

[lianduann]

Lian Duan

Research Question
Does the adoption of Generative AI (GenAI) induce a state of “Synthetic Gerontocracy” in scientific production?

Specifically, does GenAI decouple demographic diversity from Perspective Diversity, creating a paradox in which an influx of new researchers coincides with Perspective Collapse—a shift from risky, disruptive abduction toward safe, consensus-based recombination?

Motivation
Recent literature suggests GenAI may reshape the epistemic structure of innovation through three mechanisms:

1. Synthetic Gerontocracy (Artificial Aging). Cui et al. show that as scientists age and accumulate the “burden of knowledge,” they shift from disruption to recombination. GenAI—trained on large-scale historical corpora—may transmit a similar burden to junior researchers, “aging” their cognitive output toward safer developmental work.
2. Perspective Collapse. Cao et al. argue innovation depends on Perspective Diversity (distinct interpretive angles), not merely demographic background diversity. GenAI may democratize access to information yet homogenize interpretation by converging on statistically likely outputs, narrowing the space of plausible ideas.
3. Certainty Trap. Duede & Evans suggest hyper-collaboration can increase collective certainty while reducing the discovery of anomalies. GenAI may function as a consensus engine that generates highly plausible text, potentially crowding out the “alien” deviations that trigger paradigm shifts.

Proposed Research Design
The empirical strategy combines citation network analysis with a pre/post comparison around the 2022–2023 GenAI adoption surge.

1. Key Metric: The Disruption Index (D-score) (Wu et al., 2019), which distinguishes Development (consolidating the status quo) from Disruption (rendering prior references less central).
2. Event-Study / DiD Framework: Using the sharp increase in GenAI availability and salience following the release of ChatGPT (Nov 2022) as a temporal breakpoint, I will compare D-score trends in High-Exposure versus Low-Exposure fields. High-exposure fields will be defined either (i) by a field-year GenAI penetration proxy (e.g., GenAI/LLM disclosures or keyword mentions in acknowledgements/methods) or (ii) by pre-registered field categories known to be early adopters. The primary hypothesis is a relative decline in D-scores after 2023 in high-exposure fields, consistent with a shift toward recombination.
3. Identification: Biological vs. Algorithmic Age: To test the “Synthetic Gerontocracy” mechanism, I will control for authors’ biological/career age (following Cui et al.) and estimate heterogeneous effects by seniority. The critical prediction is that junior researchers in high-exposure fields exhibit a disproportionate post-2022 decline in D-scores—i.e., their output becomes structurally more “senior-like.”
4. Implementation Notes (Feasibility): To address citation-window concerns, D-scores will be computed using a fixed citation horizon (e.g., 18–24 months post-publication) and validated with pre-trend checks and placebo breakpoints.

Visualization
To illustrate the proposed “2022/2023 GenAI shock” design, I present a hypothetical event-study style trend plot of the annual mean Disruption Index (D-score) for High-Exposure vs Low-Exposure fields (2018–2025). A vertical reference line marks Nov 2022 / 2023 (ChatGPT release and the first full post period), and shaded bands represent 95% confidence intervals around group means. This figure is illustrative only—it uses simulated values to show what the empirical pattern would look like if GenAI adoption induces “Synthetic Gerontocracy.” In the real analysis, the same visualization will be generated from observed citation-network D-scores and field-level exposure classifications.

[keshavmkapoor]

Keshav Malik Kapoor

Research Question: Can government interventions in healthcare and education reduce the large-firm wage premium that discourages inventors from joining small businesses?

Motivation: In The Innovation Paradox, Prof Akcigit documents a wage premium that discourages inventors from joining small businesses: large firms can simply pay more. There’s an intuitive micro-level response—small firms can offer stronger equity compensation (e.g., stock options) to offset lower salaries and attract inventive talent. This memo asks whether government policy can counteract this dynamic and slow the broader trend toward consolidation, by reducing the need for high paying jobs in the first place. If healthcare and education costs were borne by the government, the opportunity cost of taking a higher paying job would be lower. As a result, the rate at which inventors join small/young firms should increase.

Research Design:
Each inventor $i$ at time $t$ chooses between two job types:
Large firm (L): higher wages, better benefits, lower risk.
Startup/small firm (S): lower wages, weaker benefits, higher risk, more upside (equity/learning).

I model the inventor's decision using a difference-in-utility specification. Let

where

Socialized healthcare/education changes two objects in $\Delta U_{it}$:

Channel A: benefits gap shrinks.
If healthcare/education costs are publicly financed, the advantage of employer-provided benefits declines. In the model, $\Delta B_t$ becomes less negative (moves closer to 0).

Channel B: downside risk shrinks.
If catastrophic health/education expenses are covered, the effective downside risk of startup volatility falls. This can be modeled as a reduction in $\Delta \text{Risk}_t$ (or a reduction in its effective weight for startups).

Define a counterfactual utility difference $\Delta U^{\text{policy}}_{it}$ by replacing $\Delta B_t$ and $\Delta Risk_t$ with their post-policy versions, and compute the counterfactual startup-entry probability.

The main estimand is:

and I also translate this into a wage-equivalent interpretation: how much additional wage would startups have needed to offer (absent the policy) to achieve the same increase in startup entry?

Empirical implementation
Step A: Construct the “inventor choice sample”. Identify inventors using patent–inventor links. Build worker–firm panel to observe job transitions. Define firm type: startup/small/young vs large/established.
Step B: Measure each term with proxies: Wages (w): worker earnings in linked admin data.
Benefits (B): predicted employer benefit value from firm size × industry benefit averages (or insurance offer rates).
Risk: separation/layoff risk; earnings volatility; firm failure risk (firm age + funding stage proxies).
Upside: startup stage/funding, prior growth, or a coarse equity proxy by stage.
Step C: Estimate the simplified choice equation

Predictions
Startup entry rises under socialization holding wages fixed (via the counterfactual).
The effect is \argest for high benefit-dependence inventors (e.g., older workers or those with dependents).
The policy reduces the wage-equivalent gap startups must overcome to attract inventors.

[ukaishan]

Urjita Kaishan
Research question: How evenly is automation risk from artificial intelligence distributed across industries, and does AI function as a general-purpose technology rather than a sector-specific shock?

This project studies AI as a potential general-purpose technology by examining whether exposure to automation risk is concentrated in a small number of industries or broadly distributed across the labor market. Rather than estimating causal effects, the design is descriptive and exploratory, aimed at motivating a larger empirical project on AI, growth, and labor reallocation. This project is conceptually inspired by “Destructive creation, creative destruction, and the paradox of innovation science” (Cao, Chen, Evans, 2022), which emphasizes that transformative technologies often reshape economic activity broadly rather than producing narrow, sector-specific shocks. In particular, the reading argues that modern innovation can generate widespread reorganization of work and production even when observable disruption appears uneven or modest at the aggregate level. This perspective motivates treating artificial intelligence not simply as a technology that replaces specific jobs, but as a general-purpose input whose effects diffuse across many sectors simultaneously.

Guided by this framework, the research design examines cross-industry patterns in automation risk rather than focusing on within-industry task substitution or individual job displacement. The goal is to assess whether projected exposure to automation is meaningfully concentrated in a small set of industries or instead broadly shared across the labor market. This distinction is important for understanding whether AI-driven adjustment should be conceptualized as a localized shock or as an economy-wide reorganization problem.

As a pilot exercise, the analysis uses a synthetic dataset constructed to reflect plausible labor-market patterns based on existing research on automation risk, task codifiability, and occupational characteristics. Each observation corresponds to a job role and includes industry classification and an estimated automation risk percentage. While the data are simulated, their structure mirrors real occupational datasets commonly used in empirical labor economics, making them suitable for illustrating patterns and evaluating the plausibility of the research design. The design aggregates automation risk at the industry level and compares average risk across sectors. This cross-sectional comparison is intentionally simple and differs from prior memos by shifting attention away from prediction or targeting toward dispersion and heterogeneity. In later stages of the project, this framework could be extended using real data sources such as O*NET task measures or BLS employment data to study how industry-level exposure predicts employment reallocation, wage changes, or skill upgrading over time.