Week 2: Models of Innovation and Growth

[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.

“Chapter 4: Creation, in the World Development Report 2024: Middle Income Trap”. The World Bank Group.
“What Do We Learn from Schumpeterian Growth Theory,” Philippe Aghion, Ufuk Akcigit, and Peter Howitt, Handbook of Economic Growth, ed. by P. Aghion, and S. N. Durlauf, 2014, Vol 2B: 515-563.
“Creative Destruction and Economic Growth,” 2023. Ufuk Akcigit, in The Economics of Creative Destruction: New Research on Themes from Aghion and Howitt, ed. by Ufuk Akcigit and John Van Reenen: 21-42.
“Surprising combinations of research contents and contexts are related to impact and emerge with scientific outsiders from distant disciplines.” 2023. Feng Shi & James Evans. Nature Communications.
"Simulating Subjects: The Promise and Peril of AI Stand-ins for Social Agents and Interactions." Download "Simulating Subjects: The Promise and Peril of AI Stand-ins for Social Agents and Interactions."2025. James Evans and Austin Kozlowski.
“Surprise! Measuring Novelty by Simulating Discovery.” 2020. Jacob Foster, Feng Shi and 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]

Corporate - Political Relationship Impact on Innovation
Vera Chaudhry

Research question:
How does innovation behavior change for firms that have a political relationship in the U.S. vs. do not?

Motivation: WDR 2024 cites a study in Italy that showed how politically connected firms change behavior, specifically, they innovate less. While the United States provides a much more independent market than countries abroad, there are often examples of industries, and specific companies, seeking to influence political decisions, or deepen political ties for favorable policy decisions. This project will focus on the impact of this relationship on a company’s innovation.

Research Design:
In this study, company influence in politics / political relationship will be proxied by lobbying dollars spent by a specific firm. The study will look at the following: (a) effect of corporate-political relationship on innovation within a sector and (b) effect of corporate-political relationship on innovation between industries. For effect (a), the study will compare innovation (proxied by patents filed) between different firms in an industry that have different lobbying contributions (ie, contributions that are above or below a certain threshold). For effect (b), the study will compare industries that spend more lobbying dollars to industries that spend less lobbying dollars and compare rates of innovation in these sectors. Of course, this is a slightly harder comparison because we cannot assume parallel trends between industries.

Data visualization:

Data from Open Secrets reveals that within an industry, ie, Healthcare, different drug manufacturers spend different amounts on political influence. This study can use a threshold, ie contributions above $1 million dollars, to see how innovation changes for firms above and below this threshold. An important consideration in this effect is also spillover effects (i.e. if lobbying dollars spent by one firm leads to a policy change that affects all firms).

This chart reveals the amount different sectors contribute to political influence. These sectors are broader than traditional industry groupings, but could be adjusted to compare between different industries.

[ShiyangLai]

Research Question.

Does ex-ante prediction of creative destruction accelerate firm exit and restructuring by triggering reflexive capital and organizational responses, thereby transforming innovation-driven disruption from an exogenous shock into an expectation-driven process?

Research Design.

This project studies creative destruction as a reflexive process in which forecasts of disruption actively reshape economic outcomes. Classic Schumpeterian theory emphasizes innovation as a disruptive force that unpredictably displaces incumbent firms (Joseph Schumpeter). Recent advances in AI-driven forecasting—applied to technological trajectories, firm networks, and knowledge recombination—challenge this assumption by making zones of potential disruption partially predictable ex ante. I propose to examine whether such predictions merely anticipate creative destruction or instead accelerate it through expectation-driven responses.

Empirically, the design combines sector-level forecasting signals with longitudinal firm outcomes. First, I construct a measure of predicted disruption intensity at the industry–year level using machine-learning models trained on historical patterns of patent recombination, citation network reconfiguration, and firm entry/exit. These models generate probabilistic forecasts of heightened creative destruction (e.g., elevated firm turnover or technological displacement). Crucially, these predictions are generated using only information available prior to observed restructuring.

Second, I examine whether industries flagged as “high predicted disruption” experience faster or more severe destruction, conditional on realized innovation activity. Key outcomes include incumbent exit rates, capital reallocation, employment contraction, and changes in market concentration. Identification relies on comparing industries with similar ex post innovation outcomes but differing ex ante predicted disruption levels, allowing me to isolate the reflexive effect of prediction itself. This approach is informed by sociological work on expectations and performativity (e.g., MacKenzie’s analysis of how economic models shape markets), extending it to innovation forecasting.

To strengthen causal inference, I will exploit plausibly exogenous variation in predictive visibility—such as differential analyst coverage, platform access to forecasting tools, or regulatory disclosure regimes—that affects whether disruption predictions become widely salient to investors and firms.

Contribution.

This study reframes creative destruction as an expectation-mediated process, showing how prediction technologies may intensify disruption rather than simply forecast it. By integrating Schumpeterian theory with contemporary AI forecasting, the project offers a novel account of how innovation dynamics change when the future becomes partially visible.

[erindner13]

Erik Rindner

Research Question
Does replacing human variability with AI-simulated training environments change how baseball players learn to hit, particularly their ability to adapt to pitcher mistakes and deception?

Recent advances in artificial intelligence have expanded its role beyond in-game decision-making and into training and skill development. In baseball, AI-driven pitching machines, VR simulators, and algorithmically generated pitch sequences are increasingly used to help hitters practice in highly controlled environments. At first glance, these tools appear ideal for learning: they offer consistency and the ability to isolate specific pitch types/locations. However, baseball hitting is not simply a mechanical task. Hitters are constantly reacting to deception, exploiting pitcher mistakes, and adjusting to unexpected variations in timing and movement. As a result, reducing variability in training environments may fundamentally change how hitters learn the game. This raises a deeper question about what kind of skill development AI training produces. Human pitchers are imperfect by nature; they miss spots, tip pitches, and unintentionally vary their pitches from one to another. These imperfections are not noise to be eliminated, but information that skilled hitters learn to recognize and exploit. AI-simulated environments, by contrast, are often designed to be optimal and consistent. While this may improve short-term performance or technical precision, it may also limit exposure to the very unpredictability that defines real game situations. In this sense, perfect simulation may unintentionally remove an important dimension of learning rather than enhance it. This concern closely aligns with the argument in Simulating Subjects, which suggests that AI systems trained on human behavior can reproduce and standardize that behavior, but in doing so may reduce variability and surprise. The reading emphasizes that simulated environments do not merely replicate human decision-making; they reshape it by constraining the range of experiences learners are exposed to. Applied to baseball, AI-simulated training may create hitters who excel against familiar, repeated pitch patterns but struggle to adapt when faced with novel pitchers, deceptive strategies, or unexpected mistakes.
Methods
To study this question, I propose a comparative research design examining how different training environments affect hitter learning. One group of hitters would train primarily using AI-simulated pitching environments, such as virtual reality systems or algorithmically controlled pitching machines. A second group would train primarily against human pitchers, preserving natural variability and deception. Performance would be evaluated along two dimensions: consistency against familiar pitch patterns and adaptability when facing new or unpredictable pitchers. Key outcomes could include swing decisions, contact quality, and adjustment speed during unfamiliar at-bats. This design allows us to distinguish between improvements in technical repetition and broader learning outcomes related to adaptability and situational awareness. The key outcome of interest is not simply whether AI training improves hitting performance, but how it reshapes the learning process itself. AI-simulated environments may accelerate early skill acquisition while narrowing the range of experiences hitters learn from. In an ideal setting, AI would not replace human variability in training, but complement it by exposing hitters to a broader, more strategically designed range of scenarios that preserves unpredictability while enhancing learning.

[HenryPinner7]

How does AI adoption change the Schumpeterian process of creative destruction within the renewable energy sector – does it lower barriers to entry and speed up displacement, or does it strengthen incumbents by amplifying scale and data advantages?

Background: In canonical Schumpeterian models, growth is driven by innovation that improves “quality” and can displace laggards through competitive replacement. Innovation arrives stochastically as a function of R&D effort (often modeled as a Poisson process) so innovation and reallocation are tightly linked: successful innovators earn rents, but those rents are competed away as frontier firms replace older technologies. The relevant outcome is therefore not “more innovation” in the abstract, but who innovates, and whether innovation shifts competitive advantage toward new entrants or entrenched incumbents. The renewable energy sector is a useful setting to test this logic because it combines fast technological change with sharp competition over siting, interconnection, forecasting, operations, and financing. AI plausibly changes the Schumpeterian process on two margins:

(1) it can raise R&D and operational productivity (better forecasting, predictive maintenance, materials discovery);
(2) it can alter competitive selection if its benefits are complementary to scale, data access, and portfolio depth – features more common among incumbents.

This makes the entrant-incumbent comparison central: if AI disproportionately boosts incumbent innovation capacity, it may increase innovation while muting the displacement channel that defines creative destruction.

Research Design: I would model AI as a firm-level capability that shifts innovation productivity in a Schumpeterian setting. Let firm i choose R&D effort zi​, with innovation arriving at intensity: λ i​= ϕ(AIi​)zi​,
where AIi​ is AI adoption intensity and ϕ(⋅) captures how AI changes the productivity of research and experimentation. The first empirical task is to estimate whether AI changes innovation more for entrants or incumbents. Concretely, I would estimate:

Innovationit=α + βAIit + γIncumbenti + θ(AIit× Incumbenti) + controls + εit,

where θ measures whether AI disproportionately increases innovation among incumbents. AI adoption would be measured using AI-related hiring/postings, ML references in filings/technical docs, and AI-adjacent patenting (with ML text classification to standardize measurement). Innovation outcomes would include patent counts/citations, product releases, and measurable performance improvements (e.g., forecast accuracy, O&M cost reductions). A second-stage extension would connect this to creative destruction by testing whether entrant gains (project awards/capacity share) and churn rise more in contexts where AI is “entrant-biased” versus “incumbent-biased.”

Figure: Figure 1 (shown above) plots innovation output against AI adoption intensity, separately for entrants and incumbents, with illustrative trend lines. This figure operationalizes the memo’s core ambiguity: is the marginal return to AI higher for entrants (supporting creative destruction) or for incumbents (supporting concentration)? If the entrant slope is steeper, AI looks like a barrier-lowering technology that can intensify Schumpeterian competition. If the incumbent slope is steeper or sits higher at comparable AI levels, AI appears complementary to incumbency (data/scale), consistent with “more innovation” without proportional displacement. In other words, the figure directly visualizes the key parameter in the design, heterogeneity in ϕ(AI), and motivates why the entrant–incumbent comparison is the project’s first-order test before moving to market-share churn and exit.

[nbcodes57]

Nick Baumgartner

Research Question:
Does state-owned enterprise dominance in electric power distribution slow artificial intelligence adoption among firms in middle-income countries?

Motivation:
The World Development Report 2024 (Chapter 4) documents that SOEs in infrastructure sectors crowd out private innovation and create bottlenecks for technology diffusion, noting that 60% of developing countries rely on public utilities for electricity distribution. GPTs require complementary investments in infrastructure, organizational capital, and skilled labor before generating productivity gains. These complementary investments may be systematically delayed when state-owned incumbents in enabling sectors face weak incentives to upgrade infrastructure quality and reliability. SOE dominance in energy infrastructure blocks the complementary investments that downstream firms require to adopt GPTs like AI. This creates a bottleneck: even when AI tools are available globally, firms in countries with SOE-dominated infrastructure cannot effectively deploy them due to inadequate enabling infrastructure.

Research Design:
My analysis leverages the dataset compiled by Küfeoğlu, Pollitt, and Anaya (2018) which provides detailed ownership information for electric power distribution systems across 175 countries. This dataset classifies each country's Distribution System Operators (DSOs) into three categories: public, private, or mixed.

I then compliment this data by leveraging the Microsoft Global AI Adoption 2025 report. AI diffusion is measured as the share of people worldwide that have used a generative AI product during the reported period. The measure is derived from aggregated and anonymized Microsoft telemetry and then adjusted to reflect differences in OS and device-market share, Internet penetration, and country population.

I will break down countries by income as well as DSO ownership and compare this to AI adoption rate. I will look to experiment by controlling for various other variables such as electricity access, population per DSO, GDP per capita, and mean years of schooling in order to try to isolate the effect of state owned energy infrastructure on AI adoption.

Above illustrates the causal pathway from SOE dominance to reduced AI adoption.

[sarahkarn]

Sarah Karn

Research question: To what extent is firm entry associated with patent intensity across countries, and in what circumstances can countries with low entry nonetheless sustain high innovation through incumbent firms?

Chapter 4 of the 2024 World Development Report highlights the role of creation in supporting sustained growth. This includes innovations developed by new firms as well as incumbents. A central point of this chapter is that middle-income countries often exhibit weak creative destruction - limited firm entry and growth along with a lack of abundant entrepreneurial activity lead to economic stagnation. Yet, this chapter also recognizes that there are exceptions to this pattern, and that there are high-income economies whose innovation pipeline can be attributed to dominant, established firms, rather than frequent entry. The chapter contrasts nations with different pipelines of innovation. For example, in the United States, newcomer firms heavily contribute to job creation and innovation, while in nations such as Japan and Germany, incumbent firms which invest in long-term research and development, innovating iteratively, are the primary sources of innovation. While this chapter provides a conceptual argument and illustrative examples, a useful extension would be to examine the empirical relationship between firm entry and innovation by conducting a multi-national analysis. Thus arises a key empirical question: how significant is the link between firm entry and innovation outcomes across countries?

The x-axis of my scatter plot measures firm entry as a per-capita value for each nation: the number of new firms per 1000 adults (World Bank). On the y-axis we use patent applications per one million people to represent innovative activity (Our World in Data). Overall, there is a general positive association between firm entry and innovation. However, we can tell a clearer story by examining based on income regime: among middle-income countries such as Brazil, India, and Mexico, extremely low firm entry is strongly linked to very low patent intensity. This aligns with Chapter 4’s assertion that middle-income countries often suffer from a dual failure: both limited firm entry and insufficient innovation by incumbents contribute to weak creative destruction.

Upon examining the high-income regime we see that the relationship between firm entry and patent applications is more variable. For instance, South Korea and Israel exhibit the highest indicators of innovation despite having modest levels of firm entry. The empirical evidence reveals that for high-income countries, new firm entry is not a necessary condition for high innovation. It may still be possible for incumbent firms to carry their weight via access to large amounts of capital, skilled labor, and highly capable institutions. For middle to low income countries however, inadequate firm entry in the vast majority of cases does inhibit innovation.

Overall, this plot supports the chapter’s argument that innovation pipelines and dynamics differ across countries. In particular, we see that modest to high firm entry seems to be a prerequisite for innovation in middle-income countries. Thus, a natural investigation of this study would be to examine which incentive structures are most effective in driving the development, scaling, and competitive viability of new firms. Such an analysis could help identify policy mechanisms to alleviate innovation-related stagnation in middle-income countries.

Data sources:
https://data.worldbank.org/indicator/IC.BUS.NDNS.ZS?utm_source=chatgpt.com
https://ourworldindata.org/grapher/patent-applications-per-million?utm_source=chatgpt.com

[CristianG-png]

Cristian Garcia

How can AI‑driven innovation and immigration‑driven labor mobility work together to help countries avoid the middle‑income trap? + Why blaming immigrants for job loss is a counterproductive view of the job market.

Our Tuesday reading of The World Development Report emphasizes that escaping the middle-income trap requires balancing the forces of creative destruction: Creation (innovation, new technologies, new firms), Destruction (reallocation away from outdated tasks), Preservation (institutions that maintain stability without blocking change). AI acts as a powerful creation force when used as a GPT by raising the productivity of routine tasks and lowering prediction costs. Rather than simply replacing workers, AI reallocates labor toward non‑routine, interpersonal, and creative tasks, similar to the Schumpeterian mechanism of structural upgrading we learned in class. Immigration supports creation as migrants bring skills, ideas, and entrepreneurial energy, filling essential roles in sectors that expand as AI raises productivity. Basso, Peri & Rahman in “Computerization and immigration” show that immigrants complement computerization: technology substitutes for routine cognitive tasks, while immigrants disproportionately work in manual, service, and care roles that technology does not replace. Immigration also strengthens innovative ecosystems by increasing diversity and knowledge flow. Yet public narratives often blame immigrants for “taking jobs,” even as AI automates many of the same routine tasks. This misdiagnosis undermines growth: restricting immigration weakens the very complementarities that help countries transition from infusion to innovation, the final step needed to escape the middle‑income trap.

My research design integrates the WDR’s creative destruction model, Basso–Peri–Rahman’s complementarity findings, and Kozlowski & Evans’s AI stand‑in methodology.

1. Construct occupation level AI exposure scores - Use routineness, prediction intensity, and linguistic task dependence (O*NET + OECD automation indices).
2. Build AI “labor stand ins”- Simulate how tasks in high exposure occupations would be completed under varying AI capability levels, incorporating Kozlowski & Evans’s distortions (bias, uniformity, atemporality, disembodiment, linguistic cultures, alien intelligence).
3. Simulate three development paths - Aligned with the WDR’s 1i → 2i → 3i framework:
   a. AI‑Only (Creation without Mobility) - Rapid AI adoption but restricted immigration. Innovation potential constrained by labor shortages and institutional inertia.
   b. Immigration‑Only (Infusion without Technological Upgrading) - Rising immigration but slow AI adoption. The country remains a technological taker.
   c. AI + Immigration Complementarity (Innovation Strategy) - AI accelerates creation; immigration supplies talent and fills complementary sectors. Structural transformation is faster.
4. Evaluate structural transformation outcomes - Measure routine‑task decline, non‑routine expansion, productivity growth, innovation intensity, and firm dynamism.
5. Compute an “escape probability” - Estimate the likelihood of transitioning from middle‑income to high‑income status under each scenario.
   To help support this, doing a sensitivity analysis using empirical ranges inspired by various studies on the job market reveals that if AI continues to be the only thing considered in terms of the job market, we will miss out on potential growth from AI.

The data from the above visualization are from various papers:
Immigration displacement range: –0.3% to +0.3% (from Basso, Peri & Rahman’s work), which implies immigration has a negligible effect on job loss, potentially even being a reason for new jobs.
AI‑driven routine‑task displacement range: –3% to –8%, there are many studies on this with varying results due to AI affecting job markets differently, but an article from the Federal Reserve Bank of Dallas estimated it to be over 3%, while others had as high at 10%. The range addresses both estimates.
Observed U.S. routine‑occupation decline range: –2% to –5%. Data from the Bureau of Labor Statistics shows that in December 2025, the unemployment rate had dropped to 4.4 percent, but I could not find a statistic for the employment rate; other sources claim it is low so the range is an estimate.

[zararania]

Zara Rania Suran

Research Question:
Is there empirical evidence of strategic deterrence in AI-driven biology, where the release of dominant, proprietary industry models (e.g., AlphaFold 3) causes an 'exit rate' among academic laboratories working on similar problems, effectively halting creative destruction by discouraging new entrant competition?

Background and Motivation:
This project investigates the discouraged follower effect described by Ufuk Akcigit (2023), who argues that when market leaders engage in defensive innovation and widen the technology gap too far, followers often give up on investing in R&D, leading to a decline in business dynamism. I posit that academic laboratories are currently the "discouraged followers" in the face of TechBio incumbents. As the World Development Report 2024 warns, actors unable to transition from simple "infusion" (using tools) to "innovation" (creating them) effectively lose their agency; for academic labs, this inability to compete with industry compute results in market exit. If proprietary foundation models are creating a kill zone where the cost of entry is too high, we should see a direct correlation between the prevalence of industry AI models in a subfield and the rate at which academic PIs abandon that research area. Validating this would prove that the current AI paradigm might reduce Schumpeterian competition rather than enhancing it.

Research Design
I propose a mixed-methods approach combining bibliometric tracking with direct scientist polling across five distinct biological subfields (ex: structural biology, proteomics, ecology). First, I will calculate the lab exit rate-the percentage of principal investigators who ceased publishing methodological work in their field post-2021-using OpenAlex metadata. Second, to confirm the mechanism, I will conduct a survey of these exiting scientists to determine if their departure was specifically due to discouragement from industry dominance, as predicted by Akcigit’s blocking theory. I hypothesize a strong positive correlation where fields with the highest industry AI prevalence (e.g., structural biology) will exhibit the highest lab exit rates, while fields with low industry encroachment (e.g., ecology) will maintain healthy entrant retention.

This visualization demonstrates a strong positive correlation between the prevalence of proprietary industry AI models and the exit rate of academic laboratories, showing that fields like structural biology face significantly higher researcher attrition than less digitized areas like ecology. By quantifying this discouraged follower effect, the data suggests that high-capital foundation models act as barriers to entry that stifle Schumpeterian competition. These findings could fundamentally reshape science policy by demonstrating that without training current scientists to be proficient in the use of such AI models, displacement of PIs may be likely to occur, along with a larger shift in who makes scientific discoveries and what kinds of technologies they have the ability to employ.

[felicechan]

Felice Chan
Research Question: Does artificial intelligence enable creative destruction and innovation-led growth, or does its impact depend on whether economic institutions permit entry, exit, and novelty? I focus on economic freedom as the institutional precondition for creative destruction in the age of AI, especially in low- and middle-income economies where growth often stalls.
Motivation and Theory: Schumpeterian growth theory (Aghion, Akcigit, and Howitt) identifies creative destruction as the core engine of long-run growth. Innovation is valuable precisely because it disrupts incumbents, reallocating resources toward more productive uses. Akcigit (2023) emphasizes that this process is inherently political: incumbents have incentives to block entry and protect rents, so growth depends on institutions that discipline preservation and enable destruction.
Chapter 4 of the World Development Report 2024 applies this logic to development. Middle-income countries stagnate not because they lack technology, but because firm dynamics are weak: too few entrants, too little exit, and many firms that “flat and stay.” Policies and regulations often protect incumbents, suppressing the reallocation needed for innovation-led growth. In this framework, economic freedom: low barriers to entry and exit, competitive markets, and mobility, is central to creation.
AI complicates this mechanism. AI is often framed as a general-purpose technology that lowers costs of prediction, search, and coordination. Yet the novelty and impact literature suggests that innovation depends on unexpected recombination and outsider participation, not just technical capability. Shi and Evans (2023) show that high-impact innovation often emerges from surprising combinations produced by actors distant from dominant fields. Foster, Shi, and Evans (2020) formalize novelty as “surprise,” emphasizing that innovation outcomes depend on the structure of search and evaluation systems.
At the same time, Evans and Kozlowski (2025) caution that AI systems tend to simulate dominant patterns embedded in data and institutions. If AI adoption is concentrated among incumbents, it may reduce uncertainty while narrowing the space for outsider novelty. In Schumpeterian terms, AI could strengthen preservation rather than destruction, especially in low–economic freedom environments.
Empirical Case and Demonstration: My final project will test whether AI diffusion translates into creative destruction only where economic freedom is high. Using a cross-country panel, I will measure AI diffusion via patent data and proxy creative destruction using indicators such as new firm entry, firm exit, or growth in non-AI innovation. Countries will be stratified by institutional measures related to economic freedom.
Analytical visualization: The core visualization plots AI diffusion against changes in creative-destruction outcomes, with separate trend lines for high– and low–economic freedom contexts. A positive slope only in high–freedom environments would suggest that AI enables innovation by amplifying creative destruction. A flat or negative relationship elsewhere would indicate that AI adoption reinforces incumbent dominance and limits novelty.
Contribution: This memo advances a focused claim: AI is not an independent engine of innovation; its growth effects are mediated by the institutions that govern creative destruction. By centering economic freedom, the project links Schumpeterian theory, development economics, and the novelty literature to a concrete empirical test, opening discussions on whether AI can revive, or instead bypass, the institutional foundations of innovation.

[JesseZhou-1]

Jesse Zhou

Research question: Do periods of creative destruction show up as new task and skill combinations within occupations?

Schumpeterian growth theory frames development as a process of creative destruction: new techniques displace old ones, reshaping how production is organized and how rewards are distributed. I apply that idea to occupational skill structure. If new technologies lower the cost of producing, coordinating, or experimenting with new ways of doing work, then “creation” should appear not only as firm entry and exit, but also as within-occupation recombination: new roles, new task bundles, and new mixes of skills.

Empirically, I will measure creative destruction at the occupation level using Jeff Lin’s New Work data, which reports the share of new occupational titles within detailed occupations (e.g., 1980, 1990, 2000). I interpret a higher share of new titles as revealed occupational novelty, capturing task reorganization and new role creation. To connect occupational novelty to changing skill demands, I will construct an emergent skill index for each occupation. In the modern period, I will select ONET items that capture learning, complex problem solving, creativity, and systems thinking, then apply PCA to build a one-dimensional emergent-skill score. I will backcast this index to pre-ONET decades using DOT descriptors by estimating a prediction function in the overlap period and validating out of sample.

The core empirical test is whether occupations with more “new work” also show sharper upgrading and recombination in emergent skills. To deepen the measurement, I will build two additional novelty metrics that draw on the week’s readings. First, following the idea that impact can come from surprising combinations of contents and contexts, I will measure novelty as the rarity of skill combinations within occupations, treating skills as “contents” and occupations as “contexts.” Second, inspired by “Surprise! Measuring Novelty by Simulating Discovery,” I will simulate the distribution of skill combinations expected under a baseline model and define excess novelty as deviation from that baseline. Finally, “Simulating Subjects” motivates a careful, limited use of LLMs: I can generate plausible task bundles conditional on occupation and compare simulated novelty to observed novelty as a robustness check, not as primary evidence.

[Arthur-Langlois]

The Constitutional Origins of Creative Destruction

Research Question

What constitutional characteristics enable creative destruction in Middle-Income Countries (MICs)?

Background and Motivation

Growth and innovation are driven by many factors, but a crucial component is an economic system in which creative destruction is encouraged. Creative destruction is the process by which existing technologies and firms are replaced by new, more efficient ones, raising aggregate productivity. Sustained growth therefore requires institutional arrangements that incentivize and accommodate this process.

Creative destruction can be facilitated (or constrained) by a range of institutional factors, including the legal system, political interests, and cultural norms. In this project, we study which legal characteristics---as embedded in constitutional frameworks---are associated with creative destruction. Although constitutional law does not encompass the entirety of a country's legal system, it provides the foundation of the broader legal framework. In particular, constitutional provisions may ``trickle down'' by shaping court opinions and legislation that ultimately encourage (or discourage) creative destruction. We hypothesize that provisions related to labor, intellectual property, and antitrust may be especially relevant in this context.

In an attempt to answer the broader question of how to design a legal system that encourages creative destruction, I ask: which constitutional characteristics are correlated with creative destruction?

Method

Because development and innovation are slow-moving processes, they should be studied in historical perspective, using long time horizons to better understand their drivers. For this reason, we use historical data on constitutional features for roughly 200 countries since 1789 from the Comparative Constitutions Project, paired with ex post measures of creative destruction. In particular, we use measures such as firm entry and exit rates---births/deaths per year as a percentage of active firms.

Examples of constitutional characteristics studied include: mentions of intellectual property rights; the right to a free or competitive market; references to economic planning; mentions of economic rights; the right to enjoy the benefits of scientific progress; references to science or the sciences; and academic freedom.

We follow a two-stage (post-LASSO) approach. In the first stage, LASSO performs variable selection by shrinking irrelevant coefficients toward zero. In the second stage, we run OLS on the selected set to obtain more precise estimates of the relationship between the selected constitutional characteristics and measures of creative destruction (e.g., $\frac{\text{yearly births/deaths}}{\text{active firms}}$). We then use these OLS estimates to examine which countries appear to benefit the most (as measured by creative destruction) from a given constitutional characteristic. While this design does not address endogeneity, we view it as valuable for documenting correlational patterns between constitutional characteristics and objective measures of creative destruction.

Stage 1: High-dimensional ``screening'' regression (selected potential drivers in the Comparative Constitutions Data Project).

$P_{ct} = \alpha_c + \gamma_t + \mathbf{X}_{ct}'\boldsymbol{\beta} + \varepsilon_{ct}$,

where $\mathbf{X}_{ct}$ stacks all candidate constitutional drivers listed above. LASSO is applied to this equation to select a subset of predictors for the second stage.

Stage 2: Post-LASSO estimation (selected drivers; unknown ex ante).

$P_{ct}= \alpha_c + \gamma_t + \sum_{j\in \widehat{S}} \beta_j\,X_{j,ct}+ u_{ct}$

We then estimate this specification by OLS using the predictors with non-zero coefficients selected in Stage 1.

[hcarpenter4]

Research Question:

Does the trend of large AI companies acquiring smaller companies lead to more or less innovation in the AI space?

Motivation:

In our discussion of creative destruction, we learned that incumbent firms can respond in different ways to competition from entrants. They can engage in anticompetitive behavior in an attempt to preserve their dominant position in the market, or they can contribute to innovation in order to outcompete entrants and create a technological gap between competitors. With hundreds of M&A deals being completed in the AI industry each year, it is important for the health of the industry to characterize these acquisitions by large incumbents as trying to stifle innovation by smaller competitors, or an attempt to increase innovation by scaling new ideas.

Research Design:

I propose a difference-in-differences design comparing innovation metrics for acquired AI startups versus similar non-acquired firms before and after acquisition events. This approach allows us to isolate the causal effect of acquisitions on innovation while controlling for time-invariant firm characteristics and broader industry trends.

I will construct a dataset of AI company acquisitions by major tech firms (Google, Microsoft, Meta, Amazon, Apple) from 2015-2024. For each acquired company, I will identify 2-3 control firms matched on pre-acquisition characteristics: founding year, funding stage, technical focus area (computer vision, NLP, robotics, etc.), and patent count.
In order to measure innovation, I will track multiple outputs: (1) patent filings and citations as measures of technological innovation; (2) academic publications by company researchers; and (3) product launches. For firms that are acquired, I will make sure that these metrics are attributable to the acquired company, and not just the parent company, by matching contributor names in patents, publications, and product launches to members of the acquired firm.

3 years after acquisition (Post_t), the total innovative output of the acquired company since acquisition will be measured compared to the output of the counterfactual firms during the same period. This will allow me to construct the following DiD regression:

Y_it = β₀ + β₁(Acquired_i) + β₂(Post_t) + β₃(Acquired_i × Post_t) + ε_it

Where:
Y_it = Innovation output (patents, publications, etc.) for firm i at time t
Acquired_i = 1 if firm was acquired, 0 if control
Post_t = 1 for periods after acquisition, 0 before
β₃ = The causal effect of acquisition on innovation

β₃ tells us how much more (or less) innovation output changed for acquired firms relative to control firms after the acquisition event.

If this analysis is conducted separately within different AI subfields, the difference in innovation output after acquisition may vary substantially across sectors. For example, subfields that benefit most from scaling up - such as NLP - may experience increased innovation following acquisition due to greater access to compute infrastructure, proprietary datasets, and cloud resources that are prohibitively expensive for independent startups. Conversely, robotics companies may be harmed by acquisitions because their innovation depends more heavily on rapid iteration, hardware experimentation, and maintaining diverse research directions - capabilities that can be constrained within large corporate structures.

[johnab20]

Research Question

To what extent does the transition from technology infusion (imitating established analytics) to frontier innovation (system-integrated AI) determine whether MLB teams escape the Middle-Income Trap of diminishing growth efficiency?

Research Design

The World Development Report 2024 defines the "2I" strategy (investment and infusion) as the stage where entities grow by borrowing and diffusing existing technologies. I propose that many MLB teams are stuck in this phase, hiring large analyst headcounts simply to "infuse" existing industry-standard metrics (e.g., exit velocity, spin rate) without redesigning their underlying organizational systems. These teams exhibit "flat and stay" dynamics—they maintain high staffing levels but fail to scale productivity

Conversely, "System-Innovators" represent the "3I" strategy (innovation), where they move beyond borrowing ideas to pushing the global technology frontier. Using the framework from Aghion et al., these elite organizations utilize an "escape competition" effect: because they are "neck-and-neck" at the frontier, they are incentivized to innovate more intensely to maintain their lead.

To test this, I will refine the original regression by adding a "Novelty Index" based on the Shi and Evans reading. This index will measure the "surprising combinations" of backgrounds in a team's front office (e.g., hiring outsiders from physics or biology) to see if outsider-driven novelty is the catalyst that allows high-headcount teams to break the trend of diminishing returns.

This allows me to see if the "middle-income" stagnation in MLB is caused by a lack of creative destruction—where unproductive legacy scouting methods are not being "destroyed" to make room for new AI-driven arrangements.

The regression analysis identifies a clear Middle-Income Trap where the negative slope of the league trend reveals that increasing analyst headcount often leads to diminishing growth efficiency. While "2I" teams like the Baltimore Orioles thrive by infusing standard technology, high-investment teams like the Philadelphia Phillies fall into "Innovation Saturation" when legacy institutional inertia smothers their expanded data staff. Conversely, the Tampa Bay Rays serve as the model for Frontier Innovation (3I), successfully utilizing Creative Destruction to break the saturation trend and maintain elite efficiency despite a high headcount. Ultimately, the wide distribution of residuals—from the lean efficiency of the Oakland Athletics to the stagnation of the New York Yankees—suggests that organizational success depends less on the total force of creation and more on the novel integration of those inputs.

John Butka

[TarunSethi13]

Tarun Sethi

Research Question:

Are Indian cities with greater Global Capability Center (GCC) activity characterized by higher domestic firm entry, consistent with stronger forces of creation emphasized in the creative destruction framework?

A central argument in Chapter 4 of the World Development Report 2024 is that growth in middle-income countries is constrained not by a lack of ideas but by weak creative destruction. Specifically, too few small firms enter and scale to challenge incumbents, while existing firms face limited pressure to upgrade capabilities or infuse global technologies. Consistent with Schumpeterian frameworks, countries farther from the technological frontier grow primarily through diffusion, reallocation, and entry rather than through frontier invention.

This memo studies Global Capability Centers (GCCs) as a concrete organizational phenomenon that may be associated with stronger forces of creation. GCCs are offshore hubs where multinational firms conduct software development, analytics, engineering, and business process work using digital coordination technologies. More broadly, GCCs reflect the diffusion of digital coordination technologies that reduce geographic constraints on production, enabling firms to integrate global workflows across locations and weakening the traditional link between innovation and physical proximity. While GCCs are not themselves new domestic firms, they embed global organizational practices into local economies and reshape local labor markets. By training workers in frontier workflows, management routines, and digital tools, GCCs expand workers’ outside options and increase the contestability of labor. In Schumpeterian models, it is this reallocation and mobility of skilled labor—rather than skill accumulation alone—that weakens incumbent advantage and enables entry-driven creative destruction.

Research Design

I propose a cross-city comparative analysis within India that examines the relationship between GCC presence and domestic firm entry. The unit of analysis is the city. The primary outcome is firm entry, measured using new firm registrations. The key explanatory variable is GCC intensity, proxied by the approximate number of GCCs located in each city.

The baseline analysis compares firm entry across cities with differing levels of GCC activity. Because GCCs may locate in cities already on high-growth trajectories, the analysis is interpreted as descriptive rather than causal. As an extension, the design could be expanded into a city-year panel with fixed effects to examine within-city changes over time as GCC activity grows.

Data

GCC counts are approximated using city-level shares of India’s approximately 1,600 GCCs, as reported in Knight Frank & 3AI (2024), Global Capability Centres and India Office Market Landscape. Firm entry is measured using state-level new company registrations from India’s Ministry of Corporate Affairs, drawn from the Registration of New Companies (State-wise Distribution) dataset. Major cities are mapped to their respective states (e.g., Bengaluru–Karnataka) .

Contribution

The observed association between GCC presence and higher domestic firm entry is consistent with the interpretation that GCCs—by reshaping local labor markets and increasing worker mobility—may strengthen forces of creation emphasized in the creative destruction framework. At the same time, the analysis is explicitly descriptive rather than causal: GCCs may co-locate with other growth-enhancing factors, and firm entry is used as a proxy due to the absence of publicly available city-level data on innovation outcomes such as R&D or technology adoption. These limitations motivate future work that links GCC exposure to richer firm- and labor-level measures of innovative activity.

[Rjlee22]

Ryan Lee

Question: How does the interaction between product market competition and firm-level innovation incentives differ across stages of economic development, and can this interaction explain divergent long-run growth trajectories among middle-income countries?

Motivation: This is a proposal of a research design to study how competition shapes innovation-driven growth differently depending on a country’s distance from the global technology frontier. The question is motivated by Schumpeterian growth theory and creative disruption. A key insight from this framework is that competition does not have a uniform effect on innovation: it can stimulate innovation for firms near the technological frontier while discouraging it for lagging firms, leading to heterogeneous growth outcomes.

Research Design: To address this question, I propose a firm-level, cross-country empirical design that exploits variation in competition induced by trade in middle-income economies. This design uses data to directly test the mechanisms highlighted by Schumpeterian theory. The core idea is to examine whether increased competitive pressure disproportionately boosts innovation among frontier firms, leading to accelerating growth in some countries while weakening innovation incentives among non-frontier firms in others.
The empirical strategy proceeds in three steps. First, I would create a dataset of manufacturing firms across several middle-income countries that implemented trade reforms and firm-level data on productivity, patents, R&D spending, and exit rates would be drawn from sources. Second, firms would be classified based on their proximity to the global technology frontier within narrowly defined industries, following the approach suggested by Schumpeterian models of step-by-step innovation. Third, I would estimate a difference-in-differences model comparing innovation outcomes before and after trade liberalization, interacting competition shocks with firms’ distance to the frontier.

This design directly tests two core predictions from Aghion et al. (2013): (i) that competition induces an “escape competition” effect for frontier firms, increasing innovation, and (ii) that the same increase in competition may generate a Schumpeterian discouragement effect for lacking firms. At the country level, heterogeneous firm responses can be aggregated to explain why some middle-income economies experience sustained innovation-led growth while others stagnate.